How long does it take to get an assistant live?

A basic assistant with uploaded knowledge and a deployed widget can be running quickly, because the work is configuration rather than development. The variable is knowledge preparation. Businesses with current, well-organised documentation move fastest; businesses whose policies are scattered across old files spend most of their setup time on content rather than on the platform.

Do we need developers to use NextPro AI?

Not for the core workflow. Creating assistants, uploading knowledge, configuring behaviour, customising the widget and deploying with an embed script all happen through the interface. Developers become useful for custom REST API integrations, webhook workflows and advanced widget styling.

What should our first assistant do?

Pick the single job with the highest volume of repetitive questions, which for most businesses is customer support. Narrow scope produces better answers, a faster launch, and cleaner data about what to improve. Expanding into sales, product or internal assistants is straightforward once the first one works.

How much content do we need before starting?

Less than most teams assume. Coverage of your ten to twenty most frequent questions is enough for a useful first deployment. Volume matters far less than accuracy and currency.

What file types can we upload as knowledge?

PDF, DOCX, XLSX, XLS, CSV, TXT and Markdown. Legacy .doc files need saving as .docx first. Knowledge can also come from pages imported by URL and from articles published inside the platform.

How does the assistant avoid making things up?

Retrieval-augmented generation grounds responses in your own content rather than the model's general knowledge, and source citations let you verify where an answer came from. Assistant instructions define behaviour when it does not know, and human handoff provides a route out of conversations it should not attempt. No system removes the risk entirely, which is why pre-launch testing and post-launch monitoring both matter.

What happens when the assistant does not know something?

That depends on how you configure it, which is why the instruction covering unknown questions is one of the most important settings on the page. A well-configured assistant acknowledges the gap rather than guessing, and where handoff is enabled it can route the conversation to a person.

Can we see which source an answer came from?

Yes, where citations are enabled. Responses can display the supporting sources — the specific document or article the answer was drawn from. This is what makes answer quality auditable rather than a matter of trust.

How do we keep the assistant accurate as our business changes?

Update the underlying knowledge and re-index. Imported pages can be re-synced on demand from the document itself. Knowledge insights surface the questions the assistant is failing, so you know which gaps to close first.

What happens if we upload conflicting documents?

The assistant may retrieve either version, which is exactly the problem it sounds like. Removing superseded documents matters more than adding new ones — outdated content produces confidently wrong answers, which are worse than honest non-answers.

Can the assistant use content from our website?

Yes. You can import a page by URL into a knowledge base, and it runs through the same validate, extract, process, chunk and index pipeline as an uploaded file. Imported pages keep their source URL and can be re-synced when the page changes.

Does uploading more documents make the assistant smarter?

Not reliably. Beyond a point, additional documents mainly increase the chance of retrieving the wrong passage. Precision beats volume, and knowledge usage analytics show which sources are earning their place.

Can we use the same knowledge for our help center and our assistant?

Yes. Articles, FAQs, policies and documentation published in the platform can serve a customer-facing help center and act as assistant knowledge at the same time, so one update maintains both.

Can we choose our own AI model or provider?

NextPro AI supports multiple providers and models, configured at the platform level. Tenants select from the models their subscription and purchased capabilities permit.

Why can't we see every model available?

Model availability is controlled deliberately. Without that control, model selection becomes an unmanaged cost centre where anyone can pick the most expensive option for a task that did not need it.

Does a better model fix a bad answer?

Usually not. If the assistant received the wrong context, a more capable model writes a more articulate wrong answer. Retrieval quality is the first thing to check when answers disappoint; model selection is the second.

What is RAG, in plain terms?

Retrieval-augmented generation means the system searches your knowledge for passages relevant to the question, then asks the model to answer using those passages. The model contributes language ability; your content contributes the facts.

What does re-indexing mean and when do we need it?

Indexing is the processing step that makes your content searchable by the assistant. Re-indexing repeats it after content changes, so updated documents actually reach the assistant. Any time you revise, add or remove knowledge, re-index and then re-test the affected questions.

Can we control where the widget is allowed to run?

Yes. Domain-based deployment lets you specify which domains an assistant may run on, which prevents an embed configuration being reused on unauthorised sites and keeps staging separate from production.

Does the widget work on mobile?

Yes, across desktop, laptop, tablet and mobile, and you can preview each before deploying. Mobile preview is worth taking seriously, since placement that works on a wide screen frequently obscures important controls on a phone.

Can we deploy the same assistant in more than one place?

Yes, subject to your domain configuration. A single assistant can serve several approved domains, which suits businesses running a marketing site, an app subdomain and a documentation site.

Can we run more than one assistant?

Yes, subject to the assistant capacity of your plan. Separate assistants for support, sales and internal knowledge is the recommended pattern, since each can have its own personality, knowledge scope, model, integrations and deployment.

What can an assistant do beyond answering?

With automation enabled and an integration connected, an assistant can identify a request, call a connected system, retrieve information such as an order status, and return the result in conversation. Depending on the integration it may also create requests, retrieve records, schedule actions or trigger workflows.

Do we need automation on day one?

No, and starting with it is usually a mistake. Get the knowledge layer answering reliably first. Connecting live systems to an assistant whose answers are still inconsistent multiplies the problem rather than solving it.

Is it safe to let an assistant take actions?

Start with read-only actions such as retrieving an order or checking a booking, which cover a large share of the value at a fraction of the risk. Move to actions that create or modify records once you have evidence the assistant identifies intent correctly.

When should handoff trigger?

Configure it for situations where a wrong answer is expensive: complaints, billing disputes, cancellations, anything with legal or compliance weight, and repeated failed attempts in one conversation. An explicit request to speak to a person should always route through.

Conversations, messages, active users, AI requests, credits consumed, response time, resolution rate, popular questions, knowledge usage, widget usage and unanswered questions.

What is a knowledge gap?

A recurring question the assistant repeatedly fails to answer well. You review the question, add knowledge or publish an article, re-index, test, and confirm the improvement. That cycle is the core maintenance loop.

How often should we review performance?

Weekly for the first two months, focused on unanswered questions, then monthly once the obvious gaps are closed. Quarterly, compare against the baseline you recorded before launch.

What is a good resolution rate?

There is no universal number, because it depends entirely on the mix of questions your assistant receives. Treat it as a trend line for your own deployment rather than a benchmark against other companies.

What are credits and what consumes them?

Credits meter AI usage. Requests to the assistant consume them, and consumption varies with configuration — including how much knowledge is retrieved and which model is used. Usage is tracked continuously and shown against your allocation.

What happens when we run out of credits?

Usage is tracked continuously and you are notified as you approach configured limits. Additional credits and plan upgrades are available according to the platform's commercial model. Testing before launch gives you a forecast in advance rather than a surprise afterwards.

What is the difference between a plan and a module?

A plan sets your baseline capacity and included features. Modules are individual capabilities purchased separately — advanced retrieval, advanced analytics, automation, premium widget, human handoff, help center, knowledge insights, API access, additional credits or additional storage. You pay for what you need rather than a tier bundle you partly use.

Why does a module say “Pending approval”?

Some modules require platform review after purchase. The flow is purchase, confirmation, pending approval, review, approval or rejection, notification, setup, then active. This exists for capabilities needing review or manual provisioning.

Can we upgrade later?

Yes. The plan and module structure is designed for capabilities to be added as requirements grow, rather than requiring you to choose your eventual configuration on day one.

Who inside our company can access what?

Role-based access controls permissions within your tenant, so administrators, editors and support users get appropriate levels of access to assistants, knowledge, integrations and analytics.

What administrative controls exist at the platform level?

NextGenIT Solutions manages plans, modules, marketplace availability, tenants, module approvals, AI providers, credit rules, billing, platform analytics, permissions, feature flags, notifications, audit logs and system configuration through a Super Admin environment.

Can one account manage assistants for multiple clients?

Yes — that is a core design goal. One tenant environment can carry multiple deployments, each with its own assistant, knowledge, branding, widget, domain, integrations and analytics.

Can each client get their own branding?

Yes. Widget branding, colours, avatar and conversational experience are configured per deployment, so each client's assistant looks like their product rather than yours.

Widget sessions now authenticate with short-lived signed tokens

The AI assistant platform that knows
your business, not just AI.

Upload your documents and NextPro AI turns them into assistants that answer with citations, take real actions through your systems, and deploy to your dashboard or any website — with credits, entitlements and analytics built in from day one.

Start building free See how it works

No infrastructure to run. 14-day free trial, no card required.

Customer support
Sales enablement
Internal helpdesk
Operations
Finance

Active assistants+84%

5 agents

Credits used this month+47%

6,550

Automated resolution+24%

92.4%

Tenant credit balance

$675,931

Docs ingested 1,48098% synced

Storage used420 MB

Conversation volumeLast 30 days

The gap

Most business AI stops at the conversation

There is a moment that repeats in almost every AI rollout. The assistant goes live, a customer types a real question, and the answer comes back generic. Polite, fluent, and useless — it describes returns policies in general instead of your returns policy.

The failure is rarely the language model. Models are good now. The failure is that the model was never given the two things it needed: your business knowledge, and permission to act.

Most companies already have the knowledge. It is just scattered — some in PDFs on a shared drive, some in a help center last updated two years ago, some in a spreadsheet one person maintains, and a lot of it in the heads of the support team answering the same eleven questions every week. Meanwhile the order status a customer wants lives in a backend the chatbot has never heard of.

Closing that gap is not one project. It is six.

Ingest and index the documents so retrieval actually works
Choose and configure models, then keep costs from running away
Wire the assistant into the CRM, helpdesk or store
Build a widget that matches the brand and behaves on mobile
Make the thing measurable rather than a black box
Keep improving it once real questions start arriving

Bought separately, that becomes a vector database, an orchestration layer, a model provider account, an analytics tool, a widget builder, and an engineer to hold it together. Built in-house, it becomes a roadmap item competing with your actual product. NextPro AI exists to collapse that stack into one platform.

How it works

From documents to a working assistant

Four steps, no ML engineering, no infrastructure to provision.

01
Connect your knowledge
Upload PDFs, Word docs, spreadsheets or Markdown, or point the crawler at your website. Everything is parsed, chunked and embedded automatically.
02
Shape the assistant
Set its purpose, tone and model, pick which knowledge bases it can read, and give it tools it's allowed to call.
03
Test before you ship
Run real questions in the Test Studio and inspect exactly which chunks were retrieved, which tools fired and where the time went.
04
Deploy anywhere
Paste one script tag onto any site, or use the dashboard chat. Each visitor session authenticates with its own short-lived signed token.

Product tour

A look inside the platform

Build, monitor and operate every assistant from one console.

nextproai.com/workspace/assistants/new

Assistant name

Support Copilot

Purpose

Customer support

Tone

Friendly, professional

Model

GPT-4o

Knowledge

48 documents

Tools enabled

order_lookupcreate_ticketsend_email

Live previewgpt-4o

Do you ship to Canada?

Yes — standard delivery to Canada takes 5–7 business days, with free returns within 30 days.

Shipping-Policy.pdf

Configure and preview side by side

Set instructions, tone, model and tools on the left; ask the assistant a real question on the right. Nothing ships until you have seen it answer.

A shared inbox for the hard ones

When a visitor asks for a person, the assistant pauses itself and the thread appears here with its full history — no copy-pasting between tools.

nextproai.com/workspace/inbox

Search conversations…Escalated

Order #45821 delayedSupport Copilot · 2m agoEscalated
Refund on damaged itemSupport Copilot · 14m agoActive
Bulk pricing questionSupport Copilot · 1h agoResolved

Assistant paused — a teammate can reply, resolve or hand it back.

nextproai.com/workspace/analytics

Resolution rate

92.4%

Conversations

1,284

Credits used

6,550

Conversations per day

Knowledge gaps detected

“Do you offer student discounts?”14×
“What is the SLA for priority support?”9×

See what it can't answer yet

Beyond volume and credit burn, NextPro AI groups the questions your assistants failed to answer — so you know exactly which document to write next.

Run it for many customers

The super-admin console manages tenants, plans, credits and the assistant catalogue across your whole platform.

nextproai.com/admin/tenants

TenantPlanStatus

Northwind RetailGrowthActive

Harbor LogisticsEnterpriseActive

Bright HealthStarterTrial

Each tenant is fully isolated — data, credits, plan limits and users.

Interface shown with sample data.

The lifecycle

The AI assistant platform lifecycle, in nine stages

The platform is organised around a lifecycle rather than a feature list. Every capability maps to a stage of getting an assistant live and keeping it useful afterwards.

Create an assistant

Give it a name, a purpose, a personality and a job. A support assistant and a sales assistant should not behave the same way, so they are configured separately from the start.

Add business knowledge

Upload documents, import a documentation page by URL, or publish articles inside the platform. Everything is parsed, chunked and indexed for retrieval.

Configure the model

Select from the providers and models your subscription permits. Model choice is a setting, not a rebuild.

Connect tools

Link the CRM, helpdesk, store, calendar or any REST API so the assistant can reach live business data.

Automate actions

Move the assistant from answering to doing: retrieving records, opening requests, triggering workflows.

Customise the experience

Brand the widget, set the welcome message, choose suggested questions, and control appearance down to custom CSS.

Deploy it

Generate an embed script, restrict it to approved domains, and put it on your site.

Monitor usage

Track conversations, credits, response time, resolution rate and the questions that came up most.

Improve it over time

Turn unanswered questions into knowledge gaps, fill them, re-index and re-test. This is the stage that separates a working assistant from an abandoned one.

Platform

Everything you need to put AI to work

Not a wrapper around a chat box — the operational pieces a business actually needs.

Assistants with real grounding
Answers are retrieved from your documents and returned with the source chunk cited, so staff can verify what the model said.
Knowledge bases per assistant
Scope each assistant to the material it should see. Support never reads the finance drive.
Automations that act
Turn an intent into a webhook call, a ticket, a CRM update or an email — chained as steps with run history.
Tools and integrations
Slack, Gmail, Google Calendar, HubSpot, Zendesk and Salesforce, plus a generic webhook connector for anything else.
Analytics and knowledge gaps
See resolution rate, credit burn and token spend — and the questions your assistants could not answer.
Human handoff
A visitor can ask for a person. The assistant pauses itself and the thread lands in your team's inbox.

Purpose-built

One assistant for every part of your business

A support conversation needs accuracy and caution. A sales conversation needs curiosity and momentum. Run multiple assistants in one tenant, each with its own identity, knowledge, model and deployment — sharing your account and credits, but not personalities or permissions.

Customer support

Trained on policies, product documentation, shipping and returns rules, and existing help articles. It handles the repeat questions that consume the first hour of every support shift.

Measured on deflection — questions resolved without a ticket.

Sales

Works the other direction: it opens conversations rather than closing them. Answers product and pricing questions in the moment, and helps qualify interest before a human gets involved.

Different tone, different suggested questions, knowledge weighted to product and pricing.

Product

For technical products whose documentation is excellent and unreadable at speed. Answers configuration and troubleshooting questions conversationally, with a citation back to the exact page.

A user asking about an API limit wants the number and the document it came from.

Internal knowledge

Gives employees a way to find information inside company documentation without interrupting a colleague. New hires ask about process, sales reps ask about product, ops staff ask about procedure.

Internal domains, internal tone, and usually a different set of connected systems.

E-commerce

Retail splits in two: discovery and policy questions, which knowledge alone handles, and order-specific questions, which need a live connection to your store.

The difference between a store assistant that helps and one that annoys is whether it can answer where an order is.

Agency deployments

One tenant environment, many client deployments, each with its own assistant, knowledge, branding, widget, domain, integrations and analytics.

Covered in full further down — it changes the economics of offering AI as a service.

Knowledge & RAG

Answers grounded in your own documents

Retrieval-augmented generation means the assistant quotes your material instead of improvising. Every answer carries the document it came from, so your team can verify it — and so can your customers.

Cited retrieval
Every answer carries the document and chunk it came from.
Vector search per tenant
Embeddings are filtered by tenant at the vector store, not just the database.
Reprocess safely
Re-indexing a document clears its old chunks first, so stale text can't linger in answers.

Supported formats: PDF · DOC · DOCX · CSV · XLS · XLSX · TXT · MD

UploadExtractChunkEmbedIndexReady

Example answer

“International returns are eligible for a full refund within 30 days of receipt, provided items are in original packaging.”

Return-Policy.pdf · chunk 12

Explainer

What is retrieval-augmented generation?

A language model on its own only knows its training data — it has never seen your refund policy or your product catalogue. Retrieval-augmented generation, or RAG, fixes that by looking up the relevant passage from your documents first, then asking the model to answer using it. That is why a NextPro AI assistant can cite a source and why it is far less likely to invent one.

Retrieve
Your question is embedded and matched against the chunks of your documents, scoped to your tenant only.
Augment
The most relevant passages are attached to the prompt as context, alongside the assistant's instructions.
Generate
The model answers from that supplied context, and the source document is returned with the reply.

How knowledge is built

Uploading a file is the first step, not the whole process

Every source runs through the same pipeline before an assistant can use it. Each stage exists for a reason, and chunking is the one that most affects answer quality.

01Upload
The source arrives
A file you upload, a page imported by URL, or an article published inside the platform.
02Validate
Check it is usable
Supported type, within size limits, readable. Catching a corrupted file at upload beats discovering it when a customer gets a bad answer.
03Extract
Pull the text out
A PDF is not text; it is a layout format that happens to contain text. Extraction handles that conversion.
04Process
Clean what came out
Headers, footers, page numbers and formatting noise are handled so they do not pollute retrieval.
05Chunk
Split into passages
The stage that most affects answer quality. Too fine and the assistant loses context; too coarse and retrieval returns three pages when the answer was one sentence.
06Index
Make it searchable by meaning
Someone asking whether they can send an item back reaches your returns policy even though the policy never uses that phrasing.
07Ready
Available to the assistant
The source can now be retrieved in a live conversation, and appears in knowledge usage analytics.

Formats you can upload

PDF: Policies, manuals, contracts, product sheets
DOCX: Internal documentation and process guides
XLSX / XLS: Specifications, price lists, FAQ exports
CSV: Structured data and tabular exports
TXT: Plain-text knowledge and notes
MD: Developer and product documentation

Legacy .doc files need saving as .docx first. You can also import a documentation page by URL — it keeps its source link and can be re-synced when the page changes.

Why knowledge beats model choice

Teams evaluating AI platforms over-index on which model is available and under-index on how content is processed. In practice the second question decides more outcomes.

A strong model retrieving from badly chunked, out-of-date documents produces confident wrong answers. A mid-tier model retrieving from clean, current knowledge produces reliable ones.

Your knowledge pipeline is the thing customers actually experience, which is why it is a visible, inspectable process here rather than a black box you upload into.

Diagnosis

When an answer goes wrong, four stages can be at fault

Nearly every bad answer traces to one of these, and the fix is different for each. Diagnosing which takes about a minute if you look at the retrieved sources alongside the answer — which is why the Test Studio shows both.

Retrieval failure

Symptom: A vague answer, or an admission of ignorance, when you know the information sits in an uploaded document.
Cause: Content never indexed, document phrasing far from customer phrasing, or a knowledge base so large the right passage lost out to a similar one.
Fix: Check knowledge usage to see what was actually retrieved, then tighten scope or rewrite for retrieval.

Source failure

Symptom: A confident answer that is out of date, or belongs to a different product line.
Cause: A superseded document is still in the knowledge base, competing with the current one.
Fix: Remove the old version. This is why citations matter — they make the wrong source visible.

Instruction failure

Symptom: Factually fine, but too long, too casual, off-topic, or willing to speculate outside its scope.
Cause: Retrieval worked; the assistant's configuration did not constrain the response.
Fix: Edit the assistant instructions, not the knowledge base.

Scope failure

Symptom: An answer about pricing negotiation, legal interpretation, or a complaint that needed a person.
Cause: The assistant attempted a question it should have declined.
Fix: Tighten the scope boundary in instructions and configure handoff triggers.

Behaviour you control

Two assistants on the same documents can behave completely differently

Configuration is where that difference gets defined — across identity, behaviour, capability and how the conversation starts.

Identity

Name, description, avatar, purpose

Not decoration. The name and avatar set expectations for the conversation, and the purpose statement shapes how ambiguous questions get interpreted.

Behaviour

Personality, tone, response style, language, instructions

Instructions are where the real control lives: what to do when it does not know, when to offer a human, what it must never promise. Teams who write these thoroughly get noticeably better assistants.

Capability

Model, knowledge sources, tools, integrations, automations

Scope knowledge deliberately. An assistant with access to every document you own retrieves less precisely than one pointed at the twelve that matter.

Entry points

Welcome message, suggested questions

A blank chat box asks the visitor to invent a question. Three well-chosen suggestions tell them what the assistant is good at and produce a first interaction that succeeds.

Test Studio

Validate before your customers do

The riskiest moment in an AI deployment is the first real conversation. A controlled testing environment sits between configuration and launch.

What you can check

AI responses — Against the questions you actually expect to receive
Knowledge retrieval — Whether the right passages are being found at all
Source citations — That answers trace back to the correct documents
Model behaviour — Across different phrasings of the same question
Response latency — A correct answer that takes fifteen seconds still fails
Credit usage — Per interaction — this is how you forecast cost before committing
Instructions — That the assistant does what you actually told it to
Awkward questions — Including the ones you would rather it never received

How to run a useful test round

Testing an assistant with questions you know it can answer proves nothing. A productive round deliberately includes the hard cases.

Your real top questions

Pull the twenty most common queries from your support inbox or search logs and run them verbatim, including the typos and shorthand real customers use.

The edge cases

Questions where the right answer is “I don't know”, questions touching policy exceptions, questions needing two documents combined, and phrasings your documentation never anticipated.

The failure behaviour

Ask something outside the knowledge base entirely. A well-configured assistant declines cleanly and offers a route to a human. One that invents an answer here will invent answers in production.

The cost model

Watch credit consumption while you test. If a typical conversation costs more than expected, adjust the model, instructions or retrieval scope before launch rather than after the first invoice.

Automation

Don't just answer. Resolve.

Connect an assistant to your systems so a conversation ends in an action, not a promise to follow up.

Customer asks

“Where's order #45821?”

Assistant extracts

Order ID + intent

Calls your API

Order status lookup

“Your order is arriving tomorrow.”200 OK · 180ms

Connects with

Slack
Gmail
Google Calendar
HubSpot
Zendesk
Salesforce
Webhooks
REST API

Connect what you run

From answering to doing

Answering “our standard delivery takes three to five business days” is helpful. Answering “your order shipped Tuesday and is due tomorrow” is the reason someone opened the chat.

A customer asks: “Where is my order #45821?”

Understand the request
Identify the relevant record
Call the connected system
Retrieve the result
Generate a response
Return it in conversation

Each step does real work. Understanding the request means recognising intent, not matching the word “order”. Calling the connected system means an authenticated request to whatever holds that data. Generating the response means turning a raw API payload into a sentence a person wants to read.

CRM

Customer records, deal context and lead capture

Helpdesk

Ticket creation, status and escalation

E-commerce

Orders, products, inventory and fulfilment

Calendar

Availability and scheduling

Email & chat

Notifications, follow-up and routing

REST APIs

Internal systems and anything without a prebuilt connector

Webhooks

Event-driven workflows in either direction

Other applications

As required by the deployment

The REST and webhook layer

Prebuilt connectors cover common platforms. They do not cover the internal booking system written in 2019, the regional logistics provider with a bespoke API, or the industry-specific tool serving four hundred customers worldwide. If a system can expose an authenticated endpoint, the assistant can reach it — and webhooks handle the reverse direction. For agencies and technical teams this is often the deciding capability.

Where automation needs guardrails

Read actions and write actions carry different risk. Retrieving an order status is low-stakes. Actions that create, change or cancel something need tight permissions, clear instructions about when to confirm, and handoff for anything involving refunds, cancellations or account changes above a threshold you set. Start read-only, measure on real traffic, then expand.

Embed snippet

<script
  src="https://cdn.nextproai.com/widget.js"
  data-assistant-id="asst_8f2c…"
  data-widget-key="widget_pk_live_…">
</script>

Drop it before </body> on any page.

Deployment

Live on your site in one line

No framework, no build step, no SDK to learn. The widget is a self-contained bundle that works on WordPress, Shopify, Webflow, Next.js or plain HTML.

One script tag
A ready-made loader and widget bundle — no framework or build step on the customer's site.
Signed session tokens
The embed key is public by design; every message is authorised by a short-lived signed token minted per page load.
Domain allowlisting
Name the domains permitted to embed a widget. Anything else is refused.

Website widget

The same assistant, styled three different ways. Colour, theme, shape, copy and behaviour are all configurable — no CSS required, and custom CSS available when you want it.

Northwind SupportTypically replies instantly
Hi there 👋 Ask me anything about orders, shipping or returns.
Where is my order?
Your order #45821 shipped this morning and arrives Thursday.
Type your message…
Powered by NextPro AI
Northwind Support · Light theme · Rounded
Bright HealthPatient help desk
Hello — I can help with appointments, cover and billing questions.
Where is my order?
Yes, an annual check-up is fully covered under your current plan.
Type your message…
Powered by NextPro AI
Bright Health · Light theme · Soft corners
Lumen StudioSales assistant
Hey! Want a quick answer on pricing, plans or onboarding?
Where is my order?
The Growth plan covers 5 assistants and 250k credits a month.
Type your message…
Powered by NextPro AI
Lumen Studio · Dark theme · Soft corners

Example configurations shown. Brand names are illustrative.

Appearance

Brand colour & accent
Light, dark or auto theme
Corner radius & button style
Custom launcher icon
Avatar & assistant name
Custom CSS overrides

Behaviour

Screen position (left or right)
Auto-open with delay
Welcome message & placeholder
Conversation history on/off
File upload on/off
Sound notifications

Control

Allowed domains
Rotate the embed key
Show or hide branding
Language
Per-assistant enable/disable
Live preview before publish

Shadow DOM isolated

The widget renders in a shadow root, so it can neither inherit nor leak styles with your site.

Loads asynchronously

The bundle is fetched after your page, so it never blocks your own content from rendering.

Works on any stack

WordPress, Shopify, Webflow, Next.js or hand-written HTML — one script tag, no build step.

Measurement

AI projects lose budget when nobody can say what they produced

Volume metrics tell you whether people found the assistant. They say nothing about whether it worked. Two numbers do most of the real work: resolution rate, and unanswered questions.

Metric	What it tells you
Conversations & messages	Volume and depth — adoption, not success
Active users	How many people are actually engaging
AI requests & credits	Usage tied directly to cost
Response time	The metric users feel most directly
Resolution rate	Your primary quality signal
Popular questions	What your audience actually asks, versus what you assumed
Knowledge usage	Which sources are earning their place
Widget usage	Where and how the assistant is being opened
Unanswered questions	Your improvement backlog, generated automatically

The improvement loop

Analytics tells you what happened. Knowledge insights tell you what to do about it. A question the assistant repeatedly struggles with becomes a flagged gap.

1Review the question
2Add knowledge or write an article
3Re-index
4Test the assistant
5Confirm the answer improved

Each closed gap improves the assistant, reduces escalations, and produces an article that also serves your help center and organic search. One customer question turns into three assets — and recurring questions the assistant cannot answer are frequently symptoms of a product problem rather than a documentation one.

Human handoff

Know when to step aside

A platform that pretends every conversation should be handled by AI creates worse outcomes than one that admits otherwise. Where enabled, the assistant transfers the conversation with its context, so the customer is not asked to repeat everything they already typed.

Low confidence
The assistant lacks the knowledge to answer reliably
Explicit request
A customer asks for a person — this should always work immediately
Emotional signals
Frustration or distress is evident in the conversation
High-value or high-risk
Refunds, cancellations, complaints, account changes
Repeated failure
The same question asked twice without resolution

An assistant that escalates well earns more trust than one that never escalates. Customers forgive a bot that says “let me get someone who can help with that”. They do not forgive one that loops.

Cost you can see

Usage is metered, visible, and yours to control

Plans and modules define credits, storage, assistant capacity, file capacity and conversation limits. Usage is tracked continuously and you are notified as you approach them — so consumption is something you manage rather than something that happens to you.

Model selection

The largest lever by far. Running simple FAQ lookups on your most capable model is the most common source of unnecessary consumption.

Retrieval scope

An assistant pointed at every document you own pulls more context per request than one pointed at the relevant twenty.

Instruction design

Response length follows instructions, and cost follows length. An assistant told to be concise costs less than one writing three paragraphs for a yes-or-no question.

Build one in minutes — try it here

Define purpose, tone, knowledge and model in one workspace, then test it live before anyone else sees it.

Assistant name

PurposeCustomer support

ToneFriendly, professional

ModelGPT-4o

Knowledge48 docs

StatusActive

ABC Support Assistant

gpt-4o

Do you offer free returns on international orders?

Yes — international returns are eligible for a full refund within 30 days of receipt, provided items are in original packaging.

Return-Policy.pdf

Bring your own model

OpenAI GPT-4o
Anthropic Claude
Google Gemini
OpenRouter

Security & governance

Built for more than one customer

Isolation is enforced on every request, not assumed from the session.

Per-tenant isolation
Documents, vectors, conversations and credits are scoped by tenant and re-checked on every request.
Role-based access
Owner, admin, manager, analyst and viewer roles, with per-workspace permissions.
Audit logging
Who changed what, when, and from which address — queryable per workspace.
IP allowlisting
Restrict workspace access to the networks you name.

For agencies

One platform, many clients

Custom AI development does not scale as a service line — every client becomes a bespoke project with its own infrastructure and its own way of breaking, and margins compress with each new logo. A platform model inverts that.

Recurring rather than project revenue

An AI assistant is not a website build that ends. It needs knowledge maintenance, performance review and continuous improvement — all of which support a retainer.

Results you can demonstrate

Analytics give an agency something creative work rarely produces: hard numbers on conversations handled, questions resolved and gaps closed. Renewal conversations backed by a dashboard go differently.

Faster time to value

A client can be live in days rather than a development cycle, which shortens the distance between signing and the first proof the investment worked.

Each client deployment carries its own

Own assistant and instructionsOwn knowledge baseOwn branding and widgetOwn approved domainsOwn integrationsOwn analytics

The first client deployment establishes the process. The tenth follows the same path with different content. The work shifts from engineering to configuration, which means the people delivering it do not all have to be developers.

Add AI assistants to your service catalogue

Modules

Buy only what you need

Subscription tiers force a familiar compromise: you need one capability from the higher tier and have to buy nine others to get it. Modules are purchasable individually alongside plans.

Advanced RAG

Deeper retrieval capability

Advanced Analytics

Richer reporting across deployments

Automation

Workflow actions in connected systems

Premium Widget

Advanced customisation and branding

Human Handoff

Escalation to your support team

Help Center

Public documentation powered by the same content

Knowledge Insights

Gap detection and improvement prompts

API Access

Programmatic control of the platform

Additional Credits

Higher usage allocation

Additional Storage

Larger knowledge bases

Every capability shows its state

IncludedAvailablePurchasedPending approvalActiveUpgrade required

Customers on modular platforms routinely lose track of what they bought, what is switched on, and what is waiting on something. Making state visible removes a whole category of support ticket.

Approval-based modules

Some modules require review after purchase — appropriate for anything touching sensitive integrations, elevated resource allocation, or configuration that benefits from a conversation first. For the tenant, the state is never ambiguous.

Tenant purchases module
Purchase confirmed
Pending approval
Super Admin reviews
Approved or rejected
Tenant notified
Setup
Active

Super Admin

The commercial model is configuration, not code

NextGenIT Solutions operates the platform itself through a Super Admin environment. Pricing, packaging and model availability are settings, which means the commercial model can respond to the market without waiting for a release cycle.

Plans

Create subscription plans with prices, limits and included features

Modules

Create marketplace modules with pricing, dependencies and availability rules

Marketplace

Control which capabilities are published and how they are presented

Tenants

Manage customer organisations and platform access

Approvals

Review, approve or reject module requests needing manual provisioning

AI providers

Configure providers and models, and which plans may use them

Credits

Configure usage rules and monitor consumption platform-wide

Billing

Subscriptions, purchases, revenue, transactions and payment status

Platform analytics

Usage and activity across every tenant

Security & admin

Permissions, feature flags, notifications, audit logs, system configuration

The tenant journey

Customers are never left wondering what to do next. At any point a tenant can see what they purchased, what they may use, what still needs setup, what is awaiting approval, and how much of their allocation remains.

Choose plan
Purchase modules
Receive approval
Complete setup
Create assistant
Configure AI
Add knowledge
Test assistant
Customise widget
Deploy
Monitor
Improve

Who it is for

The platform is general. How it gets used is not.

How each type of organisation typically starts, what they connect first, and where the value tends to show up.

SaaS companies

Software companies usually have the best documentation and the lowest documentation readership of any industry. An assistant trained on that library recovers value from content that already exists.

Usually starts with: A product assistant on the docs site and in-app help panel, answering configuration and troubleshooting questions with citations.
Connects first: Helpdesk first so handoff keeps context, then CRM so assistant-started sales conversations do not vanish.

E-commerce

Shipping timelines, return windows, sizing, warranty terms, payment methods and order status account for most of what customers ask. The knowledge half is straightforward; the order half needs a live lookup.

Usually starts with: Policy and product questions from documents, then order status once the store integration exists.
Connects first: The commerce platform, because “where is my order” cannot be answered from a document.

Startups & small businesses

For a team without ML engineers, the real choice is rarely between building and buying. It is between a configurable platform and no AI at all.

Usually starts with: One assistant covering product, pricing and support. Split into specialised assistants once volume justifies it.
Connects first: Whatever already holds customer conversations. Cost predictability matters more than breadth here.

Enterprises

Deployments almost always start internally, because the compliance path is shorter when no customer data is involved. Governance features usually decide whether a platform is viable at all.

Usually starts with: An internal knowledge assistant on HR policies, IT procedures and process documentation, while security review proceeds.
Connects first: Identity and internal systems. Tenant isolation, role-based access, controlled model access and audit logging are the evaluation criteria.

Digital & IT agencies

One tenant environment can carry many client deployments. That turns AI assistants into a repeatable service line rather than a series of custom projects.

Usually starts with: The first client deployment establishes the process; the tenth follows the same path with different content.
Connects first: Whatever each client already runs — which is why the REST and webhook layer matters most to this segment.

Education

Institutions handle enormous seasonal question volume against information that changes on a predictable annual cycle: admissions, deadlines, fees, accommodation and financial aid.

Usually starts with: An assistant on the current year's published material, absorbing application-peak volume.
Connects first: Student systems where available. Handoff matters more here than in most industries.

Professional services

Firms in law, accounting and consulting have expertise that is thoroughly documented and almost impossible to search. Precedent files and methodology documents accumulate faster than anyone indexes them.

Usually starts with: Internal knowledge assistants, so a consultant gets the firm's standard approach with citations instead of asking three colleagues.
Connects first: Document management. Client-facing use stays limited to process and intake, with handoff on anything needing judgement.

Real estate

Response speed to an inbound enquiry has direct commercial consequences, and enquiries arrive at all hours. An assistant answering at midnight on a listing page captures interest that would otherwise cool.

Usually starts with: Property details, neighbourhood information, viewing arrangements and process explanations.
Connects first: CRM for lead qualification — gathering requirements, budget and timeline before routing to an agent.

The alternatives

How this AI assistant platform compares to the alternatives

Each solves part of the problem. The differences show up months after the decision rather than during the demo.

The scripted chatbot

Decision trees and keyword matching. Predictable, which is its real advantage, and brittle. Every question outside the script is a dead end, and every business change means rewriting flows. It also cannot handle how people actually phrase things — someone asking to send an item back is asking about returns, but a keyword matcher looking for “return” misses it.

Handles the five questions someone remembered to script.

A general AI chat subscription

Cheap, fast, and genuinely useful for internal drafting. It is not a customer-facing assistant. It has no access to your documents, so asked about your return window it either declines or invents a number belonging to some other company. No branded deployment, no domain restrictions, no analytics, no order lookup, no handoff.

Keep it for internal productivity. It solves a different problem.

Building it yourself

A capable team can build a RAG prototype in a sprint. The prototype is not the project. The project is document parsing that survives real files, chunking that does not split a policy in half, re-indexing, a testing environment, a widget that renders on client sites, domain restrictions, conversation storage, usage metering, cost controls, role-based access, audit logging, analytics, handoff routing, and an admin interface for non-engineers. Then somebody owns it forever.

Right when the assistant is the product you sell. Rarely right otherwise.

Assembling point tools

One product for retrieval, one for the widget, one for analytics, one for automation. Each is good at its job; the problem is the seams. Knowledge lives in one system and conversations in another, so unanswered questions never reach the content team. Costs arrive on four invoices with no combined view.

Most teams consolidate within eighteen months, having learned that integration maintenance is the real cost.

The honest summary

NextPro AI is the right choice when you need a business assistant that works reliably and is maintainable by people who are not engineers. It is the wrong choice if the AI assistant is the product you intend to sell — in which case you should own the stack.

Comparison

Why not a chatbot, or your own build?

Both are valid choices. Here is honestly where each one lands.

Capability comparison between NextPro AI, a generic chatbot, and building in-house
Capability	NextPro AI	Generic chatbot	Build in-house
Answers from your own documents	Built in, with citations	Scripted replies only	You build the RAG pipeline
Triggers real actions in your systems	Tools + webhooks + automations	Rarely	You build every connector
Multi-tenant isolation	Per-tenant by design	Single workspace	Months of work
Usage metering & plan limits	Credits and entitlements built in	Flat seat pricing	You build billing
Human handoff	Assistant pauses, inbox picks up	Third-party add-on	You build routing
Time to first assistant	Same day	Hours	Quarters

Implementation

What rollout actually looks like

Vendors describe setup in minutes. The interface work genuinely is fast; the part that decides whether the assistant is any good is knowledge preparation, and that deserves an honest timeline.

Where does the authoritative answer live for your ten most common questions?

If the answer is “a document someone can find”, you are ready. If it is “in one person's head”, you have content to write before you have an assistant to train.

How current is that content?

An assistant trained on last year's pricing will confidently quote last year's pricing. Outdated knowledge is worse than missing knowledge — missing produces an honest non-answer, outdated produces a wrong one.

Who owns the assistant after launch?

Deployments without a named owner drift. Someone needs to review unanswered questions and decide when instructions need adjusting. Not a full-time role, but somebody's role.

Week 1

Scope and knowledge

Pick one assistant and one job. Trying to serve support, sales and internal teams at once is the most common launch mistake.
Gather source material for that single job — resist uploading the entire shared drive.
Review each document before uploading: remove superseded versions, check dates and figures, confirm no internal-only material is mixed in. This is the most valuable hour in the project.
Upload the reviewed set, or import pages by URL where the content already lives on your site.

Week 2

Configure and test

Set identity, personality, tone, response style, language and instructions.
Write the instructions with failure in mind — what it does when it does not know matters more than what it does when it does.
Run your twenty real support questions in the Test Studio, phrased the way customers phrase them.
Check retrieved sources, not only answers. A right answer from the wrong source will produce a wrong answer later.
Note the credit cost per conversation. That number times expected volume is your forecast.

Week 3

Deployment

Customise the widget and preview across desktop, tablet and mobile before publishing.
Write the welcome message and suggested questions carefully — they are the most underused element in most deployments.
Configure allowed domains including staging, generate the embed code, and deploy to staging first.
Launch narrow. One high-traffic page produces real data with limited exposure; a full-site launch removes your ability to fix problems quietly.

Week 4

Measure and close gaps

Read analytics with intent: conversation count is adoption, resolution rate is usefulness, unanswered questions are the to-do list.
Work the unanswered list in order of frequency.
For each gap, add knowledge, publish an article, or adjust instructions if the failure was behavioural rather than informational.
Re-index and re-test the specific question. That loop is the whole maintenance model.

Month 2+

Automation and expansion

Only connect live systems once the knowledge layer performs well. Automating an unreliable assistant compounds the problem.
Start with read-only actions — order status, account lookup, booking checks — which cover most of the value at a fraction of the risk.
Expand to write actions once you have evidence intent is identified correctly.
Additional assistants usually follow naturally from the same knowledge base.

Preparation

The work that decides everything

The gap between an assistant that impresses people and one that frustrates them is rarely the model. It is the state of the content underneath — which no vendor demo shows and every deployment depends on.

Start with questions, not documents

Write down the thirty questions customers ask most, in their words, from your support inbox rather than your site navigation. Then find where each authoritative answer lives. That exercise produces your document list, your list of answers that exist only in someone's head, and your test set.

Curate ruthlessly before uploading

Every document competes with every other for retrieval. Thirty accurate current files beat two hundred mixed-quality ones. Remove superseded versions first — old pricing sheets and previous policy revisions are the single most common cause of confidently wrong answers.

Separate internal from customer-facing

Internal process notes, margin information and candid competitor assessments do not belong in a knowledge base powering a public widget. This sounds obvious and is a recurring incident category.

Write for retrieval, not only reading

Keep passages self-contained — a paragraph beginning “as noted above” is meaningless retrieved alone. Use the words customers use at least once. Answer the question early, then qualify. Break long documents into focused sections so chunking finds clean boundaries.

Maintain on a schedule, not a crisis

Tie review to a trigger rather than a date: any time a policy changes, updating knowledge and re-indexing belongs in the same task as updating the website. Use knowledge usage analytics to prune sources nothing ever retrieves.

Instructions

Designing how your assistant behaves

Knowledge determines what the assistant knows. Instructions determine what it does with that knowledge, and they receive far less attention than they deserve.

Purpose before personality

Define scope first: what this assistant exists to do and, more importantly, what it does not. Boundaries keep it out of pricing negotiation, legal interpretation, medical or financial guidance, and competitor commentary — the four areas where a wrong answer costs most. Write the boundary explicitly; an assistant told what it handles will still attempt adjacent questions unless told what to decline.

Configure the failure case first

The default behaviour of a model asked something outside its context is to produce a plausible answer. Instructing it to acknowledge the gap, avoid guessing and offer a route out changes this materially. It is the highest-return line of configuration available.

Tone that fits the job

A support assistant meets people who are already frustrated — efficiency reads as respect, and warmth reads as evasion. A sales assistant meets people who are curious, so conversational range works. An internal assistant meets colleagues who want an answer: brevity wins and jargon is an advantage.

Welcome messages do real work

“Hi, how can I help?” tells a visitor nothing and produces silence or an out-of-scope question. “I can help with orders, shipping, returns and product questions” sets accurate expectations before anyone types.

Test the behaviour, not only the answers

Ask something out of scope and confirm it declines. Ask something ambiguous and see whether it clarifies or guesses. Ask the same question three ways and check consistency. Behavioural failures are easier to fix than knowledge failures, and easier to miss.

Evaluation

What to ask any AI assistant vendor

This list is deliberately not written to favour NextPro AI — it is written to help you evaluate properly, including evaluating us. Bring it to every demo you take.

Knowledge and accuracy

Show me the sources behind an answer, live, right now.
Demonstrate what happens when the assistant does not know — using a question your demo data does not cover.
How does content get updated, and how long until the update reaches the assistant?
What happens when two documents disagree?
Which file types work badly, not just which are supported?

Testing and validation

Can I test before deploying, and does the test show retrieved sources, latency and cost — or only the answer?
Walk me through validating an assistant against fifty of my own questions before launch.

Deployment and control

What stops my embed code working on a domain I did not authorise?
Show me the widget on mobile rather than telling me about it.
How much can I change without custom development, and what needs custom CSS?
Can I run multiple assistants, and what is the limit?

Automation and integrations

Which integrations exist today versus on the roadmap?
How does a system you have never heard of get connected?
What is the assistant allowed to do without confirmation?
What happens when an integration fails mid-conversation?

Cost

Exactly what consumes usage, and what does not?
Help me estimate monthly cost for my projected volume before I sign.
What happens when I exceed my allocation — stop, degrade, or overage billing?
Which configuration choices increase consumption?

Measurement

Show me a real dashboard, not a feature list — specifically the unanswered-question report.
How would I identify my assistant's biggest weakness after a month of live traffic?
Are analytics available per assistant and per deployment?

Security and governance

How does tenant isolation work? What is stored, where, and for how long?
Who at the vendor can see my conversations?
What role-based access exists inside my own account?
Is there audit logging, and which models am I able to control?

The exit

How would I leave? Can I export knowledge, articles and conversation history, and in what format?
This question makes vendors uncomfortable, which is exactly why it is worth asking.

How NextPro AI answers these

Citations are visible in the interface rather than on request. The Test Studio shows retrieved sources, latency and credit cost alongside every test response. Domain restrictions are configurable per assistant. Credits make consumption visible before and after launch. Analytics include the unanswered-question report. Tenant isolation, role-based access, controlled model access and audit logging are architectural rather than optional. Multiple providers are supported so model choice stays a decision rather than an inheritance. Where a capability depends on your plan or a purchased module, this page says so rather than implying everything is included at every tier.

Return

Measuring what the assistant produced

The most common measurement mistake is deploying first and looking for a baseline afterwards. Once the assistant is live, the pre-launch numbers are gone.

Record these before you launch

Support conversations per week, across every channel
Average time to resolve one
Share of them that are repeat questions already covered by documentation
First-response time outside business hours

Without them, every claim about the assistant's impact is an argument rather than a measurement.

Building the internal case

The financial argument usually rests on three components, and it is more credible when you present all three rather than only the flattering one.

Deflected volume

Conversations resolved without becoming tickets, multiplied by your own loaded cost per ticket. Use your number, not an industry figure — published figures vary by an order of magnitude and none describe your team.

Coverage expansion

The assistant answers at 2am, at weekends, and during seasonal spikes when hiring is not an option. For e-commerce and real estate, the revenue consequence of an unanswered out-of-hours enquiry is the real number, not the support saving.

Team redeployment

Staff who stop answering the same shipping question forty times a week are available for conversations needing judgement. A quality argument rather than a headcount one, and usually more persuasive to leadership.

Against those, put the honest costs: subscription, credits at projected volume, purchased modules, the internal hours spent preparing knowledge, and the ongoing hours maintaining it. A business case that omits the maintenance hours will be wrong within a quarter.

Weeks 1–8

Weekly review focused almost entirely on the unanswered-question list. Operational, not reportable.

Month 3+

Monthly reporting: volume, resolution rate, top questions, gaps closed, credits against allocation.

Quarterly

Review against the pre-launch baseline. This is where the project earns its next year of budget.

Straight answers

Common concerns, answered directly

These come up in most evaluations. They deserve straight answers rather than reassurance.

“It will make something up and we will look incompetent.”

This is the correct thing to worry about, and no platform can promise it never happens. Retrieval changes the odds substantially: the assistant answers from passages retrieved out of your content rather than from general training data. Four controls stack on top — citations make answers checkable, instructions define what happens when it does not know, pre-launch testing catches failure modes specific to your content, and handoff provides an exit. The residual risk is real: scope narrowly, test the edges, monitor unanswered questions, and keep the assistant away from topics where a wrong answer carries legal or safety consequences.

“Our documentation is a mess.”

Nearly everyone's is. You do not need to fix your documentation to start — you need to fix the part covering your ten most frequent questions. A narrow assistant answering ten things well is a better first deployment than a broad one answering forty unevenly, and it produces cleaner data about what to add next. The unanswered-question report then does the prioritisation for you.

“We will set it up and nobody will use it.”

Adoption failure usually has one of three causes, all addressable at launch. Placement: an assistant buried on a contact page gets contact-page traffic — put it where the questions occur. Expectations: users do not know what it can do, which is what suggested questions solve. Early quality: a user whose first question fails does not come back, which is the argument for testing thoroughly and launching narrow.

“We cannot predict the cost.”

Credit-based usage exists for this reason. Every AI request consumes credits, consumption is tracked continuously, and notifications fire as you approach configured limits. The forecast is available before launch: test a representative set of conversations, note the credit cost per conversation, multiply by expected volume. Costs also move with configuration — retrieval depth and model choice are yours to control.

“Our data will end up mixed with someone else's.”

Tenant isolation is architectural, not a policy applied afterwards. Users, assistants, knowledge, documents, conversations, widgets, domains, integrations, automations, analytics, credits and billing are all scoped to your tenant and re-checked on every request. Role-based access controls who inside your organisation can reach what, domain restrictions control where the widget runs, and audit logging records administrative activity.

“We will be locked into one AI vendor.”

The platform supports multiple providers and models rather than being tied to one, which matters because model pricing and capability change on a timescale of months. More practically, your investment is not in the model — it is in your knowledge base, integrations, widget configuration, instructions and accumulated analytics. Those persist across model changes.

“Nobody on our team knows how to build AI systems.”

The core workflow is configuration, not development: creating an assistant, uploading knowledge, setting behaviour, testing, customising the widget and deploying a script tag are interface tasks. Developers become useful for custom REST integrations, webhook workflows and advanced CSS. The skill that actually determines success is not technical — it is knowing your customers' questions and your business's answers well enough to scope the assistant properly.

Pricing

Plans that scale with you

14-day free trial on every plan. Upgrade as your usage grows.

Starter

For a first assistant on a single site.

$29 /month

1 AI assistant
50,000 monthly credits
1 GB knowledge storage
Website chat widget
Email support

Get started

Growth

For teams running assistants across several channels.

$79 /month

5 AI assistants
250,000 monthly credits
10 GB knowledge storage
Automations & webhooks
Custom widget styling
Analytics & knowledge gaps

Start free trial

Enterprise

For platforms reselling assistants to their own customers.

$199 /month

Unlimited assistants
1,000,000 monthly credits
100 GB knowledge storage
White-label branding
SSO & IP allowlisting
Priority support

Contact sales

Only need one extra capability? Browse the module marketplace

FAQ

Frequently asked questions

NextPro AI is a multi-tenant platform for building, deploying and managing AI assistants trained on your own business documents. It combines retrieval-augmented generation, workflow automation, an embeddable website widget, usage metering and analytics in one product.

You upload documents or point it at your website. NextPro AI parses, chunks and embeds that content into a vector store scoped to your tenant. At question time the assistant retrieves the most relevant chunks and answers from them, citing the source document.

PDF, DOCX, XLSX, XLS, CSV, TXT and Markdown. Legacy .doc files need saving as .docx first. Storage limits and per-file size caps follow your subscription plan.

Assistants can be pointed at OpenAI, Anthropic, Google or OpenRouter models, configured per assistant, so you can match cost and capability to the job.

Yes. You get a ready-made loader script and an embed snippet that works on any site. Each visitor session authenticates with a short-lived signed token, and you can restrict which domains are allowed to embed the widget.

Every document, embedding, conversation and credit balance is scoped by tenant at both the database and the vector store, and that scope is re-checked on every request rather than trusted from the client.

In credits, deducted per message based on the model and token volume. Dashboard chat and embedded widget messages meter through exactly the same path, so your usage reporting stays consistent.

Yes. A visitor can request a human, which pauses the assistant and routes the thread to your team's inbox where an agent can reply, resolve or reopen it.

Copy the embed snippet from your assistant's widget settings and paste it before the closing body tag of your site. It is a single script tag with your assistant ID and public widget key, and it works on WordPress, Shopify, Webflow, Next.js or plain HTML with no build step.

Yes. You can set the brand colour, light or dark theme, corner style, launcher icon, avatar, assistant name, welcome message, input placeholder, screen position and auto-open delay, plus custom CSS for anything else. A live preview shows your changes before you publish.

Yes — it is public by design, like a publishable API key. It only identifies which widget to load. Actual messages are authorised by a short-lived signed session token minted per page load, and you can restrict which domains are allowed to embed the widget as well as rotate the key at any time.

No. The loader is a small self-contained script that loads asynchronously after your page, renders inside a shadow DOM so it cannot inherit or leak CSS, and pulls no framework onto your site.

In detail

Everything else you are likely to ask

Getting started4

How long does it take to get an assistant live?: A basic assistant with uploaded knowledge and a deployed widget can be running quickly, because the work is configuration rather than development. The variable is knowledge preparation. Businesses with current, well-organised documentation move fastest; businesses whose policies are scattered across old files spend most of their setup time on content rather than on the platform.
Do we need developers to use NextPro AI?: Not for the core workflow. Creating assistants, uploading knowledge, configuring behaviour, customising the widget and deploying with an embed script all happen through the interface. Developers become useful for custom REST API integrations, webhook workflows and advanced widget styling.
What should our first assistant do?: Pick the single job with the highest volume of repetitive questions, which for most businesses is customer support. Narrow scope produces better answers, a faster launch, and cleaner data about what to improve. Expanding into sales, product or internal assistants is straightforward once the first one works.
How much content do we need before starting?: Less than most teams assume. Coverage of your ten to twenty most frequent questions is enough for a useful first deployment. Volume matters far less than accuracy and currency.

Knowledge and accuracy9

What file types can we upload as knowledge?: PDF, DOCX, XLSX, XLS, CSV, TXT and Markdown. Legacy .doc files need saving as .docx first. Knowledge can also come from pages imported by URL and from articles published inside the platform.
How does the assistant avoid making things up?: Retrieval-augmented generation grounds responses in your own content rather than the model's general knowledge, and source citations let you verify where an answer came from. Assistant instructions define behaviour when it does not know, and human handoff provides a route out of conversations it should not attempt. No system removes the risk entirely, which is why pre-launch testing and post-launch monitoring both matter.
What happens when the assistant does not know something?: That depends on how you configure it, which is why the instruction covering unknown questions is one of the most important settings on the page. A well-configured assistant acknowledges the gap rather than guessing, and where handoff is enabled it can route the conversation to a person.
Can we see which source an answer came from?: Yes, where citations are enabled. Responses can display the supporting sources — the specific document or article the answer was drawn from. This is what makes answer quality auditable rather than a matter of trust.
How do we keep the assistant accurate as our business changes?: Update the underlying knowledge and re-index. Imported pages can be re-synced on demand from the document itself. Knowledge insights surface the questions the assistant is failing, so you know which gaps to close first.
What happens if we upload conflicting documents?: The assistant may retrieve either version, which is exactly the problem it sounds like. Removing superseded documents matters more than adding new ones — outdated content produces confidently wrong answers, which are worse than honest non-answers.
Can the assistant use content from our website?: Yes. You can import a page by URL into a knowledge base, and it runs through the same validate, extract, process, chunk and index pipeline as an uploaded file. Imported pages keep their source URL and can be re-synced when the page changes.
Does uploading more documents make the assistant smarter?: Not reliably. Beyond a point, additional documents mainly increase the chance of retrieving the wrong passage. Precision beats volume, and knowledge usage analytics show which sources are earning their place.
Can we use the same knowledge for our help center and our assistant?: Yes. Articles, FAQs, policies and documentation published in the platform can serve a customer-facing help center and act as assistant knowledge at the same time, so one update maintains both.

Models and technology5

Can we choose our own AI model or provider?: NextPro AI supports multiple providers and models, configured at the platform level. Tenants select from the models their subscription and purchased capabilities permit.
Why can't we see every model available?: Model availability is controlled deliberately. Without that control, model selection becomes an unmanaged cost centre where anyone can pick the most expensive option for a task that did not need it.
Does a better model fix a bad answer?: Usually not. If the assistant received the wrong context, a more capable model writes a more articulate wrong answer. Retrieval quality is the first thing to check when answers disappoint; model selection is the second.
What is RAG, in plain terms?: Retrieval-augmented generation means the system searches your knowledge for passages relevant to the question, then asks the model to answer using those passages. The model contributes language ability; your content contributes the facts.
What does re-indexing mean and when do we need it?: Indexing is the processing step that makes your content searchable by the assistant. Re-indexing repeats it after content changes, so updated documents actually reach the assistant. Any time you revise, add or remove knowledge, re-index and then re-test the affected questions.

Deployment and the widget4

Can we control where the widget is allowed to run?: Yes. Domain-based deployment lets you specify which domains an assistant may run on, which prevents an embed configuration being reused on unauthorised sites and keeps staging separate from production.
Does the widget work on mobile?: Yes, across desktop, laptop, tablet and mobile, and you can preview each before deploying. Mobile preview is worth taking seriously, since placement that works on a wide screen frequently obscures important controls on a phone.
Can we deploy the same assistant in more than one place?: Yes, subject to your domain configuration. A single assistant can serve several approved domains, which suits businesses running a marketing site, an app subdomain and a documentation site.
Can we run more than one assistant?: Yes, subject to the assistant capacity of your plan. Separate assistants for support, sales and internal knowledge is the recommended pattern, since each can have its own personality, knowledge scope, model, integrations and deployment.

Automation and integrations4

What can an assistant do beyond answering?: With automation enabled and an integration connected, an assistant can identify a request, call a connected system, retrieve information such as an order status, and return the result in conversation. Depending on the integration it may also create requests, retrieve records, schedule actions or trigger workflows.
Do we need automation on day one?: No, and starting with it is usually a mistake. Get the knowledge layer answering reliably first. Connecting live systems to an assistant whose answers are still inconsistent multiplies the problem rather than solving it.
Is it safe to let an assistant take actions?: Start with read-only actions such as retrieving an order or checking a booking, which cover a large share of the value at a fraction of the risk. Move to actions that create or modify records once you have evidence the assistant identifies intent correctly.
When should handoff trigger?: Configure it for situations where a wrong answer is expensive: complaints, billing disputes, cancellations, anything with legal or compliance weight, and repeated failed attempts in one conversation. An explicit request to speak to a person should always route through.

Analytics and improvement4

What can we measure?: Conversations, messages, active users, AI requests, credits consumed, response time, resolution rate, popular questions, knowledge usage, widget usage and unanswered questions.
What is a knowledge gap?: A recurring question the assistant repeatedly fails to answer well. You review the question, add knowledge or publish an article, re-index, test, and confirm the improvement. That cycle is the core maintenance loop.
How often should we review performance?: Weekly for the first two months, focused on unanswered questions, then monthly once the obvious gaps are closed. Quarterly, compare against the baseline you recorded before launch.
What is a good resolution rate?: There is no universal number, because it depends entirely on the mix of questions your assistant receives. Treat it as a trend line for your own deployment rather than a benchmark against other companies.

Plans, credits and modules5

What are credits and what consumes them?: Credits meter AI usage. Requests to the assistant consume them, and consumption varies with configuration — including how much knowledge is retrieved and which model is used. Usage is tracked continuously and shown against your allocation.
What happens when we run out of credits?: Usage is tracked continuously and you are notified as you approach configured limits. Additional credits and plan upgrades are available according to the platform's commercial model. Testing before launch gives you a forecast in advance rather than a surprise afterwards.
What is the difference between a plan and a module?: A plan sets your baseline capacity and included features. Modules are individual capabilities purchased separately — advanced retrieval, advanced analytics, automation, premium widget, human handoff, help center, knowledge insights, API access, additional credits or additional storage. You pay for what you need rather than a tier bundle you partly use.
Why does a module say “Pending approval”?: Some modules require platform review after purchase. The flow is purchase, confirmation, pending approval, review, approval or rejection, notification, setup, then active. This exists for capabilities needing review or manual provisioning.
Can we upgrade later?: Yes. The plan and module structure is designed for capabilities to be added as requirements grow, rather than requiring you to choose your eventual configuration on day one.

Security and administration2

Who inside our company can access what?: Role-based access controls permissions within your tenant, so administrators, editors and support users get appropriate levels of access to assistants, knowledge, integrations and analytics.
What administrative controls exist at the platform level?: NextGenIT Solutions manages plans, modules, marketplace availability, tenants, module approvals, AI providers, credit rules, billing, platform analytics, permissions, feature flags, notifications, audit logs and system configuration through a Super Admin environment.

For agencies3

Can one account manage assistants for multiple clients?: Yes — that is a core design goal. One tenant environment can carry multiple deployments, each with its own assistant, knowledge, branding, widget, domain, integrations and analytics.
Can each client get their own branding?: Yes. Widget branding, colours, avatar and conversational experience are configured per deployment, so each client's assistant looks like their product rather than yours.
Can we report to clients individually?: Yes. Analytics are available per deployment, which supports per-client reporting on conversations, resolution, popular questions and knowledge gaps.

Glossary

The vocabulary, in plain language

Buying an AI platform means reading a lot of terms vendors rarely define. Here is what each one means in practice.

AI assistant: A configured conversational system with its own identity, behaviour, knowledge, model, capabilities, integrations and deployment settings. One business can run several, each doing a different job.
Assistant instructions: The written guidance shaping how an assistant behaves: what it is for, how it responds, what it refuses, and what to do when it does not know. The most consequential text in most configurations.
Audit logging: A record of administrative activity, used to answer who changed what and when.
Chunking: Splitting a document into passages small enough to retrieve precisely. Chunking quality is why one platform gives a clean answer from a policy and another returns half a sentence.
Credits: The usage unit metering AI consumption. Requests consume credits, consumption varies with model and retrieval configuration, and the remaining balance is visible against your allocation.
Embed code: The script you add to your website to run the assistant widget. Generated after configuration; no backend development required.
Human handoff: Transferring a conversation from the assistant to a person, with context preserved, so the customer does not start over.
Indexing: The processing step making content searchable by the assistant. Content is not usable until indexed, and updated content is not usable until re-indexed.
Knowledge base: The body of content an assistant can draw on: uploaded files, imported pages and articles published in the platform.
Knowledge gap: A recurring question the assistant repeatedly fails to answer. Surfaced so it can be closed deliberately rather than discovered through complaints.
Module: An individually purchasable capability, separate from the base subscription.
Multi-tenant: An architecture where each customer operates in an isolated environment on shared infrastructure.
RAG: Retrieval-augmented generation: searching your knowledge for relevant passages and having the model answer from those rather than from general training data.
Resolution rate: The share of conversations completed without needing a person. Useful as a trend for your own deployment, misleading as a benchmark against other companies.
Role-based access: Permissions assigned by role rather than individually, so administrators, editors and support users each see what they should.
Source citation: Displaying which document or article an answer came from, turning answer quality into something checkable rather than something to trust.
Super Admin: The platform-level environment through which NextGenIT Solutions manages plans, modules, tenants, approvals, providers, credits and billing.
Tenant: One customer's isolated environment within the platform. An agency's tenant may contain many client deployments.
Test Studio: The pre-launch environment for evaluating responses, retrieval, citations, model behaviour, latency and credit usage before customers reach the assistant.
Unanswered question: A logged instance of the assistant failing to answer. Individually a failure; collectively the most useful improvement roadmap you will get.
Widget: The branded interface the assistant runs in on your website, configurable for appearance, position, dimensions and conversational experience.

Put your documents to work today

Start free, connect your first knowledge base, and have an assistant answering real questions before the end of the day.

Start building free Sign in

The AI assistant platform that knows
your business, not just AI.

No infrastructure to run. 14-day free trial, no card required.

Customer support

Sales enablement

Internal helpdesk

Operations

Finance

Answers grounded in your own documents

Cited retrieval

Every answer carries the document and chunk it came from.

Vector search per tenant

Embeddings are filtered by tenant at the vector store, not just the database.

Reprocess safely

Re-indexing a document clears its old chunks first, so stale text can't linger in answers.

Supported formats: PDF · DOC · DOCX · CSV · XLS · XLSX · TXT · MD

What is retrieval-augmented generation?

Live on your site in one line

No framework, no build step, no SDK to learn. The widget is a self-contained bundle that works on WordPress, Shopify, Webflow, Next.js or plain HTML.

One script tag

A ready-made loader and widget bundle — no framework or build step on the customer's site.

Signed session tokens

The embed key is public by design; every message is authorised by a short-lived signed token minted per page load.

Domain allowlisting

Name the domains permitted to embed a widget. Anything else is refused.

Metric

What it tells you

Conversations & messages

Volume and depth — adoption, not success

Active users

How many people are actually engaging

AI requests & credits

Usage tied directly to cost

Response time

The metric users feel most directly

Resolution rate

Your primary quality signal

The AI assistant platform that knows your business, not just AI.

Most business AI stops at the conversation

From documents to a working assistant

Connect your knowledge

Shape the assistant

Test before you ship

Deploy anywhere

A look inside the platform

Configure and preview side by side

A shared inbox for the hard ones

See what it can't answer yet

Run it for many customers

The AI assistant platform lifecycle, in nine stages

Create an assistant

Add business knowledge

Configure the model

Connect tools

Automate actions

Customise the experience

Deploy it

Monitor usage

Improve it over time

Everything you need to put AI to work

Assistants with real grounding

Knowledge bases per assistant

Automations that act

Tools and integrations

Analytics and knowledge gaps

Human handoff

One assistant for every part of your business

Customer support

Sales

Product

Internal knowledge

E-commerce

Agency deployments

Answers grounded in your own documents

Cited retrieval

Vector search per tenant

Reprocess safely

What is retrieval-augmented generation?

Retrieve

Augment

Generate

Uploading a file is the first step, not the whole process

Formats you can upload

Why knowledge beats model choice

When an answer goes wrong, four stages can be at fault

Retrieval failure

Source failure

Instruction failure

Scope failure

Two assistants on the same documents can behave completely differently

Identity

Behaviour

Capability

Entry points

Validate before your customers do

What you can check

How to run a useful test round

Your real top questions

The edge cases

The failure behaviour

The cost model

Don't just answer. Resolve.

From answering to doing

CRM

Helpdesk

E-commerce

Calendar

Email & chat

REST APIs

Webhooks

Other applications

The REST and webhook layer

Where automation needs guardrails

Live on your site in one line

One script tag

Signed session tokens

Domain allowlisting

The AI assistant platform that knows
your business, not just AI.