
Most AI assistants on websites answer every question. That is the problem. An assistant that always has an answer will sooner or later give one your business never wrote.
The short answer: an AI assistant that answers from your own documents looks up what you gave it and answers only from that, lists the documents each answer came from, and says so when your documents don't cover a question. Judge any document assistant by four things: where its answers come from, whether it shows sources, what it does when it doesn't know, and what a visitor can talk it into.
Ask, part of our Tomorrow Central platform, is a chat on your website that answers visitors' questions about your products and services from your own content, shows the sources behind each answer, and says so when your content doesn't cover a question.
This post is for founders and heads of marketing, sales, e-commerce, customer experience and support at small and mid-sized businesses: what goes wrong with generic AI assistants, what to expect instead, a checklist for judging any of them, and how Ask works.
Three things, and they come from the same cause.
They make up answers. A general AI model is built to produce a fluent reply. Ask it about your returns window and it will give one, whether or not it found yours. The OWASP Gen AI Security Project lists this as a risk in its own right, misinformation: a model produces false or misleading text that appears credible. It names hallucination, content that seems accurate but is fabricated, as one of the major causes.
They show no sources. The visitor gets a confident paragraph and no way to check it. Your support team gets the follow-up ticket.
They say things the business never said. A general AI chat tool drawing on general knowledge will discuss your competitors, offer opinions or describe a policy you never had, on your website, under your name.
The cause is the same in each case: nothing ties the answer to your own material, and nothing checks it before the visitor sees it.
The fix has a name. Retrieval-augmented generation means the assistant first looks up the relevant passages in a trusted set of documents, then writes its answer from those passages rather than from memory. It is the first remedy on OWASP's list. A document assistant is retrieval-augmented generation on your own material, plus two things the technique alone does not promise: sources the visitor can open, and a check that the answer matches them.
Two more come from how Ask handles the gaps:
Product and service questions. What a product does, what a plan includes, which option suits which need, all answered from your own product pages and guides.
Policies. Returns, delivery, cancellations, warranties. These are the questions where a made-up answer costs most, and where a linked source matters most.
Documentation and help articles. Guides, how-tos and help pages that are hard to search one by one. The chat finds the right one and links to it.
Pre-sales questions. What a prospect asks before booking a call: what you cover, how you work, what is included, with a link to the page a salesperson would have sent.
Knowledge for your own people. For content only your own people should read, a chat can be limited to people in your organisation signed in to Tomorrow Central, with the sign-in checked on every message. This suits sites where your people already sign in with Tomorrow Central.
Ask runs on Tomorrow Central, our platform of business software. We switch it on for your organisation and help with the first set-up.
What goes in. Ask reads a file you upload: a spreadsheet-style file (CSV, or JSON Lines) with one row for each product, service, policy or help article. Each row holds a short reference, the item's title, its web link and its text, and it can split the text into sections. Typical sources are:
Your team can prepare the file, or we can prepare it for you. A file Ask can't read comes back with the reason. Ask does not crawl your website or read PDF or Word files, so that content goes into the file first.
Who does what.
Nothing is public until you upload content and put the chat on a website you list. The snippet holds nothing secret.
When a visitor asks a question, Ask looks it up in your content, writes an answer from what it found, checks that answer against its sources, and shows it with a "Sources" list of links underneath.
Keeping it current. Upload new files or remove old ones at any time. Each file shows whether it is waiting, being read, ready or failed. Remove a file and the chat no longer answers from it. To change a single item, upload its row again with the same reference.
Test this hardest in any AI assistant. Ask has three honest replies short of a full answer:
A chat that says "I don't know" a few times a day is doing its job. Each one points to a gap your content could fill.
Every AI agent on our platform is secure by design: it sees only what you allow, and changes only what your rules and your people approve. For Ask, that is narrow: it reads the content you chose, answers from them and does nothing else.
On top of those:
One thing to decide before you upload: if anyone can chat, anything you upload can be found by visitors on the websites where you put the chat. Upload only what you would publish, or limit the chat to signed-in people.
For the wider rules we apply to every AI agent on our platform, read how we secure AI agents.
Use this on any product, including ours. Ask the vendor each question, then try it yourself.
| What to check | Question to ask | A good answer | A warning sign |
|---|---|---|---|
| Where answers come from | Does it ever answer from general knowledge? | Never. Only from the documents you gave it | "It combines your content with its own knowledge" |
| Sources | Does every answer show where it came from? | Every answer lists its sources with links a visitor can open | Sources are optional, hidden or only in an admin view |
| Checking | Is an answer checked against its sources before it is shown? | Yes, and an answer that fails is not shown | "The model is instructed to be accurate" |
| Not knowing | What happens when your documents don't cover a question? | It says so plainly | It always produces an answer |
| Off-topic questions | Will it write a poem or discuss a competitor? | It declines anything outside your documents | It answers anything, on your website |
| Rules | Can a visitor talk it out of its rules? | Fixed checks outside the AI refuse the attempt | The only defence is the AI's own instructions |
| Separation | Are my documents kept apart from other customers'? | Stored and searched separately per organisation | A vague "your data is secure" |
| Where it runs | Can someone copy the snippet onto another site? | It answers only on the websites you list | The snippet works anywhere it is pasted |
| Actions | Can it change anything? | It only reads, or every change needs approval | It can act with no approval step |
| Who can chat | Can I limit some documents to my own people? | A signed-in mode, checked on every message | Everything uploaded is visible to everyone |
| Updates | How do I change or remove a document? | Replace or remove it, and answers follow | Old content keeps appearing in answers |
| Measurement | Can I see what it couldn't answer? | Answered and unanswered counts you can act on | Only a total number of chats |
Run the test with twenty real questions from your inbox: ten your documents answer, five they nearly answer and five they don't answer at all. A good document assistant answers the first ten with sources, is careful with the next five and says "I don't know" to the last five.
Ask is not the right fit if you need it to read your website, PDFs or Word files directly, to answer in languages other than English, to take actions such as bookings or refunds, or to hand visitors to a person. Use a tool built for that.
A chat on your website that looks up the documents you gave it and answers only from them, instead of from general knowledge. A good one lists the documents each answer came from and says so when your documents don't cover a question.
A general AI model is built to produce a fluent reply, not to stay inside your documents. If nothing forces it to answer only from your material and to check its answer against that material, it fills gaps with text that sounds right.
Only the content you upload: a file with one row per product, service, policy or help article, holding its title, its web link and its text. It does not crawl your website or read PDF or Word files, so that content goes into the file first. Your team can prepare the file, or we can.
It says it couldn't find an answer in your content, instead of guessing. If it found something but couldn't check its answer against it, it shows the sources it found instead of an answer.
If anyone can chat, anything you upload can be found by visitors on the websites where you put the chat. For content only your own people should read, choose signed-in people only.
No. Ask only reads your content and answers. It cannot change your website, your systems or anything else.
English only. Upload English content and expect English answers.
Usually within 1 to 4 days, depending on how ready your content is. We help you prepare the file, set up the chat and add it to your website.
Get in touch and we will talk you through pricing for your organisation.
If your customers ask the same questions every day and the answers are already in your catalogue, policies or help pages, Ask can answer them on your website, with the sources, so more visitors get what they need to buy. Talk to us and we will switch it on for your organisation, help prepare your content and get you started. You can also read more about Ask on Tomorrow Central.
If you need a document assistant that does what Ask doesn't, such as reading other file types or working inside your own systems, our AI and automation services team builds them to order.
ITMTB works with enterprises on AI, cloud, security, and digital transformation — from strategy through deployment.