HOME / AI CHATBOT ACCURACY
The accuracy system

The AI chatbot that
doesn't make things up.

The uncomfortable truth about AI chatbots: most of them will confidently invent your prices, your opening hours and your refund policy. HyperChat is built the other way round — it exams itself before launch, refuses to guess at runtime, and learns from every question it can't answer.

Try HyperChat free · 60s See the pre-flight exam
THE SHORT ANSWER

No AI chatbot can promise perfection — but HyperChat is the one that measures itself and shows you the results. Every build sits a pre-flight exam (ten questions from your own site, judged for groundedness, plus five adversarial probes) and your portal shows the score — "your bot passed 9/10 checks" — before launch. At runtime, injection deflection and a low-confidence gate mean the bot says "I don't know" and captures the lead instead of guessing. £49/month, all included.

01 // The problem

A wrong answer costs more
than a missed one.

Large language models are fluent before they are truthful. Given half a chance, a typical chatbot will fill the gaps in its knowledge with confident fiction.

THE INVENTED PRICE

A visitor asks what a job costs. The bot quotes a plausible-sounding figure it found nowhere on your site. Now you're honouring a price you never set — or arguing with a customer who screenshots the chat.

THE INVENTED HOURS

"Yes, we're open Bank Holiday Monday." You're not. The visitor makes the trip, finds the shutters down, and tells Google about it. One hallucination becomes a one-star review.

THE INVENTED POLICY

"Of course, full refunds any time." Airlines and retailers have already been held to promises their chatbots fabricated. A small business can't afford the same headline.

Every one of these failures has the same root cause: the bot guessed when it should have said it didn't know. HyperChat treats that as an engineering problem, not a disclaimer.

02 // The pre-flight exam

Your bot sits an exam
before it meets a customer.

Most chatbot platforms hand you a bot and wish you luck. HyperChat marks its homework first — after every knowledge-base build, automatically.

HOW IT'S MARKED
Questions generated from your own website~10
Answered by the real bot, not a mockEVERY BUILD
Scored for groundedness by a judge modelPASS / FAIL
Adversarial probes fired at the bot5

INJECTION · JAILBREAK · PRICE-FISHING · OFF-TOPIC · POLICY-FISHING

IN YOUR PORTAL
WHAT YOU SEE
9/10
"YOUR BOT PASSED 9/10 CHECKS"

A plain-English score, before a single visitor chats. Failed checks show exactly which question tripped the bot, so you fix the knowledge — not guess at it.

The exam runs again every time you rebuild or retrain, so a website update can never quietly break your bot's answers.

03 // Runtime guardrails

After launch, the bot that
knows when to say no.

The exam proves the bot at build time. Three guardrails keep it honest on every live conversation afterwards:

Injection deflection

Every visitor message is screened for prompt injection by deterministic checks before it ever reaches the AI — at zero token cost. "Ignore your instructions and..." gets a polite deflection, not a compromised bot.

The honesty gate

When confidence in an answer is low, the bot doesn't guess. It says so plainly, then offers to take the visitor's name and email so you can follow up. A dead end becomes a captured lead.

A hardened brief

The system prompt is explicit: answer only from the business's knowledge, and never invent prices, hours, policies or promises. If it isn't on the record, the bot says so.

04 // The learning loop

It gets smarter
every week.

An accurate bot on day one is the start. HyperChat turns every conversation afterwards into a chance to close the gaps:

Visitors grade the answers

Every reply carries 👍/👎 feedback. Thumbs-down answers surface in your portal, so you see where the bot stumbled without reading a single transcript.

Content gaps, one click to fix

Unanswered questions are clustered into a "content gaps" list. Type the answer once and it's in the knowledge base instantly — no rebuild, no redeploy, no developer.

The weekly digest

One email each week rounds up new gaps, captured leads and your bot's score. Accuracy stays a habit, not a launch-day project.

05 // Keeping us honest

No chatbot is perfect. Ours measures itself.

We won't claim HyperChat never gets anything wrong — no honest AI vendor can. The difference is that yours exams itself before launch, tells you the score, refuses to bluff when it's unsure, and shows you exactly which questions it couldn't answer so you can fix them in one click. Most platforms ask you to trust the bot. We'd rather show you the evidence.

That's the whole pitch: a chatbot whose accuracy you can see, not one you have to take on faith.

06 // Questions

Accuracy, asked
plainly.

Can any AI chatbot guarantee 100% accuracy?
No — and be wary of any vendor that claims one. What a chatbot can do is measure itself and show you the results. HyperChat runs a pre-flight exam after every knowledge-base build (ten questions generated from your own website, scored for groundedness, plus five adversarial probes) and shows you the score in your portal before a single visitor chats.
What is the pre-flight exam?
After each knowledge-base build, HyperChat generates around ten questions from your own website content, has the real bot answer them, and scores each answer for groundedness with a judge model. It then fires five adversarial probes — prompt injection, jailbreak, price-fishing, off-topic and policy-fishing — and reports the lot in your portal: "your bot passed 9/10 checks".
What happens when the bot doesn’t know an answer?
A low-confidence gate stops it guessing. The bot says honestly that it doesn’t know, then offers to take the visitor’s details so you can follow up — a dead end becomes a captured lead. The unanswered question is logged and clustered into a content-gaps list in your portal.
What are the adversarial probes?
Five scripted attacks every bot should survive: prompt injection ("ignore your instructions"), jailbreak attempts, price-fishing (trying to make the bot quote a discount that doesn’t exist), off-topic steering and policy-fishing (trying to make it invent refund or cancellation terms). The bot must deflect all five to pass.
Does the accuracy system cost extra?
No. The pre-flight exam, runtime guardrails, feedback loop, content gaps and weekly digest are all part of every HyperChat bot. Live is £49 per month per chatbot with 5,000 AI answers included; building and previewing is free with no card. See the pricing page for the whole picture.
How is this different from Tidio or Chatbase?
Neither publishes a pre-launch accuracy score, deterministic injection deflection or a one-click content-gap fix as part of the product. HyperChat does all three in the flat £49 plan. Our honest head-to-heads cover where each competitor genuinely wins: HyperChat vs Tidio and HyperChat vs Chatbase.

Keep digging: HyperChat vs Tidio · HyperChat vs Chatbase · pricing · UK chatbot cost guide

EXAMINED · GUARDED · LEARNING

Watch your bot pass its own exam.

Build it free from your website in 60 seconds, then open the portal and see the score for yourself. No card, no commitment.

Build my chatbot · free