12 min readBy Tommy Dempsey

When Your AI Should Say "I Don't Know": Confidence, Explained for Non-Engineers

Ask most support bots a question they have no answer for, and they answer anyway, because answering is the only thing they were built to do. The more useful behavior is rarer: the AI recognizing that it does not know, and routing the question to a person instead of guessing. That recognition has a name - confidence - and the setting that decides what happens when confidence is low is a threshold. Here is what both mean, in plain terms, for the person running the business rather than the one who built the model.

Ask most customer support bots a question they have no real answer for, and they will answer anyway. Not because they are lying, but because answering is the only thing they were built to do. A fluent, confident paragraph comes out, it sounds right, and the customer believes it. The more useful behavior is rarer and harder to build: the AI recognizing that it does not actually know, and saying so instead of guessing. That recognition has a name in the tools that do it well. It is called confidence, and the setting that decides what happens when confidence is low is a threshold. This post explains both in plain terms, for the person running the business rather than the person who built the model, because the threshold is a decision about your customers, not a technical knob to leave on default.

What a confidence score actually is

A confidence score is the AI grading its own answer before you ever see it. After it writes a reply, it also produces an estimate of how solid that reply is, usually as a number on a scale. High means the AI thinks it had good material to work from and answered the question directly. Low means it is reaching, patching together something from thin or unrelated material. Your software can then act on that number automatically: let a high-confidence answer through, and hold a low-confidence one back for a person.

The useful way to picture it is how you read a human colleague. When you ask someone a question, you can usually tell from the way they answer whether they know or are winging it. "It is thirty days, I checked the policy last week" lands differently from "I think it is maybe thirty days?" The words might even be similar, but the certainty is not, and you weight the two answers differently without thinking about it. A confidence score is the tool trying to turn that gut read into a number, so the handoff to a human can happen on its own instead of depending on you to notice hesitation in a reply you may not even be looking at.

One thing to get straight early, because it is the most common misunderstanding: a confidence score is not a measurement of truth. It does not check the answer against the real world. It is closer to a measurement of how much solid-looking ground the AI felt it had under the answer. That distinction sounds like hair-splitting, and it is the whole point of the rest of this post.

Why "I do not know" is a feature, not a failure

The instinct is to treat a bot that says it cannot help as broken. You paid for an AI, and here it is punting. That instinct is backwards. The AI that worries me is the one that answers everything, because answering everything means answering the questions it should have refused, and those answers go out looking exactly as confident as the good ones. A customer cannot tell a well-grounded reply from a fluent guess. They read a paragraph and act on it.

Compare the two outcomes directly. A confidently wrong answer is a mistake that leaves the building. The customer now believes something untrue about your refund window, your shipping time, your compatibility, and they will make a decision on it, and you will hear about it later when it is tangled up with an actual problem. An honest "I do not know, let me get a person on this" is a two-minute task in your queue. One of those costs you a customer and an afternoon. The other costs you two minutes. A tool that can tell the difference and choose the two-minute path is doing the single most valuable thing a support AI can do.

A support AI is only as trustworthy as its willingness to decline. If it never hands anything back, it is not confident, it is indiscriminate, and you will not find out which until a wrong answer reaches a customer.

What a confidence score does not tell you

Here is the limit, stated plainly so you do not over-trust the number. High confidence is not proof of a correct answer. The AI can be confidently wrong, and it happens in two ordinary ways. The first is a bad document. If one of your policy pages says the wrong thing, or says it ambiguously, the AI reads it as gospel, answers directly from it, and reports high confidence the whole time. It had solid-looking ground. The ground was just wrong. The score measured the ground, not the truth.

The second is the exception. A customer asks for something slightly outside your written policy, a few days of grace on a return, a one-time courtesy, and the AI knows the policy cold. Clear rule, clear question, high confidence, polite denial. But certainty about the rule is not the same as knowing whether the rule should bend for this particular customer, and the AI cannot feel the difference. I wrote about that whole category, where the AI sounds most sure exactly when it should be most careful, at /blog/support-questions-never-let-ai-answer.

So confidence is a filter for one specific thing: uncertainty. It reliably catches "I have nothing to stand on here," which is the case where an AI without a confidence check would invent an answer. It does not catch "I have a clear source that happens to be wrong," and it does not catch "I answered the question that was asked but missed the situation underneath it." Knowing what the number cannot do is what keeps you from leaning on it as a guarantee it was never built to be.

The threshold is a dial, and it is your decision

If the confidence score is the AI's self-assessment, the threshold is the line you draw across it. Above the line, the AI proceeds, drafting or sending a reply. Below the line, it stops and routes the message to a human instead. That line is the one real decision in all of this, and it is a business decision wearing technical clothes.

Think of it as a dial with a tradeoff at each end. Set the bar high, and the AI only acts when it is quite sure, so it hands you more messages and gets fewer things wrong, at the cost of your time on the pile it declined. Set the bar low, and the AI acts more often, saving you time, at the cost of more shaky answers slipping through. There is no universal right position, because the cost of a wrong answer is not the same for every business. A wrong answer about store hours is a shrug. A wrong answer about whether a medical or financial product is suitable is not. The higher the stakes of being wrong, the higher you want the bar.

I am deliberately not going to hand you a magic number, partly because the right one depends on your stakes, and partly because a number copied from a blog post is a number you did not reason your way to. What matters more than the exact setting is understanding which way the dial moves and why. Most good tools also keep a hard floor underneath whatever you choose, a point below which the AI always routes to a human no matter what, so that even a generous threshold cannot let a near-random guess go out. You want that floor to exist. You do not need to know its exact position to benefit from it.

How to read your "I do not know" pile

The messages the AI hands back are not wasted. They are some of the best information you will get about your own setup, if you actually read them instead of just clearing them. Every escalation falls into one of two kinds, and telling them apart is the exercise.

The first kind is a genuine handoff. The question needed human judgment, or it touched something the AI correctly stays away from, and routing it to you was the system working exactly as designed. Nothing to fix. Answer it and move on. The second kind is the one worth your attention: the AI handed back a question whose answer was actually sitting in your documentation, and it just did not find it or did not find it clearly enough. That is not the AI failing so much as a documentation gap showing itself. The answer existed, but it was buried, phrased vaguely, or split across pages in a way the retrieval could not assemble.

Sort a week of escalations into those two piles. The genuine handoffs tell you what your AI will always need you for, which is useful to know. The should-have-found-it pile is your to-do list: each one points at a document to rewrite or restructure so the answer is there next time. How the AI pulls answers out of your docs, and why structure matters more than sheer volume, is at /blog/how-ai-reads-your-docs-chunking-retrieval, and the broader habit of measuring this over time is at /blog/measure-ai-customer-support-quality-small-team.

Confidence is one guard, not the only one

It would be tidy if the confidence threshold handled everything, but it does not, and treating it as your only safety net is a mistake. Some messages should never reach the AI in the first place, no matter how confident it would be. An angry customer, a chargeback, a request to move money, a mention of a lawyer: these go to a person on the first read, because the problem with them is not that the AI lacks information. Those are handled by blunt keyword rules that fire before the AI drafts anything, not by a confidence score after the fact.

The right way to picture the two together: keyword rules catch the cases you can name in advance, and the confidence threshold is the net for the cases you did not anticipate, the weird one-off phrasing and the questions that sit just outside your documentation. The keyword list is your known list of hazards. The confidence floor is insurance against the unknown ones. You want both, because each catches what the other misses. The full version of the never-let-the-AI-answer argument, with the three families and how to route them, is at /blog/support-questions-never-let-ai-answer.

You do not need to be an engineer to get this right

If all of this sounds like it requires tuning a model, it does not. You are not adjusting the AI. You are making a handful of ordinary business judgments, and the order to make them in is simple. Start in draft-and-approve mode, where the AI writes but you read every reply before it goes out. In that mode the confidence threshold barely matters, because you are the threshold: a shaky answer still lands in front of you, and you catch it. That is the safest place to begin and the place to stay until you have watched the AI perform on your real inbox.

The threshold starts mattering the moment you turn on auto-send for a category, because now the AI is deciding on its own whether to reply or route. Before you flip that switch for a given type of question, you want to have already read a month of the AI's drafts on those questions and seen them come out right. Earn auto-send on the safe, boring, deterministic lookups first, and leave everything with judgment in it on draft or routed to you. The week-by-week version of that adoption arc is at /blog/draft-first-ai-email-first-90-days, and the auto-send-versus-approve decision by category is at /blog/when-to-let-ai-reply-autonomously-small-teams.

What Trigli does here, and what it does not

Trigli scores confidence on every reply it generates, in email and in chat, and it uses that score to route. Below a floor, a message goes to a human instead of being sent, and it will not even draft a near-random guess for you to approve, because a draft that low is noise, not help. Draft-and-approve is the default, so early on you see everything regardless of the score. In chat, where a visitor is waiting in real time, the AI escalates when it is unsure and hands off with context so the person picks up warm, and if no agent is online it captures the visitor's email rather than faking presence. Automation rules let you force whole categories or keywords straight to a human, which is how the keyword guard above gets implemented alongside the confidence floor.

Now the honest limits, because the whole point of this post is not over-trusting a number. A confidence score is not a lie detector. Trigli's floor reliably catches the AI reaching for an answer it does not have, but it cannot catch a wrong fact stated confidently from a document that is itself wrong, which is exactly why draft mode matters early and why your documentation quality is the real ceiling on reply quality. The score is a safety net for uncertainty, not a guarantee of correctness, and I would rather tell you that plainly than sell the number as more than it is. Beyond that, Trigli has no phone or SMS channel, and it cannot push a refund through your payment processor, so moving money stays a human step, which for the questions confidence is weakest on is arguably the right boundary anyway.

If you are a large CX organization that wants deep per-agent analytics and a dozen channels, a heavier helpdesk is the honest recommendation. If you are a small team on Gmail or Outlook that wants an AI which knows when to say it does not know and hands those cases to a person cleanly, that is the shape Trigli is built for.

The short version

A confidence score is the AI grading its own answer before you see it, expressed as a number. It measures how much solid ground the AI felt it had, not whether the answer is true, so it catches uncertainty and cannot catch a confident mistake from a bad document. The threshold is the line that decides what happens below a given score, and where you set it is a business judgment about what a wrong answer costs you, not a technical knob. Read the questions the AI hands back: the genuine ones tell you what it will always need you for, and the ones whose answers were really in your docs are a documentation to-do list. Start in draft mode so you are the threshold, earn auto-send on the safe categories, and never treat the number as a promise it was never built to make. The best thing a support AI can do is know when to say it does not know.


See where your AI says "I do not know" on your own inbox. The free tier is 50 emails, 25 chats, and 10 tickets a month with 3 documents, no card required. Keep it in draft mode, read what it hands back, and you will learn as much from the questions it declines as the ones it answers.

Related reading

  • How to Run a Fair One-Month Trial of AI Support on the Free Tier

    Most AI support trials fail before the AI writes a single reply, because the trial is rigged to pass. You feed the demo three softball questions, the AI nails them, and you learn nothing. A fair trial points the AI at your real inbox, your real docs, and your real last month of questions. Here is how to run that trial in about thirty days on a free tier, and three numbers to watch so the result is honest.

  • Live Chat for a One-Person Business: What to Automate and When to Say Nobody Is Online

    Live chat looks like the one channel a one-person business cannot run, because it implies someone is sitting there ready to answer. You are not, and you should not pretend to be. Here is how to run chat as a solo operator: what to let the AI handle, where it should stop, when to show the email-capture fallback, and why "no agents online, leave your email" beats a fake "an agent will be right with you."

  • How to Write a Refund Policy an AI Can Actually Answer From

    Most refund policies are written for one job: covering the business if something goes to a dispute. That is a different document from one an AI can answer a customer from. Here is how to rewrite the second kind, with before and after examples you can copy the shape of.

Ready to handle support without the headcount?

Free plan, no credit card. 14-day trial on paid plans. Takes about 5 minutes to set up.