Three Questions Before You Trust an AI Answer: Stakes, Reversibility and Checkability

⏱ 8 min read|Published 13 November 2026|By Dibyendu Choudhury

Most advice about using an AI assistant at work falls into two camps. One camp says to trust it, because it is fast and usually right. The other says to distrust it, because it can be confidently wrong. Both camps are describing something true, and both leave a busy owner with no practical rule. You cannot verify everything an assistant tells you, because then you would save no time. You cannot verify nothing, because sooner or later a wrong answer will land in front of a customer, a supplier or a tax officer with your name on it.

What helps is a short habit that sorts answers by how much checking they deserve. It takes under a minute, and it rests on three questions. How much does it matter if this is wrong? Can I undo it if it is wrong? And can I tell, without much effort, whether it is right? Asked in that order, they turn a vague worry into a decision.

Question one: what are the stakes?

Start with the cost of being wrong, measured in the units that matter to your business: rupees, relationships, time, reputation. A wrong suggestion for the subject line of a newsletter costs you a few fewer opens. A wrong figure in a quotation costs you the margin on the whole order. A wrong statement about a customer's legal rights, written in a reply from your business, can cost you far more than the sale it was meant to save.

It helps to name three rough bands. Low stakes are things where an error is cheap and quickly forgotten: brainstorming names, rephrasing an email, summarising a long thread for your own reading. Medium stakes are things that leave the building but can be corrected: a product description, a social post, a draft reply to a routine query. High stakes are things that bind you or someone else: prices, contracts, tax and compliance positions, medical or safety information, anything that moves money.

The band does not depend on how good the assistant is. A very capable tool still sits inside the same band, because the band describes your exposure, not its ability. An owner who remembers this stops asking whether the tool is reliable in general and starts asking whether this particular output is allowed to be wrong.

Question two: can I undo it?

The second question is about reversibility, and it changes the answer more often than owners expect. Two outputs can have the same stakes on paper and very different risk in practice, because one can be pulled back and the other cannot.

A post that turns out to contain a slip can be edited within the hour. A printed brochure with the same slip is in a thousand hands. A message sent to a single customer can be followed by a quiet correction. A message sent to an entire list cannot be un-sent. A draft invoice can be fixed before it goes out, while a payment released on the strength of a wrong figure may never come back in full.

So before you rely on an answer, picture the moment after it goes out. If you spot the error then, what does it take to put things right? If the honest answer is "an apology and an afternoon", you can move faster. If the honest answer is "a difficult phone call and a refund", slow down and add a check. Irreversible actions deserve a second pair of eyes whether or not an AI was involved; the assistant simply makes it easier to forget that.

Question three: can I check it cheaply?

The third question is the one most people skip. Some answers are easy to verify, and some are expensive or impossible to verify from where you sit. The difference matters more than the assistant's tone of confidence, which is rarely a useful signal.

Checkable answers have a visible test. A formula in a spreadsheet either gives the total you can see on a calculator or it does not. A list of ten items can be counted. A summary of a document you have in front of you can be compared with the document in two minutes. A draft in your own voice can be read aloud, and your ear will tell you whether it sounds like you. For these, the sensible habit is to let the assistant do the heavy lifting and spend your attention on the check.

Hard-to-check answers have no such test within your reach. A claim about a rule, a rate, a deadline, a date, a quotation, the name of a person or the details of a case sits in this group. You could confirm it, but only by going to the original source, and that is exactly the effort you hoped to save. Here the assistant is most useful as a guide to where to look, and least useful as the final word. Ask it what the source would be, then open the source yourself.

There is a subtle trap in this group. The more specific and official an answer sounds, with a section number, a percentage, a date or a neat citation, the more tempting it is to accept it as it stands. Specificity feels like evidence. It is not. A fluent, detailed, wrong answer and a fluent, detailed, right answer look identical on the screen, so for anything you cannot test, treat precision as a reason to look up the original, not as a reason to relax.

Putting the three questions together

Run an output through the questions and it usually lands in one of four places. When the stakes are low, you can undo the result and you can check it quickly, use it and move on. When the stakes are low but the answer is hard to check, use it as a starting point and do not repeat it as fact. When the stakes are higher but the answer is easy to check, do the check every time; it is cheap insurance. And when the stakes are high, the action cannot be undone and the answer cannot be checked by you, the assistant does not get the final say. A person who knows the subject does, whether that is your accountant, your lawyer, your supplier or the original rulebook.

This is not a complicated scheme, and that is the point. A rule you can apply while a customer waits on the phone is worth more than an elaborate policy nobody remembers. Write the three questions on a card beside your screen for a fortnight and you will find yourself asking them without the card.

A worked example from a small shop

Imagine a small online seller of handloom scarves who uses an assistant for three tasks in one afternoon. The first is a set of product descriptions. Stakes are modest, the text can be edited later, and the seller can read each description against the actual scarf in a moment. Verdict: use it, read it, publish it.

The second is a reply to a customer who wants to return a scarf after the return window has closed. The assistant drafts a firm and polite refusal. Stakes are medium, since a bad tone can lose a customer. The reply cannot be unsent, but the seller can read every word before it goes. Verdict: edit for warmth, then send.

The third is a question about how a certain shipping charge should be treated in the seller's accounts. The assistant gives a confident, detailed answer. Stakes are high, because it affects the books. The result is hard to check without the rulebook, and a wrong entry can echo through the year. Verdict: take the answer as a list of questions for the accountant, not as an instruction.

Same tool, same afternoon, three different levels of trust, and none of them required a technical understanding of how the assistant works. They required only an honest look at what each output could cost and how easily the owner could find out.

How to make the habit stick

Three small practices keep the habit alive. First, ask the assistant to separate what it knows from what it is guessing. Many tools will do this when asked, and even an imperfect separation tells you where to point your attention. Second, when a figure or a rule appears in an answer, copy it into a note with the word "unverified" beside it until you have seen the source. Third, keep a short log of the times an answer turned out to be wrong. After a month, the log will show you which kinds of questions your assistant handles well and which it does not, and you will have a map drawn from your own business rather than from anyone's marketing.

None of this makes you less of an enthusiast. The owners who get the most from these tools are not the ones who believe everything or the ones who believe nothing. They are the ones who know, for each task, how much checking it deserves, and who spend their attention where it buys the most safety.

One question to take away

Think of the last thing an AI assistant told you that you acted on without checking. Which band was it in, and would you do the same today? Share one line in the comments if you wish; I read every one.

If you enjoy a long read after a long working week, my Spirituality Series book Krishna-Draupadi Samvad is out now on Amazon Kindle, and you can find it on the Spirituality Series page.

---

Published 13 November 2026 · dibyenduchoudhury.com