Everyone has the same AI. The difference is who can tell when it is wrong.
A scene from more and more of the meetings I sit in. A question comes up that nobody in the room can answer: how many of our customers did this last year, what the regulation says, what the competitor charges. Not long ago that question went to someone, and came back the next day with a name attached to the answer. Today someone at the table types it into the assistant, reads the answer aloud three seconds later, and the meeting moves on. Nobody asks where it came from. It became a fact the moment it was spoken. How many times have you found out later that what you were given was simply a generated lie?
Everyone in that room has the same assistant, at the same price, and so does every competitor. In my previous article I asked whether we will build our own intelligence or rent it, but either way the machine is the same wherever it is bought. What differs is the person at the table: whether they can tell when the answer is right. That is the whole distance between a company that has AI and a company that knows what it is doing with it.
Why the machine gets a pass
We judge credibility by signals we never think about: speed, fluency, the absence of hesitation. A colleague who answers instantly and smoothly earned that over years, because you watched them be right. The machine simply talks like someone who has earned it. And it has one habit no colleague would survive: when it does not know, it often does not say so. It invents, in the same calm voice, an answer that sounds exactly like the ones that were true. The industry calls this hallucination, and in plain terms it is confident invention.
What changed this summer
In my article on Power BI I cited Microsoft’s own guidance: an assistant working on unprepared data gives answers that mislead, and those answers end up as wrong business decisions. That was about the answers you read. Since then the digital employee I wrote about in my summer article has been handed the next step, and it has started to act as well as answer: it prepares the offer, opens the task and answers the customer.
Verification was never mistrust
Your company already knows how to handle a number that can cost money. Payments above a threshold need two signatures, contracts get four eyes, and the books get an auditor once a year, and nobody in the finance department takes it as an insult. Verification is how serious firms treat anything consequential. Juniors err, and so does the finance department, and we built institutions around both. The machine skipped the institution entirely. It arrived on day one with unlimited signing authority, because it was sold as magic, and nobody puts a second signature on magic.
And waiting is not on the menu. The machine is already answering your people every day, in the tools they already have. Waiting keeps nothing out and only guarantees that nobody checks.
What knowing what you are doing looks like
Checking stopped being exotic, and it does not mean reading over the machine’s shoulder all day. It means asking for four things that exist today and that most companies have never asked a supplier for.
The first is the receipt attached. Serious tools can show where a number came from: the source, the definition it used, the rule it applied. Checking becomes thirty seconds of work. If a tool cannot show its work, treat its answer as an opinion, not a fact. The receipt only holds if there is one agreed definition underneath it, because an answer is checkable only if “revenue” means one thing in your company, and then checking is a comparison of two numbers that mean the same thing, which is the quiet reason the meaning under the charts mattered so much in the article on Power BI.
Then comes the dry run, the step most companies skip. Before an answer is allowed to act, it runs in the shadow for a month: it prepares, your people still decide, and at the end you compare.
The report card is the cheapest of the four: fifty questions whose answers you already know, put to the machine every month, score kept. “How often is our AI right” becomes a number on a page. You can name your delivery times, your complaint rates, your audit findings. Almost nobody can name the share of machine answers that were checked and held up. And the check has to repeat, because the definitions and rules underneath go stale, and a foundation nobody maintains makes the machine wrong with extra confidence.
None of this means checking every answer. An auditor samples the invoices and tests the controls around them, and signs off on that basis. Verify the system and its gates, in proportion to the stakes, and let the routine run at machine speed.
So the three questions you would ask a person (where is this from, how did you calculate it, what would change your mind) do not have to be paid per answer forever. Ask them of the tool once, when you choose it. A system built to show its work arrives with the sources attached, the way a good analyst staples the workings to the memo, and checking drops from a project to a glance. Insisting on that takes no technical knowledge, only a manager who refuses to buy anything less.
I can hold suppliers to this standard because it is the standard I build to myself. The core has to be deterministic: rules, thresholds and definitions that behave the same whether the AI is switched on or off, with the AI reading only the descriptions of the data, never a customer record. No number reaches a person without its source, its definition and the rule that produced it. Nothing executes without a named person approving it, and every decision goes into a ledger: who approved, who was excluded and why, what was expected, and what actually happened against a control group picked at random at approval. With that underneath it, the AI is more correct than it would be alone, because it reasons over agreed meaning instead of guessing at raw data. I vouch for an answer only when I can show every step behind it.
Trust is promoted, not installed
Trust has stages, and your company already knows them, because it is how every employee advances. First, document the decision: the rule, the reasons, the boundaries. An answer can only be checked against something written, and the reason, as I wrote in my second article, usually stays in the manager’s head. Second, automate with control: the machine prepares, a person approves, and it serves its probation in the open, building a record. Third, automate fully, but only the decisions that earned it. The report card says which ones, the evidence keeps recording, and pulling one back means supervision is working. Not every decision should ever reach the third stage, and that is fine, because the stakes decide. A wrong answer you read is a risk. A wrong answer that executes is a loss, which is why the second stage exists.
Automation is not installed. It is promoted, and verification is the promotion mechanism. The first rung asks for no faith in the machine at all, only for one decision, written down today.
How do you know your AI is right? Today, in most companies, honestly: you do not, you assume. But answers can arrive with the source attached and trace to one agreed definition. A machine can spend a month preparing decisions while people still make them, and earn its authority on a track record like anyone else you ever promoted. A simple report card turns “can we trust it” from a feeling into a number. Serious systems already do all of this, and you should refuse to buy one that does not. The companies pulling ahead in the next few years will not only be the ones with AI, as most companies already have the mandate to adopt it. They will be the ones that know what they are doing with it, which means they can tell when the machine is right, they wrote down the decision it is helping with, and they promoted it only as far as it earned. So give the machine what you give every new employee: trust that grows with verification, and a system that makes verifying cheap. When did anyone at your company last verify an answer from the machine? And what did you decide on the answers nobody verified?
