By Joseph Batrin
Your best customer, six years of invoices in, asks for a quote on a new project. Your AI pulls the current pricing, applies your standard terms, builds the scope, and sends it at list.
Nothing in that is a mistake. The pricing was current, the terms were right, and it fired exactly the way you built it.
And that is not the number you would have sent. They referred you two accounts last spring. A competitor has been working them since March. You would have honored last year’s price and called them yourself before anything went out. No rule would have caught that. How you feel about a customer who has stood by you for six years is not something a system can weigh for you. It lives in your head.
The Myth
The myth is autonomy. Set it up, walk away, let it run. The pitch is an employee who never sleeps and never needs managing, and the quiet promise underneath is that watching it is a phase you graduate out of.
You do not graduate out of it. The oversight is not training wheels. It is part of the machine.
The Reality
Two very different failures both get called an AI mistake, and they have opposite fixes. Telling them apart is the whole lesson.
Here is the first. A customer pays by ACH on Friday. Monday morning your collection sequence fires because the invoice is one day past due, and it sends her a notice while the deposit is still settling. Owners call that an AI problem. It is not a judgment failure. It is a bad trigger. Build a settlement window into the workflow and the problem disappears.
That is a setup problem and you fix it in the setup. Standing in front of every notice would be spending your attention to cover a bad build, and your attention is the most expensive material you own.
The second failure is the one no build prevents. The pricing was current. Every rule fired. The referral and the competitor were never in a system to begin with.
And sometimes the system knows everything it needs to know and still should not have the authority to make the decision. Information and permission are not the same thing.
So the question is not whether AI eventually gets reliable enough to run alone. It is which of your outputs can be right on paper and wrong in the room. In most businesses there are five, and your industry may add a few of its own.
A quote goes out at list price to a customer who has been with you for six years and sent you two referrals last spring. Anything that involves money, whether it quotes a price, commits a discount, waives a charge, or moves a dollar, needs your review. A simple total built out of prices that are already set does not need you to check it. When you have a loyal customer like this one that you would give a discounted price to, that is something only you are going to know, and it needs your review. The file will never be able to duplicate that information.
A customer asks whether your product handles something your documents do not quite cover. It answers, helpful and confident, and the answer reads like a commitment. That is a commitment, and commitments cover contracts, terms, guarantees, warranties, promises about performance, claims about what you do, and anything an attorney or a regulator would ever read. Nobody wrote a rule for that question because nobody had ever been asked it.
A reply goes out over your signature promising a turnaround nobody agreed to. That is your name, which covers proposals, public posts, and any answer your approved material does not already handle. The appointment confirmation is not on this list and never needs you. A sentence a customer could hold you to is on it every time. It assembled the words and you signed them.
An email arrives with the word cancel somewhere in it, and the account is closed before anybody reads the rest of the sentence. That is irreversible, which covers deleting records, closing accounts, canceling services, submitting something outside the business, and anything that cannot be quietly recalled. It is on the list because this is the one place a mistake has no recovery.
A customer of eight years writes in angry, and a technically correct answer goes back out of the policy when the moment called for you to pick up the phone. That is judgment, and judgment has no answer key. Several responses are defensible, and choosing between them takes context it cannot see.
Why This Costs You
None of it shows up as an error. The list price to the customer who has been sending you referrals. A helpful answer that turns into a commitment you never made. A canned reply to the one customer who deserved a call.
Any one of those can cost you a customer you spent years earning. What makes it worse is a system producing them at speed while everyone assumes it is fine because it has been fine so far. Nothing in the log ever flags it, because nothing errored. It did its job. The customer just goes quiet, and by the time you notice, they are gone.
A Practical Example
Go back to that proposal and ask one question about it. Could the thing that would have changed the answer have been built into your system?
If it could have, you have a build problem. Wire the source in, add the wait, fix the trigger, and you solve it once instead of every morning.
If it could not, because the answer depended on a relationship or a decision only you can make, then no build is going to fix it. It gets flagged for your review, and it goes on your list. The two referrals last spring. The competitor circling since March. The fact that this customer is having a hard year and you would rather carry them through it. Those are the calls that need your personal touch before anything goes out.
That is the whole point. Some work is automatic and belongs in the system. Some work needs your review, your judgment or your personal touch before it reaches a customer. You do not turn AI loose and let it send whatever it produces. You build oversight into it, and you spend your time on the handful of decisions where being wrong costs you something.
What Actually Works
Two passes, and the order matters.
First the build, because most of what gets called an AI mistake is a setup mistake. The wait before the notice goes out. The bank checked before the system assumes. The trigger that verifies instead of counting days. Each one is an afternoon. Doing the same job with your own attention is a few minutes every day for as long as you own the business.
A build is more than its rules, too. The rules cover the cases you thought of and the exception path covers the ones you did not. When it cannot tell whether the money landed, it holds the notice and flags the account instead of guessing. The payment clears and the notice never goes at all. The payment fails and the sequence runs on its own.
Then the gate. A rule is something the system is told, and a rule can be read and forgotten. A gate is a step it cannot skip. Before the work is allowed to leave, it stops and verifies the information, and if it cannot verify it, the work comes to you for approval. That is where the mistakes only you can catch get caught, and where the decisions only you can make get made.
And it belongs at the end, not across the whole job. It can research the account, pull the history, assemble the numbers, draft the proposal, populate the CRM and format the email with nobody standing over it. The moment that proposal is about to leave with a price on it, it stops.
What the stop looks like decides whether any of this saves you time. The proposal arrives finished, with the price it means to send, what this customer has bought over six years, and what you charged them last time. You approve it, change the number, or send it back. Then it goes, the decision gets recorded, and the follow up runs without you. Twenty minutes of assembling a quote became a short review of the only part that needed you.
And if you make the same change for the same reason every time it stops, that is not the gate doing its job. That is something the system should have known, and you just found your next build.
Money, commitments, your name, anything irreversible, and judgment calls stop for a person. Everything you can safely systemize keeps running. That split is what buys you the speed you are paying for.
Do This Today
Write your never alone list. These are the things you need to see before they go out to a customer. It might be three lines or twenty five. The size of the list depends on your business, how you run it, and what you are not willing to risk.
Then take each line and ask whether it could run on its own and still be right every time. If it can, without risking the wrong price, a commitment you never made, or a customer you worked hard to earn, build it into the system and let it run. If it cannot, build a gate around it so it stops for your eyes before it leaves.