The Escalation Is the Instrument
The minute your system gets stuck is the most valuable minute of its week. The standard design throws it away.
AI that works at scale runs on one loop, and it is duller than the headlines suggest. The system does something. An outcome scores what it did — the customer bought, or did not; the quote was accepted, or was not. The score corrects the system. Round it goes. Strip away the architecture and that loop is the whole engine. Which half you fund first is a real decision, and the scoring half is the one I would fund.
The loop needs volume you do not have
Here is the part that rarely gets said out loud to a business your size. That loop runs on volume. It corrects itself because it is wrong a thousand times a week and each of those leaves a mark in the numbers.
So when you are told that the system “learns from use”, the sentence is true, and it is not true of you in the way it is true of them. You will not accumulate enough quotes, enough jobs, enough refusals for a pattern to surface statistically. Fifteen instances is an anecdote. The honest version is that a small business cannot buy its way into that loop, and pretending otherwise is how you end up with a confident system and nothing to catch it. I would rather run no system than a confident one with no brake — it is the same failure as an unchecked spreadsheet, only faster and better spoken.
Which leaves the real question: if not volume, what corrects it?
The thing that gets thrown away
Watch what happens when one of these AI systems meets something it cannot handle. It hands over. A ticket appears, a person picks it up, the person sorts it out, the ticket closes. Everyone behaves well and the customer is served. It feels like the responsible design.
It is also where the entire value goes out of the window.
Because what just happened is this. The system reached a decision it could not make. A person made it. That person had a reason. And the reason — not the answer, the reason — is the only correcting signal your business is ever going to produce. You had it in your hands for about four minutes and you closed the ticket on it.
I do it too. My own system logs every decision that escalates to a person — how many, what kind, and why it came to them rather than staying put. It has run every week for months. Nowhere does it record what the person decided, or why. I built the ledger and left out the column that mattered.
Kavak, the Latin American used-car business, made this point from the opposite end. Alejandro Maza Ayala, their chief product and AI officer, speaking in August, said the reason they refuse to let a stuck case drop into a second-line queue is that the queue breaks the loop. Which I read as: you solve the customer’s problem and you lose the data that would have stopped it happening again. So their systems escalate by asking for help, explicitly, and the answer comes back into the system rather than dying in a support tray. He also recommends roughly matching the time, tokens and money spent on scoring to what is spent on the agents themselves. Separately, on pace: he likes to move extremely fast, and to move fast you need brakes. I am quoting the mechanism and not their numbers, which are self-reported and would not survive the scrutiny I would want to apply to my own. The mechanism stands on its own.
The inversion
An escalation rate is a number an AI project is set up to drive down — the one that embarrasses it in month three. Turn it over.
An escalation is the system telling you, with a timestamp, exactly where the judgement in your business lives. Not judgement in general — this decision, this Tuesday, this customer, with the actual circumstances attached. It found the spot without anyone being asked to look inward, which matters more than it sounds, because looking inward is the part that does not work. I wrote a fortnight ago about the system of record that only exists in the owner’s head, and about why asking for it never works. Ask cold and you get a shrug — not evasion, just that a decision made often enough stops having steps. Ask at the moment the decision is being made, with the specifics in front of you, and you get the real answer, because there is something concrete to be right about.
That is the whole trick. The escalation is not an interruption to the build. It is the instrument.
What this looks like on a Tuesday
Three things get captured at the point of escalation, and only three. What the system could not decide. What the person decided. Why — in a sentence, in their words, at the time. The third is the asset and the first two are the filing.
The discipline is that the reason goes back in. If it does not, you have built a well-documented queue and you will be having this conversation again next year. Closing that loop is unglamorous work, and it is the work.
Then something useful happens without anyone planning it. After a few weeks you have a ranked list of the places your business depends on one person’s judgement — ranked by how often it actually came up, not by how important it felt in a workshop. That list is what an away-day with a whiteboard is meant to produce, and in my experience the away-day version records the decisions that feel weighty rather than the ones that recur. Yours was generated by observation. It is not anybody’s opinion.
Which flips the sequence people expect. The advice is normally to write everything down first and automate afterwards — and if you can, do; I have argued for it plainly enough. But if the writing-down never starts because the blank page is too big and there is a business to run, then put something modest in front of a real process and let it get stuck. Where it stops is where to start writing. You are not asking the business to document itself in the abstract. You are letting it point.
What you are actually building
None of this needs a large budget and none of it is exotic. It needs the handover to be treated as the valuable event it is rather than a tidy-up, and it needs someone to insist that the reason gets recorded while it is still fresh, which is a management habit rather than a technology.
Two honest limits. This does not lift thirty years of judgement out of someone’s head — it reduces what the business depends on them for, at the margin. And it is not an asset you capture; it is a habit that produces one, slowly.
What accumulates is not a smarter system, or not only. It is a written record of how this business decides, built from real decisions rather than remembered ones, in priority order, at the pace the business actually throws them up. That record is worth having if the AI project succeeds. It is worth rather more if it does not — because it is the part that stops the business needing one particular person in the building to know what good looks like.
If your systems are already handing things over and nobody is writing down why, you are throwing away the only correcting signal the business produces. If you would rather not work out the mechanics of it alone, you know where to find me.