Every company has one escalation that keeps coming back. Start there.

The Routing Layer is my monthly letter on execution inside AI-era companies: one problem per issue, worked down to the mechanism, specific enough to act on without hiring anyone. This one is for founders who can name their recurring escalation before they finish the first paragraph.

Ask a founder which decision keeps landing on their desk and the answer arrives before the question is finished. A pricing exception on a strategic account. A go-live date that Sales committed and Delivery found out about later. A discount somebody promised on a call that nobody wrote down. It is always the same one, or the same two, and it has been the same for a year.

Each time it shows up it gets handled as an event. Someone flags it, a thread forms, you read the thread, you make the call, everyone moves on. Resolved. Until the next instance, usually three or four weeks out.

The mistake is in the word resolved. A decision that has reached you three times is not a problem you keep solving. It is a seam in the company that was never drawn, and every instance carries the same message: nobody below you owns this.

Here is why it recurs. The decision lives between two functions. Finance sees a margin call. Customer Success sees a retention risk. Both are right, both assume the other holds the pen, and no rule anywhere says which one does. So the decision does what unowned decisions do. It travels upward until it hits someone with enough authority to end the ambiguity. That is you, with the least context of anyone involved and the least time, making the call in eleven minutes because the thread has already cost eleven days.

Bain's May research on tech company operating models calls this the default design. Product and engineering teams own the work, management resolves what falls between them, and anything still unresolved goes up. It holds until speed and scale increase. Then the meetings needed to hold it together multiply, upward referral becomes the norm, and the more traffic the seams carry, the longer everything takes. One caveat: Bain studied around 300 established tech companies, so read it as a pattern, not a benchmark for a 90-person business.

AI is the speed increase. Deloitte's 2026 predictions put the gap plainly: 48 percent of respondents say their organization introduced AI without redesigning the workflows or roles it sits inside, and 12 percent report redesign at scale with a new operating model behind it. Put AI in Finance and margin analysis gets faster. Put AI in Customer Success and account research gets faster. Neither team gets any faster at deciding who owns the call, because that was never a throughput problem. What you get instead is two accelerated functions producing more exceptions and more commitments, all arriving sooner at the same undrawn boundary, and all of them still routing to one desk.

The good news is that the recurring escalation is also the cheapest diagnostic you will ever run, because the company has already done the sampling for you.

Pick one. Not the process map, not a RACI for the whole business. The single decision that escalated to you most often last quarter. If two compete, take the one with a customer on the other end of it.

Reconstruct the last three instances. Four questions per instance, written down, because the point is the pattern across three and not the argument about one.

1. Who was supposed to own it? Check what the org chart, the SOW, the approval policy, the CRM field actually says. Frequently the honest answer is that nothing says anything.
2. Who actually acted? Pull the thread, the ticket, the record. Who moved it, who sat on it, who got tagged in and went quiet.
3. Where did it travel? Count the hops between the moment the question surfaced and the moment someone decided. Count the days as well. Elapsed days, including the ones it spent waiting.
4. Where did escalation stop? If the answer is you, three times out of three, the diagnosis is done.

One discipline while you do this: keep three kinds of evidence apart. What people remember happening. What the documents say should happen. What the system trail shows actually happened. Do not blend them into a theme. When they disagree, the disagreement is the finding, and the most senior person's version does not get counted twice.

Then draw the seam. Four decisions, one page.

The owner. One name, not a team.

What travels with the handoff. The three or four facts the owner needs to decide without calling a meeting. For the pricing exception: margin floor, contract value, renewal date, what was actually promised and by whom.

Where escalation stops. A named person below you with the authority to end it, plus the narrow condition under which it goes higher. Written as a condition, not as "use judgment."

What needs no human at all. The band inside which the answer is automatic: any discount under X on a contract above Y within Z days of renewal, approved. AI and automation can run that band once it exists. They cannot draw it. That part stays with you, once.

Then measure two things, both of them on Bain's list of operational measures: cycle time on that decision, question to answer in days, and the number of times it escalates per month. Success looks like the fourth instance resolving before you hear about it, and you finding out in a weekly review instead of a thread at 11pm.

One decision a month is twelve seams a year, and most companies at this stage do not have twelve. They have three or four, hit repeatedly, mistaken for twenty because each instance looks new.

Designing the path was always the founder's job. Being the destination was the accident.

If the reconstruction turns up more than one seam that ends at you, that is what the 7-Day Clarity Sprint maps.