How the Simulation Works

A company decides it wants to use AI to do some of its work. There are two very different ways to go about that. The question is simple: which way ends up with less of a mess?

You can’t answer that honestly in the real world, because no two companies are alike. So we build a pretend company inside a computer, make an exact copy of it, and run both copies at the same time. One copy adopts AI the first way; the other the second. Everything else is identical — same customers, same staff, same good and bad luck. Because the two copies differ only in their approach, any difference at the end came from the approach. That is the entire trick.

We picked an insurance company because its work is easy to picture. Three ideas you need:

That’s it. Requests come in, and they travel through a chain of desks until they’re finished.

Every time work passes from one desk to the next, that handoff is a seam. Two words carry most of this project:

Put those together and you get the two kinds of seam the experiment turns on:

Now remember the childhood game of telephone: a sentence whispered down the line turns into nonsense. The same happens at a seam. The work arrives at the next desk slightly garbled, missing a detail. The next person has to stop and untangle it before doing their own job. That wasted untangling time is what we measure.

Both teams run an identical copy of the insurance company. Only how they bring AI in differs.

One team learns by doing. The other plans, then acts. Which produces less mess? That’s the race.

The simulation runs in rounds. Each round is a stretch of time, like a couple of weeks at the office. The exact same round happens to both teams at the same moment.

Repeat for many rounds and you get two full stories: how each version of the company fared over time.

Step 4 is the heart of it. A single claim’s trip when it reaches an AI desk:

A smooth trip costs almost nothing. A garbled one costs real cleanup time and chips away at trust. Across thousands of claims over many rounds, the small differences add up to a clear winner — or a genuine tie.

Whether the AI succeeds or stumbles on a given handoff is decided the exact same way for both teams, at the exact same moment. If Team A’s AI trips over a claim, Team B’s AI trips over that identical claim too. Neither gets luckier. So any lasting difference can only come from how each philosophy fixes its handoffs over time — not from a good roll of the dice.

We let both companies run all the way to the end. Then we ignore the early rounds — every new setup is clumsy at first — and look only at the settled second half. Whichever team has clearly less translation debt over that stretch wins. If they’re too close to tell apart, we call it an honest tie. We don’t crown a winner on a hair’s difference, because a hair’s difference is noise.

Nothing below is a metaphor — it’s what the engine does, sprint by sprint, and every number is a real setting in the code. A claim runs through a fixed chain of desks — intake, routing, damage estimate, coverage check, approval, payment, closure — across auto claims, commercial underwriting, and subrogation. Each handoff is a seam stamped with two numbers set before the run.

This is why the company leaks value unevenly. The cheap, standardized steps are safe; the expensive human-judgment steps are where a confident-looking AI handoff is most likely to be quietly wrong.

Both teams close the same gappy handoffs — in opposite ways. The difference in the code is small, specific, and decisive.

The whole contest in one line: A is slow, narrow, and permanent; B is fast, broad, and leaky. Which wins depends on the seam mix and the weather — which is what the experiment is built to find.