What happens to my company when its AI model changes? Model succession in 3 steps

Published July 12, 2026 · 25 United Capital · También en español

If your company runs on AI agents, you have a problem that hasn't hurt you yet: models expire. The one operating your business today will be replaced by a better one within months — or discontinued. And when that happens, the question won't be "is the new model smarter?" (it will be), but: does your company survive the handover without losing its standard, its memory and its judgment?

If your operation lives in a chat's memory, you don't have a company: you have a conversation. Conversations aren't inherited.

We solved this and we PROVED it: two blind exams, verified and on record, in which a model with zero context took over our technical chief's role and made the right calls on real cases. This playbook is the 3 steps of that method. (Scope honesty: this solves MODEL succession. The succession of the human in charge is a different problem, with different tools — we treat it separately.)

Step 1 — Write the role, not the agent

The root mistake is treating the agent as the asset. The asset is the ROLE: its identity, its method, its decision rules and its learned cases — written outside the agent, in documents any future model can cold-load.

Our technical-chief/verifier role lives in a single-piece Successor's Manual: who you are in the company (and what you are NOT: a people-pleaser) · the method in one sentence with its corollaries · the bank of real cases with each one's lesson · the quick decision rubrics (when to escalate to the human, when to stop, how to score) · how to work with OUR specific CEO, not an abstract one · and the exact cold-boot order: what to read, in what sequence, upon waking.

The key almost everyone skips: the manual includes the INCIDENTS with their lesson — not just the rules. Rules without their scars are advice; with them, they are transferable experience.

Apply it tomorrow: open a document called "Manual for role X". First section: "who you are and what your standard is". Second: "the mistakes we already made and their rule". Third: "what you read at boot, in what order". With that you already have more succession than 99% of AI companies.

Violation smell: onboarding a new agent means "read the chat history".

Step 2 — Blind-test the successor against real cases

An untested manual is a hope. The test: give a model ZERO context — only the manual — and put your REAL hard cases from the past in front of it. If it decides well, the role transfers. If it decides badly, you fix the manual, not the candidate, and repeat.

The receipt (two exams, on record): in the first (July 5, 2026), a blind model, armed only with the manual, resolved 3 of 3 real cases from our history — it re-detected a validator the executor had edited to approve itself, blocked a fake installer by checking its signature against the official source, and told the CEO an uncomfortable truth without dressing it up. The exam's finding: one wrong environmental assumption in the manual — THE MANUAL was corrected. In the second (July 6, 2026), another blind model, given only our operating constitution and statute, resolved 3 of 3 citing exact articles — and detected, unprompted, something we hadn't asked: that the very text wasn't ratified yet and therefore didn't rule. The successor applied the law TO the document containing it. That's when we knew the system thinks, rather than recites.

Apply it tomorrow: save your 3 worst agent incidents (you have them). New session, zero history, only the manual: present them as if they were happening today. Compare its decisions with what good judgment would have done. Every candidate failure is a paragraph your manual is missing.

Violation smell: "the new model is really good, it will surely adapt". Adaptation without an exam is faith — and faith with your company inside is the illusion of capability.

Step 3 — The handover rite: succession as law, not as emergency

Succession is not improvised the day the model changes: it is a rite with its own law, executed on EVERY handover — planned or not. Ours: (a) a boot order written BEFORE the handover (what the successor loads, in what order, which laws bind it, which decisions it must NOT touch without the human); (b) an operating state with a freshness gate — the successor first checks whether the state is recent and coherent, and if it isn't, DECLARES it before acting; (c) a mandatory handover declaration — the successor recites what it recovered: active mission, next action, constraints; if it can't recite it, recovery is incomplete and it doesn't work; (d) a dry rehearsal before the real handover.

The receipt: the rite is a chapter of our operating statute (every handover runs it). The dry rehearsal of our latest handover ran on July 11, 2026 scoring 4 of 4, with the successor's boot order written and verified BEFORE the change; and the formal rehearsal of the real handover, one day later, scored 5 of 5 — and caught two outdated documents, corrected that same day. The exam improves the document, always. The first exam's conclusion was written into its record: "handover day is not a cliff; it is a relay."

Apply it tomorrow: write your main agent's boot order TODAY, while the current model is alive — it is infinitely easier to write it with the incumbent in the seat than to reconstruct it without them. And rehearse it once in a clean session before you need it.

Violation smell: treating a model handover as a software update ("I'll point to the new model and done"). What changes is not a version: it's the employee who knew everything.

What this method buys you (and what it doesn't)

It buys you: freedom of provider and version — riding every improvement of the AI frontier without fear, because your company lives in its documents, not in a model's memory · continuity of standard (the successor inherits the judgment, not just the data) · and a VERIFIABLE proof of maturity almost no one can exhibit.

It does not buy you: the human's succession (that is will, vault and law — another playbook) · nor does it exempt you from verifying the successor during its first real weeks: the blind exam is the entrance door, not the graduation.

Why you can trust this (and how to verify it)

Because it isn't a framework designed on a whiteboard: it is the mechanism by which our own company — a non-technical human in charge, AI operating everything — has already survived its model handovers, with dated records, public cryptographic fingerprints of its law documents in the Receipts Room, and building in public with truth labels. The method can be copied for free (this playbook IS the copy); the track record backing it cannot.

Frequently asked questions

What happens to my company when the AI model that operates it changes? If the role is written outside the agent and blind-tested, nothing: the successor loads the documents and continues. If it lives in a chat's memory, you start from zero.

How do I transfer knowledge from one AI agent to another? Don't transfer the chat: write the ROLE (a manual with identity, method, cases with their lesson and cold-boot order).

How do I prove a new model can replace the current one? Blind exam: zero context, only the manual, against your real past incidents. If it fails, fix the manual and repeat.

What is a model handover rite? Pre-written boot order + freshness gate on the state + mandatory declaration of what was recovered + dry rehearsal. On every handover, no exceptions.

Does this require technical skills? No: documents and discipline. A non-technical founder can run the exam by comparing decisions against what actually happened.

25 United Capital · one human directs, AI operates, written rules govern — built in public at 25united.com.