These essays were supposed to be about architecture. Read them back and notice what they actually did: they kept issuing rules. Approvals are not data; verifications are. Govern by consequence. Do not promote the model; promote the workflow. A consequence that does not change the next run is only an incident, not learning. Rules stated that flatly are not tips. They are articles. And somewhere in the writing, the conclusion stopped being optional: an enterprise that intends to run on AI needs a constitution.
The AI labs reached this conclusion before the rest of us. Anthropic wrote a constitution for its models — an explicit statement of principles the system is trained to hold — on a theory that applies far beyond model training: guidance that lives in no document ends up living nowhere. Now look at the enterprises deploying those models. Most have pilots, a policy scattered across security reviews, and instincts carried by whoever sat through the last vendor briefing. The models have written principles. The institutions running them do not.
Why a constitution, and not another policy? A policy is written for a technology, and this technology changes every quarter; a constitution is written for an institution, and states what holds while everything underneath changes. Policies multiply until no one can hold them; a constitution stays short enough that a manager can carry it into a meeting. And a constitution admits amendment — openly, on evidence — where policy just accretes.
So here is mine: fifteen articles, in three sections. Statements only — the reasoning behind each, and the strongest objections to each, deserve more room than an essay allows. I am publishing them now, before that longer treatment, precisely so they can be argued with.
I. How the Work Runs
Article 1. Outcomes are the unit of value. An AI program is judged by what happens to the work: faster, fewer errors, cheaper, less human effort, and control that holds. Count those. Do not count agents deployed.
Article 2. AI proposes, humans decide, automation executes. Agents read context, handle variation, gather evidence, and prepare. Humans hold the decisions that carry consequence, and the accountability that follows. Automation commits the approved change.
Article 3. Do not spend intelligence on work that does not need it. Wherever work can be written down and checked, it runs as automation — no model in the loop, nothing to pay per run, the same result every time. Models are reserved for judgment, language, and variation.
Article 4. Orchestration holds the work together. Work that spans people, agents, automations, and systems is coordinated by a declared structure that owns progression, state, and recovery. Where the path is known, declare it. Save goal-seeking freedom for outcomes that genuinely outrun paths — and bind it hardest exactly there.
Article 5. Authority is earned by the work, never granted to the model. An agent may assist anyone on supplied context. It receives real authority only over work it has been given the context to know — described, bounded, with written limits — and only as evidence accumulates. Authority shrinks automatically when performance falls.
Article 6. Design for the exception factory. The happy path is what the process chart shows; the exceptions are where the company actually lives. A design that handles only the happy path automates the easy eighty percent and collapses on the twenty percent where the cost lives. Escalation, rollback, and recovery are decided at design time, not discovered in production.
II. What the Institution Builds and Keeps
Article 7. Build the map and the rails as you go. Every AI deployment should leave two things behind: a better map of how the work actually runs, and automations for the parts that proved stable. Build both with the AI, at the same time, from the first project — not as a separate program later.
Article 8. Capture judgment, not conversations. Every interaction with a model either lands attached to the work it belongs to — the process, the case, the decision — or evaporates when the session ends. Capture starts with the first supervised delegation, not with autonomy.
Article 9. Rent the models; own the memory. Never marry a single model — switch by capability, cost, and where your data must live, without rewriting the work. What the institution owns is the memory: the map of its work, the record of its decisions and corrections, and in time models trained on its own validated work, inside its own walls.
Article 10. Learning is governed, not self-modifying. Evidence proposes; a named owner approves; nothing becomes operational without that approval. The record of decisions is append-only.
III. What Happens to the People
Article 11. One program, not two. AI adoption and workforce transformation are the same program. Every deployment decision is a workforce decision; every workforce decision changes what the AI can safely do.
Article 12. Keep both ledgers. The productivity ledger records what got faster and cheaper. The institutional ledger records what was lost, preserved, or rebuilt: trust, mentorship, customer memory, judgment under pressure, culture. Report both.
Article 13. Redesign roles around the work that remains. Every role produces two outputs: the visible work product, which AI increasingly absorbs, and the developmental one it does not — trust built over time, mentorship, customer relationships, judgment in the exceptional case. Redesign roles around the second before deleting them for the first.
Article 14. Rebuild the apprenticeship on purpose. The old apprenticeship — juniors learning the trade by doing the routine work — dies when the routine work goes to machines. Build its replacement deliberately: the map as the training ground, seniors teaching judgment, juniors teaching the machines’ native speed. Do not cut the juniors.
Article 15. Democratization, not anarchy. The people who own the work author their own work — apps, automations, agents, delegated tasks, corrections to the map — inside guardrails the institution defines, and its platform enforces.
The Amendment Clause
A constitution that cannot say what would amend it is a creed. Article 3 rests on a limit I do not expect to fall — exact execution has no acceptable variance at any model capability. Articles 8 through 10 rest on how institutions learn and answer for themselves, which changes more slowly than any technology. The articles most exposed are the people articles — not because they are weak, but because the pressure to ignore them will be strongest exactly when quarterly numbers reward the hollow version. Amendments will be earned by evidence, and made in public.
So argue with it. Tell me which article your institution could not sign, and why — that objection is exactly the debate this document exists to start. The reasoning behind each article, and the strongest cases against them, are coming in a longer form. But do not wait for it: put your institution’s own rules in one short document — what it will always do, and what it will never do — sign them, and amend them in public when the evidence demands it. When this technology settles — and it will settle — the institutions still standing will be the ones that could say, in writing, what they would not do and what they would never stop doing, while everything else changed.
Credit: Source link
