Humans hold the gates

Ask what makes an AI-native strategy firm different and most people will point at the model. The model matters. It reads faster than any analyst, drafts options in minutes and argues against itself without getting tired. But a board does not adopt a strategy because it was fast to write. It adopts a strategy because someone it trusts stands behind it.

That is why the most important part of strategy1 is not the model. It is the signature. Every decision we help a client make passes a named person before it moves. Not a committee in general, not “the team”, but a person whose name is on the approval record and who can be asked, later, why they signed.

A model can draft a recommendation in minutes. It cannot be accountable for it. Accountability is the part of strategy work that does not compress, and we do not try to compress it. We try to compress everything around it.

Our work runs on the S1 Method: six stages in a loop, not a line. Sense, Model, Simulate, Decide, Deploy and Learn. AI does the heavy lifting between the stages. Humans hold the gates. Nothing advances past a gate without a signed approval record.1Show source for note 1

There are three gates, and each one asks a different person a different question.

• H1 Model. An ontology engineer confirms the structure of the Strategy Twin, the living model of the client's market, company and options, before any option is built on it. If the model is wrong, everything built on it is wrong, so this is the first signature.
• H2 Decide. The Strategy Council signs and, where liability exists, a Domain Fellow signs too. Then the client chooses. The output is a signed decision memo, not a deck.
• H3 Deploy. The client's named owner approves a dry-run of every change before anything is written to a client system.

After Deploy comes Learn. KPIs are tracked against the decision that set them, and when results drift, the scenarios re-run and the Council sees a proposed next decision with its evidence. The loop returns to Sense. The gates do not disappear on the second pass. They are there every time.

Claude is the AI model made by Anthropic. We use it for what it does well: reading documents and public sources, tagging each claim with its source, drafting options, red-teaming them and explaining the result in plain language.

We do not use it to calculate. Every calculation, from NPV and IRR to break-even and Monte Carlo runs, happens in our own deterministic code.2Show source for note 2 The LLM explains; it does not calculate. That rule matters for a reason that has nothing to do with whether the model is good at arithmetic. A number from code can be tested and re-run. A number from a language model can only be trusted or not.

The same thinking applies to facts. The narrative cannot add a fact that is not in the evidence ledger.3Show source for note 3 Every number and claim links to a source, a date and the agent that produced it, and every key figure carries a confidence grade: A for a verified primary source or client data, B for a credible secondary source, C for a model estimate.4Show source for note 4 A C-grade figure is labelled as an estimate wherever it appears.

None of this replaces judgement. It makes judgement checkable. When a Strategy Partner signs a decision memo, they are signing something they can trace, line by line, to its basis.

A gate that anyone can pass is not a gate. The value of a signature comes from the person behind it: their authority to change the recommendation, their responsibility if it is wrong, and their willingness to say no.

So our gates name people. A strategy1 pod is one Strategy Partner and Forward Deployed Strategists, supported by ontology and AI engineers. The Strategy Partner owns the client's trust and signs the decision memo. Domain Fellows, who are chartered engineers, CAs, lawyers and sector veterans, review and sign the parts that carry professional liability.5Show source for note 5 They are paid per review.

This is also where we draw a line on what we claim. strategy1 provides decision support and blueprints; licensed professionals sign technical, financial and legal conclusions. A model can structure a problem a specialist would recognise. It does not hold a licence, and we never describe its work as if it did.

The hardest gate to hold is the last one. Strategy that runs has to reach the systems where a business actually operates: CRM opportunities, budgets, briefs and run-sheets. That is exactly where an AI system acting on its own would do the most damage.

Before any write, the change passes a policy check, the client's owner approves a dry-run of exactly what will change, and the approval is recorded in a hash-linked audit trail.6Show source for note 6 Policy check, human approval, audit record. In that order, every time. Approved decisions become actions only after the named owner accepts them.

We are careful with language here. We never describe a write as automatic when a person approves it. If a page, a deck or a sales conversation implies autonomy where an approval exists, it is wrong, and we correct it.

The test of a gate is whether it answers the questions a board will ask. Who approved this? What was it based on? How sure are we? How does it fail?

Each has an answer on the record. The approver is named at the gate. The basis is in the evidence ledger. The confidence is in the grade beside each figure. And the failure modes are in the red-team report, because every strategy ships with its strongest counter-arguments and the conditions under which it fails. “How this fails” is a required section, not an appendix.

Humans hold the gates; AI does the heavy lifting between them. That is not a slogan about caution. It is a description of where the time goes. The reading, drafting, modelling and arguing get faster. The signing stays deliberate. We think that is the right trade, and we would rather show it on one real decision than argue it in a proposal.