Strategy as Code is not the point

Last month, a director at a Nordic AI company posted a thought on LinkedIn: Strategy as Code. A strategy.md file for every agent, every repo, owned by a director or VP. Agents read it. Agents question it against implementation reality. Gaps surface as they appear. The post got a hundred reactions and sixty comments, most of them practitioners describing what they were already doing. A lead product designer scoring discovery ideas against strategy.md. A CEO maintaining a company-ai-context repository as a single source of truth. An engineer with axioms.md scored against design docs. A governance architect encoding business strategy, missions, and policies into agentic workflows.¹

Earlier this spring, BCG's Henderson Institute published a nine-page analysis concluding that more than 80% of strategy function tasks face high or medium AI exposure, and that the first half of 2025 saw more strategy-related AI tools launched than the previous two years combined. Their core finding: "Harnessing AI to alter the system of strategy development itself will be the true differentiator."² The World Economic Forum and Bain put it more plainly: the strategy machinery must spin faster.³

The instinct is catching. Strategy should be structured, machine-readable, present where decisions happen. That much is settled.

The substrate question

What is not settled is what the strategy lives in.

The simplest answer is a file. A markdown document per team, per repo, per agent. Versioned, reviewable, delivered to AI tools. Several companies are now building exactly this: scan your Confluence, Notion, GitHub, Slack, and Jira, find the real decisions, detect contradictions, deliver the governed context to every tool via MCP. The pitch is clean and the problem is real. Teams using different AI tools are building from different versions of reality. A text layer that harmonises the decisions solves that.

These practitioners got there first because their tools already support it. Decisions live in repos, context lives in docs, agents read both. That's a genuine head start, and the instinct is exactly right.

But what happens when the organization is larger than what any collection of documents can hold?

A single strategic choice in a 500-person enterprise fans out across dozens of objectives, each owned by a different team, each depending on other teams for delivery, each measured by different KPIs, each supported by different vendor investments. The connections between these entities are the substance. A markdown file can describe a decision. It cannot hold the dependency chain that determines whether that decision produces an outcome. The dependency chain is not text. It is structure: typed, directional, queryable.

Two substrates

The substrate is forming in layers.

The first is settling: text. Decisions captured, versioned, governed, delivered to AI tools. The moat is the scan: how many sources you index, how well you detect contradictions, how fast you propagate updates. The AI reads documents and produces better-informed documents. When a better model arrives, the documents get better. Intelligence resets on every query.

Documents alone don't ground the model. Researchers at Esade, Imperial, and NYU ran 15,000 trials feeding industrial context to seven leading LLMs; the bias in their strategic recommendations shifted by only about 11%, with the models still defaulting to differentiation, augmentation, long-term, collaboration, regardless of the situation in front of them.⁷

The next layer is harder: a graph. Strategic choices connected to objectives, objectives connected to teams, teams connected to each other through dependencies, all measured by KPIs that trace back to the choice that justified them. The moat is the graph itself. A year of use produces a denser graph than anything a competitor can offer on day one. When a better model arrives, the traversals get deeper. The AI diagnoses: tracing a KPI drop through the dependency chain to the team whose delayed commitment caused it, and recommending what to do about it. The graph sits above how work actually flows, and each query writes back into it. Intelligence accumulates.⁴

Both layers are forming, at different points in the picture.

What is settling

The question of whether strategy needs a machine-readable substrate is over. BCG says it. The practitioners say it. Multiple venture-backed teams are building it. Salesforce rebuilt its entire platform as APIs, MCP tools, and CLI commands so agents can access everything humans can. Their reasoning: both humans and agents need the same data, the same workflows, the same trust layer.⁵ Weeks later, SAP published its own Autonomous Enterprise architecture, with a knowledge graph at the centre⁸ and a named substrate for the operational knowledge the organization runs on.⁶ The surface changes. The platform doesn't.

The substrate is forming in layers. The question is what each one is built from.