Don't pre-build the AI layer

There is a small subculture of AI development advice telling agencies to do something structurally similar with their codebases. Build the orchestration layer first. Wire the agent runtime. Bolt a tools registry into every service. Be ready for the AI that the client has not yet told you they want, in the shape they have not yet told you it will take.

It is a fine instinct. It is also expensive.

The economic shape of pre-building the AI layer is unflattering once you look at it directly.

An engineering hour at the agency rate carries a margin. Add a wrapper tax that doubles the implementation effort, because every service is being shaped twice (once for the human callers, once for a hypothetical agent) and the margin compresses. At enough endpoints, the margin goes negative. The agency is now paying to be AI-ready on the client's behalf, billing for ten hours of work it is doing in fifteen.

Multiply by Phase 1's typical surface area, and you have built a perfectly polite way to overrun the budget before anyone has talked about Phase 2 at all.

When the real anchor tenant arrives, a Phase 2 LLM use case with a specific operational shape, what does it need from the building it walks into?

It needs utilities. Event logs to embed against. Queues so scoring and downstream calls run async without rewiring everything they touch. Service layers that own their business logic, so a future tool call lands in one place per concept rather than scattered across route handlers. Vector columns already in the schema, so the embedding pipeline has a home on the day it ships. Role checks and audit trails wired in, so the anchor's questions about "show me Factory A's quarter" do not slip across user boundaries.

None of those are capabilities. None of those appear in the AI section of any brief. They are what any capability needs in order to exist at all. They are the building's utilities, not the building's tenants.

A platform refactor recently went through exactly this shape. The team walked thirty-one production templates expecting to rewrite each one. Twenty-three already worked through then new path without a single change. Four needed one property preserved. Three needed a small adapter. A five-week migration shipped as a four-day change. A pre-built orchestration framework would have been four weeks of code adapting templates that needed no adapting. The substrate is what made the work small.

The honest agency move is not to skip Phase 2 thinking. It is to sell what the buyer is actually paying for.

Phase 1 is the operational core. Phase 1 is also the AI prep, because the substrate is what makes Phase 2 cheap. The agency that pours utilities in Phase 1 is the agency that, six months later, ships Phase 2 in weeks instead of quarters. The agency that pre-builds the orchestration framework is the agency that, six months later, rips it out because it does not fit the anchor that finally arrived.

In twelve months, the agencies still standing will be the ones that have done substrate work for at least three clients and ridden the AI wave with each of them when it arrived. The ones that built the orchestration framework upfront will be on their second rewrite, looking for clients who pay for code that gets thrown away. (My track record on twelve-month predictions in this space is mixed; revisit this paragraph in May.)

The mall stays empty until the anchor signs. The utilities keep the building useful until it does.