Architecture as expensive relic

A comfortable assumption is spreading: models write code, map integrations and produce docs in seconds, so enterprise architecture belongs to a slower era.

“Generation is the cheap part. Deciding what deserves to become permanent is not.”

Concede the true part

Reference diagrams, capability maps, target-state decks, standards nobody read past page four. Agents produce those faster and often better. Architecture staffed to make artifacts should shrink.

Review boards deserved automating

Three weeks to return cosmetic comments on a diagram was already a broken model. That correction was overdue long before anyone heard of a transformer.

Judgment before commitment

What survives is the ability to see across domains, spot when locally rational calls create enterprise consequences, and arbitrate before capital and data are committed.

“It is the only part of the work that cannot be pushed down or coded away.”

Code was never the cost

Software is expensive because you operate it, secure it, staff it, integrate it, govern its data and live with its decisions for a decade. Cheap generation just gets you there faster.

State is what accumulates

Rewriting is cheap, so why worry? Because the data already written, the integrations built against your interface, the regulatory obligations and user habits do not regenerate in an afternoon.

“None of that regenerates in an afternoon, and none of it is in the repository.”

The asymmetry is the point

Implementation gets easier to replace while everything around it stays as persistent as ever. We now create commitments faster than we discover what they cost.

The BI tool I was proud of

Two weeks for a job scoped at three months. Drag-and-drop reports, near zero cost, delighted users. By every measure applied at the time, a win.

Years later: a silo

Jeanne Ross of MIT mapped four sequential stages of architectural maturity. Skipping one just adds another silo, at a speed that looks like progress.

“You do not skip a stage; you pay for each one in accumulated foundational discipline.”

Nobody asked why

The analysis was being done by hand in spreadsheets. He killed the spreadsheets without asking whether the data model could ever answer the question the business was really asking.

“I wrapped the gap in two weeks of custom code and made it harder to see.”

The bill arrived later

When the business wanted longitudinal reporting and cross-domain aggregation, the tool integrated with nothing, scaled to nothing, and had to be rebuilt from scratch.

The question one level out

Not how do we solve this well, but is this the right problem, for whom, and what must the answer become in three years. An hour on that changes the decision.

“It is a deliberate widening of the frame before commitment, and it is worth nothing unless it happens before.”

Narrow window, wide window

A squad sees its own domain in high resolution. The enterprise frame is the inverse: an enormous window at low resolution. Both are needed, and no organization gets the second one free.

Fifty teams, all succeeding

Marketing ships a vector store. Logistics ships routing. Customer service bridges platforms. Every demo lands, and every initiative makes sense inside its own boundary.

Then the seams appear

Competing definitions of the same customer. Sensitive data in unclassified storage. Duplicate capabilities. Inference costs nobody saw until adoption.

“Enterprise incoherence does not require bad local decisions.”

The paradox ahead

Cheaper production means more independent decisions, more interfaces, more data dependencies. The bottleneck moved from building to deciding coherently.

“The cheaper software becomes to create, the more expensive incoherence becomes.”

Give the agent a wider frame

Fair objection, and models do challenge assumptions well. But whoever builds the frame already made the decisive call: this request deserves to be implemented.

An excellent pipeline, wrong problem

Ask for a pipeline between two systems that disagree and you get a clean, tested, permanent one. It may just automate a symptom and institutionalize the dysfunction underneath.

Analysis is not standing

Telling a funded sponsor their request cements a dysfunction is not an analytical act. And when a pipeline leaks records or fails an audit, someone still has to sign.

“Someone owns the decision. Someone signs.”

Three sources of authority

Own the paved road so the governed path is the easiest one. Hold a seat where capital is allocated, before funding closes the option. Carry named, signed risk.

“If going around your road is easier than using it, you do not have a road.”