/ A. Pozo

The case for funding architecture

Code got cheap. Being right didn't.

AI collapsed the cost of producing code. It did nothing to the cost of producing the wrong system. This page is about that gap — and why the person who controls it is the highest-leverage thing you can fund.

AI-generated · 0.4s
// discount enginefunction applyDiscount(total, pct) {  return total - pct;}
Plausible. Compiles. Wrong.

The shift

AI scaled output. It never scaled coherence.

Drag the slider. As AI capability grows, the volume of code an organisation can produce explodes. Whether that code forms a system that holds together does not move on its own — a person moves it.

Code volume Architectural coherence
RISK 0 AI capability →
58%

The distance between these two lines is risk. It is paid later, with interest.

Specification

The same request. Two completely different systems.

A model is a function of its input. Give it a vague prompt and it fills the gaps with guesses. Give it a precise specification and it fills them with your intent. Toggle the input below.

input

“Build me an orders module.”

Tangled. No boundaries. The bugs are structural — you cannot test them away.

input

Order is an aggregate. It cannot ship below available stock. Totals are derived from lines — never stored raw. Cancellation after fulfilment is forbidden. Money is an explicit value object.

Bounded. Invariants enforced at the type level. The model resists misuse by construction.

Sales Inventory Billing

Spec-driven domain design

The specification is the asset. The code is a build artefact.

When the spec is the source of truth, code can be regenerated — but the thinking behind it cannot. Activate each layer of the specification and watch the architecture compile.

Domain Sales Billing Inventory
spec → architecture

An empty spec compiles to nothing you can trust.

The irreducible

Four things a model cannot decide for you.

The compounding cost

Architectural debt does not add up. It multiplies.

Scroll through eighteen months of the same project. One path had an architect shaping every boundary. The other shipped whatever generated fastest.

With architectural discipline Without it
Month 0 Month 18 Cost of change ↑

Both teams used the same AI. Only one stayed cheap to change.

The leverage equation

AI is a multiplier. A multiplier needs something worth multiplying.

Trustworthy output is AI throughput multiplied by human judgment. Throughput is now effectively unlimited. Drag judgment toward zero and watch what unlimited throughput is worth.

AI throughput
×
78
Architectural judgment
=
78
Trustworthy output

Unlimited speed times zero judgment is still zero. This is the case for funding the judgment.

The decision

Fund the person who makes the AI worth the spend.

I design domain-driven backends — DDD, Hexagonal, CQRS — for systems where being wrong is expensive. If you're investing in AI-accelerated delivery and want it to compound instead of decay, let's talk.