Ten-layer diagrams of the AI stack are accurate and slightly misleading, because they draw every layer the same size. Building a compliance agent that now handles most of our cases end to end, the first four took a couple of weeks. The rest took the year.
Embeddings, a vector store, retrieval, a model. That’s the demo, and it demos well enough that people start asking about a rollout date.
Then the actual work starts.
Tools. The moment it can act instead of answer, every capability needs a contract, a failure mode and an idempotency story. Charging someone twice because the response timed out isn’t a model problem.
Autonomy limits. We capped ours deliberately. Anything touching money, legal standing or an angry human routes to a person. That costs a better-looking automation number and buys an audit trail.
Memory. Easy to add, hard to bound. State that carries across sessions is state that can be wrong across sessions.
Security. Prompt injection stops being theoretical the first time the input is a customer email nobody on your team wrote.
Evals. The one I’d build first if I started over. Thousands of real scenarios running as regression tests, because an agent you can’t measure is a demo, not a feature.
The layers aren’t equal. Four of them get you something to show. The other six decide whether you can leave it running.