The demo hides most of the system

A convincing prototype can be one prompt, one model call, and a polished interface. A production feature has to work with real permissions, incomplete data, slow dependencies, unexpected inputs, changing models, and people who do not know how the system was designed.

That gap is why many AI projects feel almost finished for weeks. The visible intelligence is present, but the ordinary engineering that makes it safe and repeatable is still missing.

Seven layers around the model

I use seven questions to expose the real scope before a team commits to implementation.

  • Workflow: what outcome is the user trying to reach, and where should AI make a decision?
  • Context: which data may the model see, how current is it, and who is allowed to access it?
  • Tools: which actions can the system take, and what contracts do those APIs enforce?
  • Validation: how is model output converted into something the application can trust?
  • Evaluation: which examples represent acceptable, dangerous, and ambiguous behavior?
  • Operations: what is logged, measured, retried, timed out, or escalated to a person?
  • Economics: what do latency and model usage cost at realistic volume?

Start with the workflow, not the framework

Framework selection feels concrete, but it is downstream of the business process. First describe the job as a sequence of decisions, data lookups, actions, and approval points. Deterministic steps should stay deterministic. Reserve model judgment for the places where ambiguity is real and useful.

This also produces a better first milestone: one bounded workflow, connected to real data and tested against representative cases. It reveals integration and quality risks without creating a platform before the team understands the product.

Dependability is a product feature

Users do not experience your architecture diagram. They experience whether the feature is predictable, explains failure, respects access, and hands control back when confidence is low. Validation, fallbacks, approval gates, and observability are therefore part of the product — not invisible backend polish.

The most valuable AI integration work often looks like senior software engineering because that is what converts model capability into an accountable system.