Workflow before model.
Most AI projects lose the plot before model selection. The useful question is simpler: what decision improves, who trusts the output, what can fail, and who owns the result?
A model is not a workflow.
A stronger model can improve an answer. It cannot repair an unclear job, missing data, weak handoff, vague success metric, or customer experience nobody owns.
The architecture work starts before provider selection. It starts with the human decision, repeated task, trusted inputs, review point, cost constraint, and failure mode.
That is the difference between AI theater and useful AI: the system has a job, a boundary, and an operating owner.
The first four questions.
Name the moment where AI helps a person decide, respond, summarize, remember, learn, escalate, or refuse.
Define the trigger, inputs, output surface, handoff, and what happens when confidence is low.
Separate verified data, retrieved context, model judgment, user edits, and final responsibility.
Track latency, token cost, review effort, support load, conversion, retention, margin, and risk reduction.
Give AI a job description.
What sources are allowed, what must be retrieved, what is stale, and what should never enter the prompt.
Whether AI drafts, explains, ranks, routes, summarizes, recommends, escalates, or refuses.
Where the user reviews, corrects, approves, or takes back control before the system affects a customer.
The metric that proves value: time saved, error reduced, lead captured, issue resolved, or decision improved.
The same rule shows up everywhere.
The AI is useful only when it helps a reader remember, review, compare, and ask better questions about a book.
Personal finance AI must distinguish explanation, recommendation, user action, and regulated or high-risk decisions.
A food marketplace needs reliable ordering, subscriptions, pickup, delivery, and chef tools before AI can add leverage.
A voice check-in system needs consent, careful memory handling, escalation rules, and family confidence before automation matters.
Before choosing a model, write this page.
Write one page with seven lines: user, repeated job, trusted inputs, AI responsibility, review point, failure path, and business metric.
If those lines are weak, the model choice will not save the product. If those lines are strong, model selection becomes an engineering decision instead of a strategy substitute.
This is the architecture layer I care about: useful AI that earns trust inside real work.