Ten weeks is long enough to ship a real AI product and short enough that you cannot hide behind a platform program. The constraint is the point. Here is the arc we run at Spinfluence Labs for copilots and bounded agents.
Weeks 1–2 — Discovery, not a model bake-off
Write the job to be done in one sentence. Write the failure you will not tolerate (wrong medical advice, a hallucinated price, an email sent to the wrong customer). List the systems of record you can actually touch. If the data is not there, the product is not an agent — it is a data project, and you should say so in week one.
End with a signed one-pager: user, job, out-of-scope, success metric, and the human approval gates.
Weeks 2–4 — Three directions, one spike
Most AI products fail at the interface. Chat is the default and usually the wrong first surface. We prototype three interaction patterns (inline copilot, structured form-with-draft, task agent with a review queue) and spike the riskiest retrieval or tool path. You pick a lane. We freeze a build spec.
Weeks 4–10 — Build in public, every Friday
A senior pair owns design and engineering. Evals are not a phase; they are how we know a change is not a regression. Observability, prompt and policy versioning, and a boring deployment path are part of “feature complete.” Week ten is a launch runway: who is allowed to use it, how you roll back, what you measure in the first 30 days.
What we refuse to build in v1
- An agent that can “do anything” in your SaaS with no audit trail.
- A chat window bolted onto a PDF dump and called RAG.
- A model fine-tune before you have an eval set and a job.
For the product-vs-internal-tool split, read agents vs. copilots. To start the work, use the AI product practice or the brief.
Continue in the practice: related service.