Most AI projects don't fail on technology. They fail on scope: a vague "AI-powered" brief that means everything to everyone, growing until the timeline collapses. Our typical engagement is about 14 weeks, and hitting that consistently comes down to how the first two weeks go.
Week one is ruthless narrowing. We force one sentence: "The system does X for user Y so that Z." If Z isn't measurable — hours saved, turnaround time, error rate — the project isn't scoped yet. Then we list everything the system will explicitly NOT do. The not-doing list is the most important document in the project. AI invites scope creep because "can't it also..." is always one prompt away.
Week two is the slice. We pick the smallest end-to-end slice that delivers real value and can be demoed to a skeptical user: real data, real interface, one complete workflow. Not a model benchmark, not a UI mockup — a working slice. This does two things: it proves the riskiest technical assumption early (usually data quality or integration, never the model), and it gives stakeholders something concrete to react to instead of slideware.
Then we sequence by risk, not by logic. The standard mistake is building the easy 80% first and discovering in week 10 that the hard 20% is impossible. We build the scariest part first — the integration with the legacy system, the accuracy bar on messy real documents — while there's still time to change course.
We also fix the eval criteria before building. For each capability, we write down what "good enough" means and how we'll measure it, agreed with the client. Without that, launch criteria drift forever and the project never ships — it just gets "one more iteration."
Fourteen weeks isn't magic; it's what happens when scope is a decision made in week one instead of a negotiation in week twelve. That's how we run AI product delivery at QuantaloomAI: fixed scope, fixed timeline, measurable outcomes.
Field note: A real week-one deliverable from a recent engagement: one sentence ("The system triages inbound vendor invoices for the finance team so that approval time drops from 5 days to 1"), a not-doing list with 11 items (no PO matching, no multi-currency, no mobile app — all requested, all deferred), and three named risks with owners. The client pushed back on two of the not-doing items; we negotiated one back in and priced the change explicitly. That negotiation — in week one, with numbers — is the whole game. Compare it with the alternative: discovering the disagreement in week ten, when it's a schedule crisis instead of a pricing conversation.