Our approach page carries one promise with money attached: working demo by the end of week two, or we refund the sprint fee. Not slides, not a staging link with mocked data — real code running against real data, in front of the people who will use it.
The refund clause reads like marketing. It is actually an engineering constraint, and it changes how the first ten working days of every engagement are spent.
What it forces on us
A demo deadline this early makes horizontal build plans impossible. There is no time to lay out the data layer, then the services, then the UI, and hope they meet in week five. The only plan that survives is a vertical slice: one thin path from real input to visible output, cut through every layer, ugly everywhere except where it matters.
It also drags the slowest conversations into week one. Data access, VPN credentials, API allowlists, the security questionnaire — on a twelve-week plan these sit comfortably in the backlog until they block someone. With a refund riding on day ten, we ask for credentials in the kickoff call and escalate on day two, not day twenty. The deadline does the nagging for us.
And it forces scope honesty before the first commit. A demo-able unit has to exist on paper at kickoff — that is why the week-1 demo goal is written down before we open an editor. If we cannot name the slice, the scope is not ready, and it is better to discover that in a planning doc than in week four.
What it catches
Almost every expensive engagement failure we have seen announces itself in the first two weeks, if anything is listening. The data is not what everyone believed — the CRM export is missing half the fields the workflow needs. The integration has a wall — the vendor API exists but the client's plan tier does not include it. The workflow is wrong — the demo works and the room goes quiet, because the tool answers a question nobody actually asks on a Tuesday.
On AI work specifically, the week-2 demo catches the most common failure of all: the model is fine and the surrounding system is not. Retrieval that returns the wrong documents, permissions that block the useful ones, latency that makes the feature unusable in the real workflow. A slide deck hides all of this. A working slice against real data cannot.
Why the refund is the point
Without the refund, a missed demo becomes a status-meeting apology and the engagement drifts on. With it, week two is a real gate for both sides: either the slice works and the plan is validated, or the engagement stops while the sunk cost is one-third of a sprint instead of a full quarter. We priced the risk and put it on ourselves, because we are the ones claiming to know how long this takes.
Where the rule does not apply
Some questions are genuinely open — whether the data can support the feature at all, whether to build or buy. Promising a demo there would be dishonest. That is Diagnostic Sprint territory: a fixed week of investigation with a build/no-build memo at the end, and the demo promise starts only when a build begins. The rule is not "always demo in two weeks." It is "never let uncertainty hide inside a build contract."