The two-week look before the six-month build

The biggest commitment in an AI project usually gets made with the least information. Someone sees a promising demo, gets excited, and greenlights a build measured in quarters. The two weeks that would have told them whether the thing can actually reach production never happen, because two weeks feels like a delay. It is the cheapest insurance in the whole project, and teams skip it to save time they then lose many times over.

The demo is not the information you think it is

An engineering leader who scaled large consumer systems described the trap. A demo answers one question. Can the model produce a good output for a chosen input. It stays silent on every question that decides production. Where does the data come from at three in the morning. What happens when the upstream service is slow. Who is allowed to see the output. He watched a team spend two quarters building against a demo that had answered none of that, then discover in month five that the data they needed was locked behind a system nobody had permission to touch.

Real AI feature costs tend to run two to four times the first estimate, and the gap is almost always these unasked questions. A short look answers them while they are still cheap. You do not need to build the feature to learn that the data path is broken. You need two weeks and someone willing to poke at the parts the demo skipped.

Two weeks is enough to find the wall

A founder who runs an AI data-security company made the same case from the security angle. Before he commits engineering time, he spends a short window trying to find the wall the project would hit. Sometimes the wall is a compliance rule that forbids the data from leaving a region. Sometimes it is that the useful version needs access nobody will grant. He would rather find that wall in week one than in month six, and he treats the search as the most valuable part of the whole process.

What a focused two-week look actually produces:

A map of the data path, including where it is dirty, late, or restricted.
A read on the integration seams, the systems the feature must touch and who controls them.
A feasibility verdict, whether this can reach production or hits a wall that changes the plan.
A starter test set, real inputs with expected outcomes the eventual build can run against.

That is not a small deliverable. It is the difference between a build aimed at a real target and a build aimed at a demo. Legacy upkeep already eats 60 to 80% of most IT budgets, and AI features that plug into that legacy inherit its constraints. Two weeks spent finding those constraints reshapes what the six months should even attempt.

How we approach it at Density Labs

The AI Opportunity Assessment is exactly this look, fixed at two weeks and $2,500. We are not trying to build anything in that window. We are trying to learn whether the build is worth starting and, if it is, what it should aim at. Teams come in expecting a green light and sometimes leave with a redirect that saves them a quarter of wasted engineering. Both outcomes are wins, because both replace a guess with a finding.

The two-week look feels like a delay only until you compare it to the six-month build it corrects. Spend the two weeks. It is the cheapest part of the project and the only part that can save the rest.