How much of a retail AI project is estimate and how much is discovery
In a first retail AI project, a large part of the timeline is not estimate. It is discovery. You do not know the real data state until you are inside it. Whoever promises a single number, as if everything were estimable, blows up when discovery arrives.
The project starts with the right hypothesis: the problem is clear, the technology exists, the timeline seems calculable. What the team does not measure before starting is how much of the path is still unknown.
The part that is estimate
The estimable part is known engineering. Integrating with a documented system, training a model on clean data, shipping an interface. That is quoted with reasonable confidence, within a range.
If the whole project were that, the timeline would be honest and short. It is almost never the whole project.
The part that is discovery
Discovery is everything you only learn by looking at the real data. How many months of history actually exist, clean. Whether shelf-out is flagged or hidden in the sales. How many codes the same product has. Where the registry lies.
In a first project, discovery is usually 30% to 50% of the time, and it is invisible in the proposal. The vendor who ignores it quotes only the estimable part, wins the bid on the shortest timeline, and later “discovers” what was always there.
Why discovery becomes delay, not learning
When discovery is not budgeted, it shows up as delay and starts a fight. The client thinks they were misled; the vendor thinks the data was worse than agreed. Both are right, and the project joins the projects that fail.
The honest move is to separate the two parts in the timeline. Discovery becomes its own phase, short and cheap, before the timeline commitment. That is the Diagnóstico: discovery done on purpose, upfront, instead of mid-project as a surprise.
The criterion before the timeline
The number that decides is not “when is it ready”. It is “what we still do not know about the data, and how long it costs to find out”. Whoever answers that in week one quotes a timeline that holds. Whoever does not quotes a fiction.
Take the AI project on your roadmap and the timeline you were given. Ask how much of it is undone discovery. If the answer is “the data is probably fine”, you bought an estimate on ground nobody looked at.
Send me the problem and the timeline on the table. In one hour I will send back where the hidden discovery sits in that timeline, and which phase has to come before the commitment. If the shape confirms, the two-week Diagnóstico is discovery done cheap and upfront, and it ends with a timeline that separates estimate from what is still a question.