What a two-week AI assessment actually produces
The first step of assess, pilot, ship gets the least respect and does the most work. Here is what comes out of it.
Most teams asking about AI already have a list. Twelve ideas, sometimes thirty, gathered from an offsite or a vendor deck. The list is not the problem. The problem is that nothing on it has been costed, and so the one that gets funded is whichever one the loudest person described most vividly.
Two weeks is enough to fix that. Not enough to build anything — enough to know what to build, and what it costs to be wrong.
Week one: subtract
We spend the first week removing candidates. An idea comes off the list when the data it needs does not exist, when the workflow it improves is not actually painful, or when a deterministic system would do the job for a tenth of the money. That last one takes more ideas off the list than anything else.
What survives has three properties: a person whose day gets measurably better, data already sitting somewhere we can reach, and a failure mode nobody has to apologize for in writing.
Week two: cost it
The survivors get priced. Model and inference cost at realistic volume, integration work, the evaluation harness, and the ongoing cost of someone owning it after we leave. That last line is the one most estimates skip, and it is usually the largest.
We also write down the accuracy the use case actually needs. “As good as possible” is not a target; “wrong less than two percent of the time, and visibly unsure the rest” is. A number makes the pilot falsifiable.
What lands on the last day
- A ranked shortlist — usually three — with the reasoning for every rejection kept.
- A cost model per candidate, including the year-two run cost.
- The accuracy bar each one has to clear, and how we would measure it.
- A pilot scope for the top candidate: what gets built, in what order, and what proves it.
- The list of things we found that are not AI problems at all.
That last item surprises people. It is also frequently the most valuable page in the document, because it is the one that saves a quarter.
When to skip it
If you already know the use case, the data is in hand, and someone senior owns the outcome, skip the assessment and start the pilot. Paying for a two-week study to confirm a decision you have already made is theater. We will tell you that on the first call.
The assessment is step one of how our AI consulting practice works: assess, pilot, ship, and hand over so your team can run it without us.
Written by Tommy Shrove, BluuAlpha Technologies.