Most AI readiness assessments sold as standalone engagements run six to eight weeks, bill somewhere between €40,000 and €120,000, and end in a scored framework your own engineers could have produced in a fortnight. The alternative costs roughly a quarter of that: put two senior engineers on a thin vertical slice of the real use case for ten to fifteen working days and judge readiness by what breaks.
This memo sets out what an assessment is genuinely testing, what each shape costs, and the cases where the long advisory version is the correct purchase anyway.
Readiness is four questions, not forty
Scoring frameworks with thirty dimensions produce a number like 2.7 out of 5. Nobody can act on that. Four questions decide whether an AI build reaches production, and all four are answerable by writing code against your actual systems:
- Can an engineer get authorised, repeatable, programmatic access to the source data inside this quarter? Access approval is the most common place pilots stall, and it stalls for months, not weeks.
- Is the data captured at the grain the use case needs? A churn model needs event-level records. If the warehouse only holds daily aggregates, the project is a data engineering project first and a modelling project second.
- Does a lawful basis exist for the processing, and does it survive the data leaving your tenancy? For EU personal data going to a hosted model endpoint, that question has a specific answer and somebody in legal has to give it.
- Who operates the thing on day 91? Name the team and the rota. An unnamed owner means the pilot is the deliverable.
Two shapes, compared
| Advisory assessment | Engineering probe | |
|---|---|---|
| Duration | 6 to 8 weeks | 2 to 3 weeks |
| Staffing | Part-time partner, two analysts | Two senior engineers, full time |
| Evidence base | Interviews and workshops | Code run against production-like data |
| Output | Scored framework, roadmap, business case | Working slice, plus a written blocker list |
| Indicative cost | €40k to €120k | €15k to €30k at European senior day rates |
| Residual value | Deck ages in a quarter | Code becomes the first increment |
| Estimate quality | Derived from what people said | Derived from observed throughput |
The cost gap is not a discount. It is a difference in what gets bought: six weeks of senior advisory time with a brand signature on the cover, against four to six engineer-weeks. Current European senior day rates for AI and data engineers sit on our live day rate page, and you can model a build increment with the project cost calculator.
What the probe should hand back
A two-week probe that cost you €20,000 and produced only a verdict was badly run. It should return: one end-to-end path from source system to a rendered output, however ugly; measured retrieval or inference latency against real volumes, not synthetic ones; a count of records that failed validation and why; and a named list of access, licensing and legal blockers with the person who can clear each one.
That last item is what makes the next estimate defensible. Most AI cost overruns trace back to data access and data quality discovered in month three rather than week two, which is also the pattern behind the drivers in what actually moves AI development cost.
Where renting engineers is the wrong answer
Three cases, honestly:
Your board or regulator wants an independent opinion with liability attached. Rented engineers give you evidence, not an indemnified signature. If the point of the exercise is defensible governance, buy the firm whose professional indemnity cover is the product.
You are prioritising across eight business units with no single use case chosen. That is portfolio analysis, and engineers are the wrong instrument. Pick the use case first, then probe it.
You have no internal engineer who will own the result. A probe hands back findings that someone has to act on. Without an owner, you get a better-evidenced version of the same shelf-ware.
Three questions before you sign either one
- What will you have written or run by day 10, and against which system?
- If the answer turns out to be “not ready”, what does the invoice look like and does the work stop?
- Who on your side holds the data access decision, and have they agreed to be available in week one?
A supplier who cannot answer the first question with a system name is selling interviews. If the use case is already clear and you want engineers on it rather than a roadmap about it, tell us what you are building and we will scope the slice.