We have written several pieces on why a national operator should build its own AI talent rather than rent it: the handover plan that left an operator running its own models, the localisation-by-design argument, the case for sovereign subsurface-AI capability. Every one of them made the outcome argument. None of them put a price on the input. This piece does, because at some point a finance director stops asking whether upskilling is a good idea and starts asking what a head of it costs. We answered that question once in a real proposal, and the calculations sheet behind it is specific enough to be worth reading straight.
The proposal priced an enterprise AI upskilling programme by tier. Not a course fee, a per-participant number, and the number moves a lot depending on which tier you are standing up. At the bottom, a junior-analyst tier came in at 26,500 EUR per participant. At the top, a professional-data-scientist tier came in at 146,500 EUR per participant. That is roughly an order of magnitude across the same programme, and the interesting part is not the headline spread. It is that the spread is built, line by line, out of costs you can name.
The number is a stack, not a rate card
The temptation with a training price is to treat it as a rate someone made up and then defend or discount. The calculations sheet does not work that way. It builds the per-head price from cost lines, and each line has a unit you can check against reality.
Start with compute. Every learner needs their own cloud footprint to train on, and the sheet carries a cloud cost of 3,750 EUR per learner at the junior tier, rising to 6,250 EUR per learner at the professional tier. The professional tier trains bigger models on more data for longer, so the compute line grows with the ambition of the tier. This is the same marginal-compute logic that governs the serving side of subsurface work, only pointed at people instead of logs.
Then mentor time, which is where the middle of the number actually lives. The sheet books mentor support at 12,000 EUR for the junior tier and 20,000 EUR for the professional tier, and it is explicit about why: those figures cover 120 to 200 offsite mentor-days at roughly 1,800 EUR per candidate-day. Mentoring is not a fixed overhead you spread thin. It is a day-rate multiplied by a day-count, and the day-count climbs from 120 to 200 as the tier gets more demanding. A professional data scientist is expensive to build mostly because someone senior sits next to them for two-thirds of a working year.
Two more lines round it out. Assessment runs 500 to 1,000 EUR per participant - the cost of actually testing that the capability landed rather than assuming it. And travel is about 11,000 EUR per location, the flights, hotel and meals that on-site delivery requires. Neither is large next to mentor time, but both are real, and leaving them out is how training budgets quietly overrun.
Why the tiers spread by an order of magnitude
Drag the tier lever in the instrument and the per-head total climbs from 26,500 to 146,500 EUR. That climb is not compute alone - compute barely doubles, 3,750 to 6,250. It is the combination: more mentor days, a larger training package, assessment at the top of its band, all stacking at once. A junior analyst needs enough supervised practice to become useful on a defined task. A professional data scientist needs enough to design the task, so the mentor-day count, the compute, and the package each move to their upper bound together. The order-of-magnitude spread is what happens when four cost lines all scale in the same direction at once.
This is the honest version of a claim the outcome pieces made in passing. When we wrote that a national operator should build rather than rent, the build side had a number we never showed. Here it is, and it is not small. A cohort of ten professional-data-scientist heads is not a training line item. It is a seven-figure capability investment with a compute bill, a mentor roster, and a travel schedule attached.
What the price does not include, and why that matters
The per-head number is a programme cost, not a total cost of ownership. It does not carry the learner's salary while they train, the opportunity cost of the mentor's time away from delivery, or the infrastructure the operator keeps after the programme ends. A finance director reading only the 146,500 EUR line would under-count the real commitment, because the expensive part of upskilling a senior practitioner is the months of their own time, not the invoice.
It also does not promise the outcome. The programme prices the inputs - compute, mentoring, assessment, travel - and inputs are not results. The assessment line exists precisely because the sheet does not assume the capability landed; it pays to check. That is the difference between this piece and the ones that came before it. The Omanization and capability-transfer pieces argued that the outcome is worth having. This one prices the ingredients and leaves the outcome where it belongs, downstream of whether the mentoring actually took.
The value of writing the build-up down is that it turns an intimidating headline into a set of decisions a buyer can actually make. You can cut mentor-days and accept a shallower capability. You can co-locate cohorts and share the travel line. You can push more of the compute onto shared infrastructure. Each lever is visible because the price was never a rate card. It was a stack, and stacks can be re-costed line by line.
References
[1] Squicciarini, M., and Nachtigall, H. Demand for AI skills in jobs: Evidence from online job postings. OECD Science, Technology and Industry Working Papers. Documents the labour-market scarcity and wage premium behind enterprise AI-skills demand, the backdrop against which build-versus-rent decisions are priced. https://doi.org/10.1787/3ed32d90-en
[2] World Economic Forum. Future of Jobs Report 2023. Reports employer reskilling and upskilling expectations and the share of the workforce organisations expect to retrain, context for programme-scale capability investment. https://www.weforum.org/reports/the-future-of-jobs-report-2023/
Limitations
The per-participant prices, the cloud-per-learner range, the mentor-support range, the 120-to-200 offsite-day span, the roughly 1,800 EUR candidate-day rate, the assessment band and the per-location travel figure are all taken from a single unpublished training-programme proposal and its embedded calculations sheet. They are one bidder's build-up for one enterprise programme, not an industry benchmark, and the tiers reflect that proposal's role definitions rather than a standard job ladder. The instrument reads only the two tier endpoints and the individual cost-line ranges as sourced; the smooth interpolation between junior-analyst and professional-data-scientist tiers is a reading aid, not a quoted intermediate price. The figures are programme costs and exclude learner salaries, mentor opportunity cost, and retained infrastructure, so they undercount the full commitment of standing up in-house capability.
