Most AI project budgets are built around the wrong number. Companies fixate on model or API costs, which are usually the smallest line item, and systematically underestimate the three that dominate: the people who build the system, the data work required before any model is useful, and the operating cost of keeping the thing alive after launch. A realistic AI budget is mostly a talent budget with a data-preparation surcharge, and any calculation that does not reflect that will be wrong in the expensive direction.
The four cost blocks of every AI project
- Talent, the dominant block: engineers, data specialists and product time, whether on payroll, on day rates or embedded. On typical applied projects this is the majority of total cost, and on LLM-based projects the share is often even higher because infrastructure is rented rather than built.
- Infrastructure and compute: cloud, GPUs or model API usage, vector stores, environments. Meaningful, growing with scale, but at pilot and single-feature scale usually a clear second to talent.
- Data preparation, the chronically underestimated block: discovering, accessing, cleaning and restructuring the data the project actually needs. Practitioner experience consistently puts a large share of total project effort here, and it lands as engineering hours, which means it lands in the talent budget.
- Operations after launch: monitoring, evaluating quality drift, retraining or re-prompting, adapting to model deprecations and API changes. A production AI system is a running commitment, not a finished deliverable.
Rough cost ranges by project type
Ranges below are indicative orders of magnitude for the DACH market, assuming senior-level staffing at market rates; individual projects vary widely with scope, data condition and integration depth. Their purpose is to calibrate expectations, not to replace a scoped estimate.
| Project type | Typical duration | Typical team | Indicative build cost |
|---|---|---|---|
| Pilot / proof of value | 6-12 weeks | 1-2 people | EUR 30k-100k |
| Production feature (e.g. RAG assistant, forecasting) | 3-6 months | 3-5 people | EUR 150k-500k |
| Platform / multiple use cases | 6-18 months | 6-10+ people | EUR 500k-2M+ |
| Ongoing operations (any of the above) | per year | fractional to 2 FTE | commonly 15-30% of build cost/year |
The hidden line items that break budgets
- Integration with legacy systems: the model works in a notebook in week four, and then months go into connecting it to the ERP, the identity system and processes designed in 2009. On Mittelstand projects this is routinely the largest unplanned item.
- Change management and enablement: training the people whose workflow changes, redesigning the process around the tool, handling the dip in productivity before the gain. Skipping it converts a working system into shelfware.
- Evaluation and quality assurance: building test sets, human review loops and acceptance criteria for probabilistic systems takes real engineering time that deterministic-software budgets never included.
- Compliance and governance: GDPR reviews, the EU AI Act's requirements where they apply, works-council involvement, security review of vendors, weeks of calendar time even when the direct cost is modest.
- The cost of vacancy: every month a key AI role stays unfilled, the burn continues while the deliverable slips, slow staffing is a real cost even though no invoice ever names it.
How the staffing model shifts the calculation
The same project can carry very different cost curves depending on how it is staffed. Full-time hires have the lowest steady-state monthly cost but the highest commitment: recruiting takes months, and the cost continues after the project peak passes. Freelancers invert this: day rates for senior AI talent look expensive against salaries, but you pay only for the phase you need, with no severance risk and no idle cost, for a six-month build with an uncertain follow-on, the flexible model frequently wins on total cost despite the higher rate. Embedded experts sit between: a premium over payroll, in exchange for immediate start, pre-vetted capability and, done properly, deliberate knowledge transfer to your own people, which is the part that keeps paying after the engagement ends. The honest comparison is never day rate versus monthly salary; it is total cost over the project's real duration, including recruiting time, idle risk and what remains in the building afterwards.
A simple method for a defensible budget
- 1Scope one use case, not the AI vision: a budget for "AI in operations" is fiction, a budget for "automated intake-document triage" can be estimated.
- 2Price the team first: define the phase-by-phase team shape, then apply market salaries or day rates, this anchors the dominant block before anything else.
- 3Add a data-honesty buffer: have someone technical inspect the actual source data before the budget is final, and size the data-preparation line from what they find rather than from hope.
- 4Budget operations from day one: put a yearly operating line, commonly 15-30% of build cost, into the plan before approval, an AI feature without an operations budget is a future outage.
- 5Stress-test with the hidden items: legacy integration, change management, compliance lead times, if the budget only survives when all three go perfectly, it is not a budget.
