Short answer. Plan AI as a recurring operating line, not a one-time project cost. Benchmarks put AI near 3% to 5% of the banking technology budget in 2026, and model inference is only 15% to 25% of the real cost. Integration, monitoring, governance, and people carry the rest. Set the number from a shortlist of use cases you can put into production this year, cost the full path for each, and let that total drive the budget.
An AI budget that counts only model API fees under-funds the work that actually puts AI into production. In Ableneo’s 2025 portfolio, 94% of projects used large language models, yet the model call was the smallest line. The budget that ships a system pays for data preparation, integration with core systems, monitoring, human oversight, and change management.
Treat the plan as a portfolio, not a single bet. Across 34 production AI projects delivered in 2025, roughly 4 of 5 reached production. A realistic budget funds the full path to production for the use cases you expect to ship, and caps spend on the ones still proving value.
Repeat business is a planning signal. Ableneo’s average client runs 1.7 projects, so a first production use case usually seeds a second. Budget for the expansion, not only the pilot.
Regulation puts fixed dates in the budget. Under the EU AI Act, high-risk stand-alone systems such as credit scoring and insurance risk assessment and pricing (Annex III) must meet their obligations by 2 December 2027. The conformity work, technical documentation, human oversight, and event logging, carries a cost that belongs in the plan for the year ahead, not in a later surprise.
Supervisors now expect AI governance to be funded, not improvised. The ECB’s supervisory priorities for 2026 to 2028 name banks’ AI strategies, governance, and risk management as a focus area, and the ECB asked banks to have AI-cyber action plans ready by 31 October 2026. DORA already requires operational resilience for the systems AI runs on. A budget that omits these lines is a compliance gap, not a saving.
Plan AI as a recurring operating line. Model inference is only 15% to 25% of the real cost; integration, monitoring, governance, and people carry the rest.
The model is the visible cost and often the smallest one. Independent estimates put base inference at 15% to 25% of the running budget for a production system. The rest is integration with core and data systems, monitoring and drift detection, audit logging, human review, security controls, and the people who run and use the system. A plan built on per-token pricing alone understates the true figure and sets up a mid-year overrun. Cost each use case across all five lines before you commit a number to the board.
Separate one-time cost from recurring cost inside each line. Integration and data preparation are largely one-time. Inference, monitoring, security, and people are recurring and grow with usage. A budget that books the build cost but forgets the run cost looks cheap in year one and breaks in year two, when a successful system carries more traffic and the monitoring and oversight it needs scale with it.
Fund the path to production, not an open-ended pilot phase. The common failure is a budget that pays for many proofs of concept and none of the engineering that puts one into daily use. A better split sets a small, capped amount per pilot with a clear go or stop gate, then concentrates the larger share on taking the winners to production. Ableneo’s 2025 delivery rate, roughly 4 of 5 projects reaching production, comes from that discipline: pilots stay cheap and short, and production is where the money and the return sit.
As a working ratio, keep pilot spend to a minor share of the AI budget and reserve the majority for production and the run cost that follows it. A pilot that cannot name the result it is testing, the data it needs, and the owner who will run it in production does not get funded. That gate is cheaper than a year of proofs of concept that never ship.
Benchmarks are a starting point, not an answer. The Capgemini Research Institute puts AI near 3% of the banking technology budget in 2025, rising toward 5% in 2026. For a mid-market bank, the right number is the one tied to a shortlist of use cases with a measured return, not a percentage copied from a larger peer. Start from the two or three use cases you can put into production in the year, cost the full path for each, and let the total set the budget. A percentage benchmark then becomes a sanity check, not the plan itself.
Lead with a result the board already trusts. A defensible AI budget ties each line to a measured or expected business outcome, not a technology wish. One production reference does more than a forecast. Ableneo’s work with T-Mobile Czech Republic cut document validation time 14x, from 7 hours to 30 minutes, at over 90% accuracy, with a client-approved result. A budget framed as “this spend removes this much time or cost, in production this year” survives a CFO review. A budget framed as “AI capability” does not.
Ableneo plans AI budgets the way it delivers, around the path to production. Across 2025 the team shipped 34 production AI projects for 20 clients in regulated industries, with roughly 4 of 5 reaching production and an average of 1.7 projects per client as first use cases expanded into second ones. That record lets a budget conversation start from measured results instead of estimates. See Ableneo’s AI transformation guidance for how production discipline shapes each stage, from first pilot to funded operating line.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.