Short answer. A proof of concept for a single banking AI use case typically costs between EUR 15,000 and EUR 60,000 and runs 6 to 12 weeks. Moving that same use case into production usually multiplies the figure 5 to 10 times, because the cost sits in integration to the core system, model risk controls, and monitoring, not in the model. The EU AI Act adds a further 6 to 10 percent to the development cost of a high-risk system, and a third-party conformity assessment runs EUR 10,000 to EUR 40,000 per system. The number that decides the budget is everything the model has to plug into.
The cost of an AI system in a bank splits into three layers, and only one of them is the model. The first layer is the build: the pilot that proves the use case works on the bank’s own data. The second is integration: wiring the model into a core banking platform that is often 15 or 20 years old, plus the security review, the model risk sign-off, and the monitoring stack. The third is the run: inference cost, retraining, human oversight, and the compliance evidence that never stops. Banks that budget only for the first layer are the ones whose projects stall.
Ableneo shipped 34 production AI projects in 2025, and roughly 4 of 5 reached production. That number matters to cost more than any price list, because a pilot that never ships is 100 percent wasted budget. Across those projects, 94 percent used large language models, so the run-cost question is real, not hypothetical.
A realistic cost picture for one production use case looks like this:
In financial services, most high-value AI use cases are regulated by default. Credit scoring, creditworthiness assessment, and fraud decisioning fall under Annex III of the EU AI Act as high-risk systems, which triggers conformity assessment, technical documentation, logging, and human oversight before the system goes live. Those obligations are not optional line items. The European Commission’s own impact assessment put the compliance uplift at 6 to 10 percent of the development cost of a high-risk system, and that is before the recurring documentation burden.
DORA has been in force since 17 January 2025, and it adds a second layer of cost that many AI budgets miss: the Register of Information, incident reporting readiness, and resilience testing all apply the moment an AI model supports a critical or important function. For a mid-market bank in Central Europe with EUR 50M to EUR 500M in revenue, a EUR 1 million surprise on a single use case is not a rounding error. The ECB found that 38 percent of euro area firms are already at an advanced stage of AI adoption, so the buyers who estimate cost accurately now are the ones who scale without a budget freeze later.
A banking AI proof of concept costs EUR 15,000 to EUR 60,000, and production typically costs 5 to 10 times more.
Four factors move the number, and the model API is usually the smallest of them. Data readiness comes first: if the data is scattered, poorly labelled, or locked in a legacy core, the work to make it usable can exceed the model work several times over. Integration is second: every system the model reads from or writes to adds connection work, and connecting to existing systems can add EUR 30,000 to EUR 200,000 depending on documentation quality. Governance is third: model risk management, conformity assessment, and audit trails are engineering work, not paperwork. People are fourth: the specialists who build, validate, and operate the system are the largest recurring line.
Put together, the model license or token spend is often less than 20 percent of the two-year total. Banks that negotiate hard on the model price and wave through the integration scope are optimizing the wrong number.
A proof of concept is cheap and fast on purpose: EUR 15,000 to EUR 60,000 over 6 to 12 weeks buys evidence that the use case works on real data. Production is where the figure jumps 5 to 10 times, and the jump is predictable. It pays for integration to the core, a security and model risk review, a monitoring and alerting stack, incident procedures, and change management for the people who will use the system. Published cost breakdowns understate the total by 30 to 50 percent, because those lifecycle costs only appear after the pilot succeeds.
The practical lesson is to treat the pilot as the first 15 percent of the spend, not a preview of it. A bank that approves a pilot without a production budget behind it is building a demo, and a demo that cannot ship returns nothing.
Compliance cost is recurring, and that is what budgets miss. The EU AI Act conformity assessment lands at EUR 10,000 to EUR 40,000 per high-risk system, but the technical documentation, logging, and post-market monitoring continue for the life of the system. DORA adds the Register of Information, submitted annually, plus the incident desk and resilience testing that a critical AI function now falls inside. Each of these has a first-year setup cost and an annual carrying cost, and both scale with the number of AI systems in production. A bank running one AI use case can absorb this by hand. A bank running ten needs a governance operating model, or the per-system cost compounds.
Run cost is the line that surprises finance teams, because it did not exist for classic software. For an LLM-based system it has four parts: inference or token spend that scales with query volume, the monitoring and observability stack, periodic retraining or fine-tuning, and the named human oversight the EU AI Act requires for high-risk systems. Inference alone can move from a few hundred euros a month for a light internal tool to tens of thousands for a high-volume customer-facing system. This is why cost governance, sometimes called FinAIOps, matters: without a per-use-case view of run cost, a bank cannot tell a EUR 500-a-month assistant from a EUR 30,000-a-month one until the invoice arrives.
Estimate across the whole lifecycle, not the pilot. Score each use case on four axes: data readiness, number of systems it integrates with, regulatory tier under the EU AI Act, and expected query volume. A low-risk internal assistant with clean data and one integration is a small, fast project. A high-risk credit decisioning model on a 20-year-old core is a large, multi-quarter one, and pricing them the same is how budgets break. The strongest cost control is the production rate itself: at roughly 4 of 5 projects reaching production, the money spent on pilots is money that ships, which is a different economics from a portfolio where most pilots are sunk.
Ableneo has shipped 34 production AI projects across four countries, with roughly 80 percent reaching production and an average client running 1.7 projects. That track record in regulated financial services is the real cost lever, because the largest hidden cost in bank AI is the pilot that never ships. Ableneo scopes build, integration, run, and compliance as one number from the start, so the production budget is visible before the pilot begins. See Ableneo’s AI transformation work, and for the return side of the same equation, how to build a board-ready business case for AI in a bank.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.