Short answer. Allocate AI costs the way you allocate any other shared IT service: start with showback for 4 to 6 weeks to see who is actually spending, then move to chargeback once at least 80% of spend is attributed to a real owner. Banks that skip straight to chargeback without a cost model usually end up disputing invoices instead of governing usage. A production AI program at a mid-market bank typically runs dozens of models and agents in parallel, so cost visibility by team, model, and use case has to exist before governance can.
Most banks start their AI program with one pilot and one invoice. That changes fast once the pilot works. Ableneo shipped 34 production AI projects in 2025 across 20 clients, and 94% of those projects use large language models, which means token spend, not headcount, becomes the variable cost line finance has never had to model before. Once a bank runs five or six AI systems in production at once, a single shared cloud or model bill stops answering the only question the CFO actually asks: which business unit is generating this cost, and is it generating value back.
In practice, cost allocation shows up in three recurring situations:
None of these are technology problems. They are governance problems that show up as technology bills.
Banks and insurers cannot treat AI spend the way a retail company treats a marketing SaaS subscription. Under DORA, financial entities have had to maintain a register of ICT third-party dependencies and assess concentration risk since January 2025. An AI system with unclear cost ownership is also, almost by definition, a system with unclear operational ownership, and that is exactly what a DORA audit tests for. If nobody can say what a model costs to run per business unit, nobody can credibly say who is accountable for it either.
The EU AI Act adds a second pressure point. GPAI obligations and Article 50 transparency duties have applied since 2 August 2026, and high-risk standalone systems under Annex III, including credit scoring and insurance risk pricing, have to comply by 2 December 2027. Conformity assessment, human oversight, and logging all carry real operational cost. A bank that has not built a cost model for its AI estate will discover these compliance costs as a surprise line item instead of a budgeted one, at the exact moment a supervisor is asking questions.
Start with showback for 4 to 6 weeks, then move to chargeback once attribution covers roughly 80% of AI spend.
Showback reports estimated cost back to the team that generated it without moving a single euro between cost centers. Chargeback posts that cost as a real, binding entry against the business unit’s budget. Showback is the right starting point because it builds the habit of looking at usage data before anyone has to defend it in a budget meeting. Most organizations run showback for 4 to 6 weeks, until cost attribution covers the majority of total AI spend, then switch to chargeback once coverage clears roughly 80%. Skipping showback and jumping straight to chargeback is the single most common reason AI cost allocation programs fail in year one: business units dispute numbers they have never seen before.
A usable AI cost model joins three data sources: the model or API provider’s billing data, request-level telemetry (who called the model, for what feature, how many tokens), and the business context (which product, which customer segment, which use case). Without the third layer, a bank can see that it spent 40,000 euros on inference last month but not whether that spend produced fewer manual underwriting hours or just more failed retries. Build the model at the level of team, environment, model, and use case from day one, even while running showback, so the eventual move to chargeback does not require rebuilding the plumbing.
Cost allocation and compliance governance are the same discipline seen from two directions. A DORA-compliant register of ICT dependencies needs to know which AI systems are running and who depends on them, which is exactly the ownership map a chargeback model requires. An EU AI Act risk classification exercise needs to know which business unit deployed which model for which purpose, which is exactly the attribution layer a cost model requires. Banks that build the cost model and the compliance register as two separate projects end up maintaining two versions of the same ownership data. Build them once, from the same source of truth.
Ownership tends to split unevenly if nobody assigns it deliberately: engineering and platform teams often hold the most influence over the decisions that drive cost, while a separate finance or FinOps function is left accountable for the number without control over the usage that created it. The workable model puts a single accountable owner, usually inside a FinAIOps or platform engineering function reporting into IT finance, in charge of the cost model and the allocation policy, with business unit leads accountable for their own usage once they can see it. Neither side should own it alone.
Start with the highest-spend AI system already in production, not a new pilot. Instrument it for team, model, and use case level telemetry, run showback for one full budget cycle, and only then decide whether to move that one system to chargeback. Trying to build an allocation model across the entire AI estate on day one is why most attempts stall before they produce a single usable report.
Ableneo builds AI systems for banking and insurance clients with the assumption that they will still be running, and still need to be paid for, a year after go-live. That means cost telemetry, usage attribution, and DORA-aligned ownership mapping get designed into the architecture during the pilot phase, not bolted on after finance asks for the first invoice explanation. Ableneo’s AI transformation work for regulated clients pairs this cost governance layer with the same production discipline behind the 4 of 5 delivery rate across 34 AI projects shipped in 2025, so a bank scaling from one AI use case to ten never loses the ability to answer who is spending what, and why.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.