The Short Answer
A bank can run AI agents on infrastructure it owns rather than rents. On ibl.ai you own all the code and the data, so KYC files, transaction histories, and customer records are processed inside the bank's own perimeter and every tool call is logged where auditors already look. The platform is model-agnostic and has no per-seat pricing, so you can deploy anywhere, including air-gapped.
The ambition is now standard. Banks describe AI "workforces" — agents handling know-your-customer checks, onboarding, transaction monitoring, adverse-media screening, and fraud detection.
The delivery is not. Capgemini Research Institute's World Cloud Report in Financial Services 2026, surveying 1,100 leaders across 14 markets, found that only 10% of financial institutions have deployed AI agents at scale, while 80% remain in ideation or pilot.
The same research shows where banks most want them: customer service (75%), fraud detection (64%), loan processing (61%), and customer onboarding (59%).
Every one of those four touches regulated customer data. That is not a coincidence, and it is most of the explanation for the 10%.
Why do bank agent pilots stall before production?
Not because the agents fail. Because the review does.
A fraud-detection agent that reaches production has to answer questions a demo never asks: where does the transaction data go while the agent reasons over it, who else can reach it, how long is it retained, what is logged, and can that log be handed to an examiner.
When the agent runs in a vendor's cloud, each answer is a contractual assurance rather than an architectural fact. That distinction is survivable for a pilot on synthetic data and becomes the sticking point the moment real customer records are involved.
So the pilot works, the risk committee asks where the data goes, and the programme returns to ideation. Ten percent.
What does SEC, FINRA, SOX, and PCI DSS actually demand of an AI agent?
Not that you avoid AI. That you can evidence control over it.
Under SEC and FINRA supervision obligations, a firm must be able to show what its systems did and why. SOX demands auditable controls over anything touching financial reporting. PCI DSS constrains where cardholder data may be processed at all.
An agent complicates each because it acts without a human in the loop. The reconstruction of what happened is the log, and the log's completeness is a property of the platform, not the model.
This is why the ownership question is a compliance question rather than a philosophical one. If the platform is yours, the audit trail lives beside your existing access logs, retention follows your policy, and there is no third party whose incident response you inherit.
If it is not, you are evidencing someone else's controls.
We worked through the deployment options in detail in Private AI for Financial Services, on Your Own Servers and the vendor comparison in Self-Hosted AI vs ChatGPT for Financial Services.
Does an AI workforce make sense under per-seat pricing?
No, and the mismatch is structural rather than a matter of negotiation.
Per-seat licensing — roughly $30 per user per month for Microsoft Copilot, about $60 for ChatGPT Enterprise — prices a human at a keyboard. An AI workforce is by definition not that.
An agent screening transactions overnight occupies no seat, and a bank that wants a hundred of them is not asking for a hundred more employees.
Vertical tools price even more sharply against scale: legal AI has been reported at $300–500 per lawyer per month, which is the same shape applied to a smaller population.
The alternative is to pay for the work rather than the headcount, or to own the platform outright and pay only for the compute.
At a bank with 40,000 staff and an ambition to put agents into four core processes, the pricing model is not a line item — it decides whether the programme is affordable at all. The arithmetic is laid out in AI Cost Math for Financial Services.
What should a bank actually own?
Three layers, in order of how hard they are to replace later:
The data path. Where KYC files, transaction records, and adverse-media results travel while an agent reasons over them. This is the layer regulators ask about first and the hardest to change after the fact.
The audit trail. Every prompt, retrieval, and tool call, in a form an examiner accepts, retained under the bank's own policy.
The model choice. Not because any one model is best, but because a bank that cannot change models has accepted a single vendor's pricing, deprecation schedule, and capability curve as a permanent input to a regulated process.
None of that requires building an AI platform from scratch, which is its own multi-year trap. It requires a platform the bank runs and holds the code for. The distinction between owning and renting is set out in Renting Enterprise AI Costs Far More Than the Invoice.
Where does this leave the 90%?
With a solvable problem, and a specific first question.
The banks in the 10% did not find better models. They resolved where the agents run before scaling them, which meant the risk review happened once at the architecture level instead of separately for every use case.
That is the practical sequencing: decide the data path and the audit trail first, then add agents against it. Doing it the other way — proving value in a pilot and negotiating the architecture afterwards — is precisely how eighty percent of a market ends up stuck in ideation.