FROM THE 2026 ANNUAL REPORT · SECTOR SUMMARY
What the BANKING sector shows
Based on this methodology, banks are the strongest-disclosing sector among those reviewed .The typical posture reaches an evidenced-control standard on both the logical and operational handling of models, reflecting disclosed mechanisms such as model-registration gates, review-and-approval bodies, and technical controls over data exposure. The sector is also the only one in which any institution discloses command of the infrastructure layer — dedicated compute housed in an institution's own facilities — though this remains the exception rather than the category norm and sits at an evolving stage where disclosed. As elsewhere, agentic operation is disclosed at an evolving stage: even where agent activity is extensive, public disclosure describes it as emerging or supervised rather than governed in production.
Heat-map classifications reflect publicly available information reviewed under the methodology described in this report. They are not assessments or certifications of any institution’s actual internal AI capabilities or controls. Grey (Not Disclosed) indicates the absence of public disclosure, not the absence of control.
The report does not evaluate, rate, certify, or benchmark any individual institution; the tiers reflect the completeness of public disclosure as our review found it, not an assessment of any institution's actual controls.
The Central Finding
The Central Finding
Sectors differ in how manyof their institutions can demonstrate control — but far less in where the typicalinstitution sits. The defining pattern is dispersion: within-sector dispersion exceeds between-sector dispersion. The strongest and weakest institutions inside a single sector differ more than the sector medians differ from one another — which is why this report is read sector by sector, and firm by firm, rather than as a single industry verdict.
That dispersion has a structural source. Disclosed control is strongest where duty is retained and the model environment is owned, and thinnest where duty is delegated down a chain of managers, servicers, and providers — fading to non-disclosure at the physical base of the stack, which nearly every institution rents rather than commands.
PUTTING BANKS IN CONTROL OF AI
Before any specific exposure, three questions establish whether the board is looking at control correctly. The first asks whether the banks's AI policies are enforced in the infrastructure or merely written down — the distinction between governance and control. The second and third are the two dimensions every board must hold at once:
• Deep — the full stack. An institution's AI runs on five layers: the agents that act, the models that decide, the data centers that hold the data, the compute it runs on, and the power beneath. Controlling the top layer while renting the four below is not control. The deep question asks which layers the bank has verified, and which it is assuming.
• Wide — the full chain. In financial services the institution rarely operates the AI it relies on. The valuations it books and the decisions it answers for are increasingly produced by AI running inside the managers and servicers it delegates to, on the providers beneath them. The bank holds the fiduciary duty; the delegates hold the AI. The wide question asks whether the bank has verified their control, or delegated the work, kept the liability, and hoped.
Deep and wide are not alternatives. A bank can hold complete control of its own five-layer stack and still fall short of its duty, because most of the AI it depends on runs inside firms it does not operate.