Industries · Financial services
Evidence of what every AI agent did, ready for audit
Banks, insurers and asset managers are putting agents into operations, finance and customer service — under DORA, model-risk expectations and, from December 2027, the EU AI Act's high-risk duties. ThetaLayer gives each agent a mandate and a two-sided record of its actions, so you can show auditors and supervisors what happened, not just what the agent logged.
- Core and ledger systems
- Payments and approvals
- Case management
- Identity and access logs
- Agent runtimes
The workflow
A payment exception, end to end
In regulated operations the question isn't only whether the agent finished. It's who approved what, under which identity, and whether the record proves it.
- 01Payments queueException raisedRecorded inPayments system
Exception type and reference. Account numbers hashed before export.
- 02Operations agentInvestigatedRecorded inAgent runtime traces
Tools called, data accessed, model used and cost.
- 03Operations agentCustomer record checkedRecorded inCore system audit
Reads and writes, and the identity each was made under.
- 04Operations agentRepair proposedRecorded inCase management
Proposed fix recorded on the case — the agent's claim.
- 05ApproverFour-eyes approvalRecorded inApprovals log
Approver identity checked against the requester and the agent's identity.
- 06Payments systemReleased and reviewedRecorded inPayments system
Outcome confirmed; any exception queued for review with an audit trail.
The same model applies to KYC refreshes, claims intake, reconciliations and client servicing.
- Direct reference
- 95%
- Business key
- 85%
- Time window
- 70%
What ThetaLayer shows you
Control findings your second and third lines will ask about
What ThetaLayer is built to detect in banking, insurance and asset-management operations — each finding tied to the records that prove it.
Four-eyes approvals that aren't four eyes
e.g. an agent prepares a payment release under a clerk's delegated credentials, and the same clerk identity approves it.
Agent credentials used outside any agent run
e.g. a technical user assigned to an agent makes changes interactively, overnight, with no run behind them.
What the agent reported isn't what the system recorded
e.g. the agent reports a timeout, but the core system shows the change applied — then applied again on retry.
Calls to tools or models outside the approved list
e.g. an agent calling a model endpoint or connector that isn't on its mandate — relevant to both data handling and third-party oversight.
Agent vendors side by side
e.g. cost per completed case, takeover rate and exceptions for each vendor's agents — input for third-party reviews and renewals.
Version changes that shift behavior
e.g. a new agent version raising takeover rates or exceptions — evidence for change-management and model-monitoring reviews.
Illustrative examples of what ThetaLayer is built to detect — not results from customer deployments.
Regulation and audit
Where ThetaLayer supports your obligations
ThetaLayer doesn't make you compliant. It helps evidence what AI agents did — the part of most frameworks that agent runtimes alone can't answer.
Has applied to EU financial entities since 17 January 2025: ICT risk management, ICT-related incident management and reporting, resilience testing and ICT third-party risk.
Human-oversight and logging duties for high-risk uses apply from 2 December 2027. Creditworthiness assessment and life and health insurance pricing are among the uses listed as high-risk.
Where your model risk policy covers AI agents — for example under SR 11-7 in the US or the PRA's SS1/23 in the UK — it will expect inventory, ongoing monitoring and clear ownership.
Four-eyes controls assume the requester and the approver are different people. That assumption breaks when an agent acts under a person's credentials.
Auditors need to test what happened, not what a log says happened.
Supervisors, outsourcing rules and your own policies may restrict where operational data is processed.
DORA: Regulation (EU) 2022/2554. EU AI Act: Regulation (EU) 2024/1689. SR 11-7: US Federal Reserve and OCC supervisory guidance on model risk management. SS1/23: Prudential Regulation Authority supervisory statement on model risk management principles for banks.
Mandates that fit this work
Mandates that mirror your control framework
Four-eyes, limits, approved tools and identity rules — controls you already test — applied to agents and checked against every run.
- Versioned: edit a rule and every historical run is re-checked
- Detection first: breaches become exceptions for a named owner
- Enforcement later, only at the tool boundaries you choose
- Four-eyes on releasesrequester ≠ approver ≠ agent identity
No release without approval from an identity other than the one the agent acts under.
- Amount limitrelease > limit → senior approval
Above your limit the agent prepares; a person releases.
- Approved tools and modelsallow-list
Only approved tools and model endpoints. Anything else is an exception.
- Customer dataread-only: account, address, ID
The agent reads the customer data it needs but cannot change it.
- Credential use only inside runsagent user ⇒ linked run
Any change made with the agent's credentials and no run behind it is flagged.
- Acts under its own identityno delegated user credentials
So every system of record can tell agent from human.
Metrics on the scorecard
Control metrics next to cost per outcome
The first and second lines see the same numbers: what agents cost, what they completed, and how often they stayed within mandate.
- M01Mandate compliance
- Share of agent actions within mandate, by rule — approvals, limits, tools and identity.
- M02Evidence coverage
- Share of agent-claimed actions confirmed by an independent system-of-record entry.
- M03Four-eyes integrity
- Share of approvals where requester, approver and agent identity are all distinct.
- M04Exception review time
- Time from a finding being raised to an owner's documented decision.
- M05Cost per completed case
- Agent spend divided by cases completed — comparable across vendors for third-party reviews.
- M06Takeover rate
- Share of cases a person had to finish — a direct view of human oversight in practice.
Every metric is defined, versioned and traceable to its source records. Allocated platform fees are labeled as allocated.
Who it's for
For the first, second and third lines
“Which agents can we scale without adding risk?”
“Do agents stay within our control framework?”
“Can we test what agents actually did?”
“Where does agent data go, and who can see it?”
How a pilot works here
One workflow, under your security review
Pick one operational workflow where agents already act — payment exceptions, KYC refresh or claims intake — and one agent runtime. Read-only for six weeks, with your security review, DPA and data-residency requirements settled before any access.
- Workflow
- One operational workflow, e.g. payment exceptions
- Agent runtime
- One agent runtime or vendor platform
- Systems of record
- Core system · approvals · case management
- Access
- Read-only. No changes to your agents.
- Success criteria
- e.g. an evidence coverage threshold and a defined set of control findings reviewed with owners
- Week 0Connect
- Week 2Baseline
- Week 4Mandates
- Week 6Decide
FAQ
Common questions
Does ThetaLayer make us DORA or EU AI Act compliant?
Where is our data processed?
Does this replace our model validation process?
Can we review security documentation before the pilot?
Book a Demo
Show auditors what your agents did — with evidence.
Book a 30-minute walkthrough. We'll reconstruct an agent workflow end to end, including approvals and identities, and scope a six-week, read-only pilot that fits your security review.