AI agent inventory for banks

Every conversation about AI governance in a bank starts with the same question: what AI do you actually have? Supervisors ask it, auditors ask it, and the board should ask it too.

The answer they expect is an inventory. Not a slide with the ten projects the data science team knows about, but a complete and current record of every AI system, model and agent, each with an owner. This article sets out what supervisors and auditors have published on this, the fields worth keeping, and why agents make the job harder.

What supervisors say

ECB (euro area). In February 2026 the ECB's representative to the Supervisory Board said more than 85% of large banks under European supervision already use AI in some form. He named "fragmented ownership" as a recurring concern: responsibility split across IT, data science, business lines and control functions without a clear accountability framework. He asked for clear accountability for AI-driven decisions, assessment before material use cases go live, and monitoring and reporting afterwards [1]. You cannot do any of that without knowing what is in use.

PRA (United Kingdom). Supervisory Statement SS1/23 applies to UK banks with internal model approval and covers AI used in models. Principle 1.2 says: "A comprehensive model inventory should be maintained." It should cover models in use, under development and decommissioned, firm-wide, and record each model's purpose and use [2]. Principle 2 asks firms to give responsibility for the model risk framework to a named senior manager [2].

MAS (Singapore). The Guidelines on AI Risk Management issued on 7 October 2026 apply to all financial institutions and all forms of AI. They expect firms to "identify their AI use, maintain inventories at an appropriate level of granularity" and assess the risk materiality of each use case. They take effect on 7 October 2027 [3]. In its response to the consultation, MAS addressed two hard cases: AI embedded in third-party services, and how often to update. It chose not to prescribe a fixed frequency, but said updates may be triggered by events such as material changes to a use case or the systems behind it [4].

CBUAE (United Arab Emirates). The Central Bank's Guidance Note on AI, issued on 11 February 2026, asks licensed financial institutions for an inventory of "all AI models, systems or technologies developed or deployed", with at least the name, purpose and risk rating, including systems developed or hosted by third parties. It also asks for the "clear and immediate ability, with human intervention, to cease use" of an AI system [5].

EU AI Act and DORA. The AI Act has no single article titled "inventory" for deployers. But from 2 December 2027 a bank must know which of its systems are high-risk under Annex III, who oversees them and where their logs are kept [6]. DORA, which has applied since January 2025, already requires a register of information on contractual arrangements with ICT third-party service providers [7]. A list of AI systems helps with both. It does not replace the DORA register.

What auditors ask

Audit firms have published similar expectations.

PwC US tells clients that "if AI is part of your process, it becomes part of the audit". It recommends "a complete and accurate inventory" that links each use case to business processes, flags regulatory relevance and records the assessed risk level. It adds that the inventory "shouldn't rely solely on self-reporting": firms should also look for shadow AI, AI embedded in third-party tools and AI features added through system updates [8]. As evidence that controls actually operated, PwC lists items such as validation sign-offs, approval workflows, exception logs, samples of human review and records of overrides [8].

The external auditor's lens is narrower. Writing for ICAEW in September 2026, Grant Thornton's Jon Morris says the auditor's question is where AI produces information that feeds the financial statements, for example an expected credit loss model. For those uses the auditor asks who owns the model, how it was validated, what checks cover its inputs and outputs, and how it is monitored [9]. A bank's inventory is how it answers the first question quickly and shows that the rest of its AI is out of scope.

The fields worth keeping

Drawing these sources together, an AI inventory in a bank should hold at least:

Field	Why
Unique ID, name and type (agent, model, tool, AI-enabled system)	One record per asset; no duplicates
Platform and provider	Links to the DORA register of information [7]
Named owner and business unit	ECB accountability; SS1/23 Principle 2 [1, 2]
Purpose and business process	SS1/23 1.2; CBUAE minimum fields [2, 5]
Risk rating and AI Act classification, with reasoning	MAS risk materiality; Annex III [3, 6]
Personal data and data sensitivity	GDPR; PwC metadata [8]
Human oversight model and the people assigned	AI Act Art. 26(2); CBUAE [5, 6]
Approval status, approver and date	Evidence that controls operated [8]
Connections: models, tools and data an agent uses	Change impact; concentration risk
Status: planned, in use, retired	SS1/23 covers decommissioned models too [2]
History of every change, viewable as of a date	Audit sampling; supervisory questions about the past

Why agents make this harder

Traditional models come from a small number of teams with a validation process. Agents do not. Staff can build them on low-code platforms, connect them to tools and data, and change them without a release cycle. Three problems follow.

1. Discovery. Self-reporting misses agents, which is why PwC warns against relying on it [8]. Platform vendors have noticed: Microsoft made its Agent 365 registry, a tenant-wide view of agents, generally available on 1 May 2026 [10]. That helps for one platform. Most banks use more than one.
2. Change. An agent's model, tools or permissions can change in an afternoon. MAS's event-triggered updates [4] fit agents better than an annual review.
3. Ownership. The person who built an agent is often not the person accountable for the process it serves. The ECB's "fragmented ownership" concern applies here first [1].

Keeping the inventory current

• Find, then confirm. Read the platforms where agents are built, and ask each owner to confirm what was found. Do not overwrite what an owner declared with what a connector read.
• Load what you already have. Most banks have spreadsheets. Start from them, check them row by row, and do not run a second list in parallel.
• Route by risk. A minor change should take a light path; a high-risk change goes to risk, compliance and data protection.
• Separate requester and approver, and keep each decision with the facts and policy version it relied on.
• Keep history. Supervisors and auditors ask about the past. Be able to show the register as it stood on any date.

Where Signoria fits

Signoria is a register built for this job. It holds the AI agents, models and tools a regulated institution connects, each with a named owner, approvals and an audit ledger, and it runs in the institution's own infrastructure. Read-only connectors find agents on Microsoft Copilot Studio and Amazon Bedrock and ask owners to confirm. Existing inventories load from Excel or CSV. The register can be viewed as it stood on any date, and the approver is never the requester.

It does not monitor agent traffic and does not replace your GRC tool. It is the system of record that the rest of your governance reads from.

Request the free 90-day pilot and build the inventory on your own estate.

Sources

1. Pedro Machado, ECB, "Technology is neutral, governance is not: AI adoption in the banking sector", 24 February 2026. https://www.bankingsupervision.europa.eu/press/speeches/date/2026/html/ssm.sp260224~6c5b64a77a.en.html
2. Bank of England, PRA Supervisory Statement SS1/23 "Model risk management principles for banks", current version effective 23 April 2026, Principles 1.2 and 2. https://www.bankofengland.co.uk/prudential-regulation/publication/2023/may/model-risk-management-principles-for-banks-ss
3. Monetary Authority of Singapore, "MAS Sets Out Supervisory Expectations on Responsible AI Adoption by Financial Institutions", 7 October 2026. https://www.mas.gov.sg/news/media-releases/2026/mas-sets-out-supervisory-expectations-on-responsible-ai-adoption-by-financial-institutions
4. MAS, "Response to Feedback Received on Guidelines on Artificial Intelligence Risk Management", section 6. https://www.mas.gov.sg/-/media/mas-media-library/publications/consultations/bd/2025/response-to-feedback-received-on-guidelines-on-artificial-intelligence-risk-management.pdf
5. Central Bank of the UAE, "Guidance Note on the Consumer Protection and Responsible Adoption and Use of Artificial Intelligence and Machine Learning by Licensed Financial Institutions in the U.A.E.", issued 11 February 2026, §§2(f), 6(f), 9(c). https://rulebook.centralbank.ae/en/rulebook/guidance-note-consumer-protection-and-responsible-adoption-and-use-artificial-intelligence
6. AI Act, Articles 26 and 113 and Annex III (consolidated text as at 27 July 2026), AI Act Service Desk. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-26
7. Regulation (EU) 2022/2554 (DORA), Article 28(3). https://eur-lex.europa.eu/eli/reg/2022/2554/oj/eng
8. PwC US, "Responsible AI: Now that AI is part of your business, what does it mean for audits?" https://www.pwc.com/us/en/tech-effect/ai-analytics/responsible-ai-audits.html
9. Jon Morris (Grant Thornton UK), "What does client use of AI mean for the audit?", ICAEW, 24 September 2026. https://www.icaew.com/technical/audit-and-assurance/faculty-resources/audit-and-beyond/2026/what-does-client-use-of-ai-mean-for-the-audit
10. Microsoft Security Blog, "Microsoft Agent 365, now generally available, expands capabilities and integrations", 1 May 2026. https://www.microsoft.com/en-us/security/blog/2026/05/01/microsoft-agent-365-now-generally-available-expands-capabilities-and-integrations/

This article explains regulatory expectations in general terms. It is not legal advice.