For a decade, identity in financial services mostly meant human identity: KYC, sanctions screening, customer due diligence. The control frameworks, audit expectations, and examiner questions all assumed a person on the other end.
Autonomous AI agents break that assumption. An agent can now open the position, send the payment, or answer the regulated customer, often with no human in the per-action loop. The identity question does not go away. It moves to the machine.
What the NCCoE is actually doing
The National Cybersecurity Center of Excellence, the applied arm of NIST, turns standards into reference implementations. Its work on machine and non-human identity formalizes the same expectations that govern human access — identification, authorization, auditing, and non-repudiation — for software identities, including AI agents.
This matters because NCCoE outputs often become the vocabulary auditors and examiners adopt. Once a framework is published and a reference build exists, “how do you manage non-human identity?” stops being a research question and starts looking like a control you are expected to evidence.
Why US fintech cannot wait for “an AI law”
There is no single federal AI statute as of 2026, and that fact is often misread as meaning there is no obligation. In reality, the obligations that already bind a US financial institution apply to its agents the moment those agents touch regulated activity.
- GLBA: safeguarding customer financial data, including data an agent reads or writes.
- SOX: controls and auditability over financially material systems, including automated ones.
- SEC recordkeeping obligations such as 17a-4: accurate, tamper-resistant records of activity regardless of whether a human or an agent produced it.
- PCI DSS: where agents handle cardholder data.
None of these frameworks need to mention “AI” explicitly. They still require that you can identify the actor, constrain it, audit it, and prove what it did.
The four pillars, in fintech terms
Identification asks which agent executed a trade, provably and not by inference. Authorization asks what the agent was permitted to do and who delegated that authority. Auditing asks you to reconstruct the incident and show both what the agent did and what it said. Non-repudiation asks for a record a third party can verify without trusting your word for it.
A certificate answers the first. Scope and delegation chains answer the second. Behavioral monitoring plus content-risk scoring answer the third. A WORM log and signed forensic exports answer the fourth.
What to do before the question arrives
- Give each agent a real identity — a per-instance X.509 certificate, not a shared key.
- Bind and enforce scope, and make delegation explicit when agents spawn agents.
- Audit what the agent does and says, and keep the record in a form you can hand over and a third party can verify.
Do that, and the NCCoE’s work stops being something you read about and becomes a box you have already checked — in the US today, and in every other market that maps to the same four pillars.