Banks Use AI Agents to Catch Compliance Drift Earlier

Watch more: Digital Shift With Flagright’s Chris Phillips

Financial institutions have spent years stalking the risks and opportunities that artificial intelligence brings to their security-critical space. Then came AI agents.

These agents could give banks something their existing financial crime operations struggle to produce: investigations conducted consistently, documented exhaustively and audited continuously rather than months later through a small sample of cases.

“The AI doesn’t get bored. It doesn’t get fatigued. It doesn’t forget to include that link for the one search it did,” Christopher Phillips, director of financial crime industry engagement at Flagright, told PYMNTS. The result, he said, can be “a robust output that is easy to audit.”

That turns the conventional risk equation around. More automation does not necessarily mean less control. Properly designed, more autonomy could create more visibility into how financial crime decisions are made. Traditional financial crime quality assurance operates largely through sampling. Compliance or audit teams might examine a relevant percentage of cases and extrapolate findings across the broader program.

Agentic AI, however, makes another model possible.

“With AI, you can do all of them,” Phillips said. “You could literally do a QA or an audit of every single case.”

“What we need to do is holistically rethink the investigative process, beginning to end,” he said, including documenting the workflow and identifying where controls belong throughout it.

The AI Agent Is Becoming a Bank Employee

As financial institutions move AI from recommending actions toward executing them, much of the debate has focused on what machines should be allowed to do without human approval. But in the agentic era of AI, a new question is emerging: What happens when AI makes every investigation observable, reproducible and auditable?

“The technology is mature,” Phillips said. “Out of the box, if done right, if applied correctly, it is a game changer.”

The implication is bigger than productivity. Artificial intelligence could move financial crime operations from sample-based supervision toward continuous control, provided banks resist the temptation to treat agents as another piece of software bolted onto an existing compliance stack. Maturity does not mean autonomy without boundaries.

“If you’re going to close an account, if you’re going to make a recommendation for SAR filing, if you’re going to take any action against the customer that has consequences, then that needs to be reviewed by a person,” Phillips said.

The emerging operating principle is one where machines can investigate, search, identify patterns and assemble evidence, but consequential decisions require a different standard. Humans need to provide what Phillips calls “effective challenge,” or understanding why an agent reached its conclusion and documenting why they agreed or disagreed.

“AI agents are basically employees,” Phillips said. “They need to be monitored like employees and need to have strictures like employees.”

Controls Are Becoming the Agentic Financial Crime Investigation Accelerator

Consider something as simple as investigators overriding an agent’s recommendation. A small disagreement rate might be normal. But if overrides move from 5% toward 10% or 15%, the institution suddenly has a measurable signal that something may have changed. Perhaps customer behavior shifted. Perhaps underlying data drifted. Perhaps the model requires retraining. The important point is that the institution does not need to wait months for an audit sample to discover it.

“You can see that quickly and you can retrain the model quickly,” Phillips said, noting that in an agentic environment, audit stops being primarily retrospective and becomes part of the operating system.

“We tend to think of controls as kind of regulatory blockers,” he said. “It’s not the right way to think about controls. Controls help make everything better as well.”

Governance procedures can establish who retrains a model, when intervention becomes necessary and who approves the resulting changes. The difficult part becomes deciding those things before something goes wrong.

“Who’s watching that creep? At what point does it become material? Who’s responsible for retraining the model?” Phillips said. “And then not only that, but who signs off on the retrained model?

“Every agent is a threat factor,” he said. “Every agent can be hacked, training data can be poisoned.”

That is particularly important because responsibility ultimately does not transfer to the machine.

“Where does accountability lie? It lies to the institution,” Phillips said.

Watch the full PYMNTS TV interview with Christopher Phillips to hear more about:

• Why AI agents could make financial crime investigations more auditable than humans. Phillips argues agents can standardize investigative workflows, produce more complete documentation and enable QA across every case rather than a limited sample.
• Why human oversight is shifting from approval to effective challenge. Phillips says consequential actions such as closing accounts or recommending SAR filings should still be reviewed by people, with humans expected to explain why they agree or disagree with an agent’s conclusion.
• Why the real control problem is ownership, not autonomy. As agents gain access to sensitive systems and begin executing more work, institutions need clear thresholds for overrides, retraining, escalation and model signoff — because accountability ultimately remains with the financial institution.

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.