Agent Traffic at the Bank Door: How Fraud and Risk Teams Tell Sanctioned AI Agents From Bad Bots
For years, fraud and risk teams at banks and payment providers had a fairly clear picture of who was on the other side of a session: a customer, a fraudster or a bot. AI agents have blurred it. A customer's assistant checking a balance, a shopping agent trying to pay and a credential-stuffing script can all look like automated traffic. Treat them all as bots and you block legitimate customers. Treat them all as customers and you open the door to abuse. The industry's first reaction has been to block. Bank of America told customers not to give banking credentials to AI assistants, according to TechCrunch, and JPMorgan has been reported to block agent card payments, though that report rests on a paywalled summary. A vendor-authored article named several UK banks as blocking agents; those names are unconfirmed. Now detection tools aimed specifically at agent traffic are launching. This analysis sets out how fraud and risk teams can move from blanket blocking to a policy that detects, classifies and then decides.
Why agents break the old categories
Traditional bot defences assume that automation is either malicious or irrelevant. Good automation, such as screen-scraping aggregators, was pushed onto sanctioned APIs years ago through open banking. Everything else could be challenged or blocked.
AI agents do not fit that model, for three reasons.
They act for real customers. A customer who asks an assistant to check their spending or pay a bill is a legitimate customer using a new interface.
They look like humans and bots at once. Agents can drive browsers, fill forms and navigate apps, producing patterns that sit between human behaviour and classic bot signatures.
They can be manipulated. An agent can be redirected by content it reads, so even a legitimate agent may behave in unintended ways.
That means fraud and risk teams need to answer a more nuanced question than "human or bot?". They need to know whether automated traffic is a sanctioned agent acting within its authority, an agent behaving oddly, or an attack.
Step 1: detect agent traffic
The first requirement is visibility. Many organisations do not know how much of their traffic already comes from AI agents.
Useful signals include:
- Declared identity. Some agents identify themselves through user-agent strings, signed headers or registered credentials. Declared identity is useful but can be spoofed, so it should be verified where possible.
- Behavioural patterns. Interaction timing, navigation paths and input patterns can distinguish agent-driven sessions from humans and from classic scripts.
- Device and environment signals. Headless browsers, automation frameworks and cloud hosting can indicate agent activity.
- Access route. Traffic arriving through sanctioned APIs, such as open banking, is very different from traffic driving the web or app interface.
Vendors are launching detection products aimed at agent traffic. A fraud-platform vendor announced in November 2026 that it had expanded its digital trust product to help banks detect AI agents accessing customer accounts. Firms should evaluate such tools on their own evidence and integration needs.
Step 2: classify what you see
Detection on its own produces a number. Classification turns it into something you can act on. A practical scheme has four classes.
- Sanctioned agents using a recognised route, with verified identity and a known accountable owner, acting within agreed scope.
- Known but unsanctioned agents from recognisable assistants or platforms, acting through the customer interface without a formal route.
- Unknown automation that behaves like an agent but cannot be identified.
- Hostile automation showing patterns of fraud, credential stuffing or abuse.
Classification links directly to agent identity. The more reliably you can verify which agent is acting and who stands behind it, the more traffic you can move into the sanctioned class.
Step 3: apply a tiered policy
With classes defined, the policy becomes a set of decisions per function and per class. Typical tiers are:
Allow. Sanctioned agents performing in-scope actions, such as reading balances through consented access.
Allow with confirmation. Sanctioned agents initiating sensitive actions, such as payments, with explicit customer confirmation. Starling's "Talk to Transact", which lets customers start a payment by voice and confirm with a tap, according to FF News, illustrates a bank keeping confirmation in its own channel.
Challenge. Known but unsanctioned agents, or unusual behaviour from sanctioned ones, trigger step-up authentication or a request to use a sanctioned route.
Restrict. Unknown automation is limited to low-risk, read-only functions or slowed down.
Block. Hostile automation is blocked and investigated.
The policy should be set per function, not per channel. Reading a balance and initiating a payment carry very different risks.
Step 4: offer a sanctioned route
Blocking without an alternative pushes customers towards workarounds, such as sharing credentials with assistants. The more durable answer is to offer agents a sanctioned way in.
In the UK, open banking already provides consented, scoped and revocable read access through regulated providers. For payments, options under discussion include pre-authorised mandates with customer-set limits, scoped agent credentials from card schemes and issuers, and in-app confirm-before-pay flows. For teams building on our Agentic Commerce & Payments agenda, the aim is to move as much legitimate agent traffic as possible into routes where identity, authority and limits are clear.
Avoiding the alert trap
Agent detection can generate a large volume of new alerts. If every agent session raises an alert, fraud teams will drown and genuine threats will be missed.
- Automate the routine decisions. Sanctioned agents within scope should not raise alerts.
- Alert on change, not presence. Flag sanctioned agents behaving outside their normal pattern, not their existence.
- Feed outcomes back. Use confirmed fraud and false-positive outcomes to tune classification.
- Measure customer impact. Track how many legitimate customer agents are challenged or blocked, not just how many threats are stopped.
A worked example: three sessions on a Tuesday morning
Imagine three automated sessions reaching a UK bank's channels in the same hour.
Session one arrives through the bank's open banking API from a regulated provider, on behalf of a customer who consented to read access for a budgeting assistant. Identity is verified, scope is read-only and behaviour is normal. Classification: sanctioned. Policy: allow, with no alert.
Session two drives the bank's web interface. It identifies itself as a well-known AI assistant acting for a logged-in customer and is trying to set up a new payee. There is no formal route for this assistant to initiate payments. Classification: known but unsanctioned. Policy: challenge with step-up authentication in the bank's own app, and show the customer how to use a sanctioned route next time.
Session three also drives the web interface, but with rapid login attempts across many accounts from cloud infrastructure, and no declared identity. Classification: hostile. Policy: block and investigate.
Under a blanket-block policy, sessions one and two would both have been refused, frustrating two legitimate customers. Under no policy, session two might have added a payee without the customer's explicit confirmation. The tiered approach treats each appropriately.
What regulators are likely to expect
UK regulators have not issued agent-traffic rules. But existing expectations apply. The FCA's October 2026 document on AI in financial services says existing rules apply and frames AI use as support, automate or abstain, with live monitoring as a concrete expectation. Consumer Duty requires firms to avoid foreseeable harm and to support customers in pursuing their financial objectives, which includes not unreasonably blocking customers who choose to use legitimate tools. Firms should be able to explain their agent policy, show how it protects customers and show that it is monitored and reviewed. Innovify is a technology partner and not FCA-authorised; firms should take their own advice on how these expectations apply.
Who owns agent policy?
Agent traffic sits across fraud, digital channels, security, product and compliance. Without a clear owner, decisions are made inconsistently in each channel. A named owner, with input from each function, should hold the agent-access policy, review it as standards and threats change, and report on its effect.
Designing the customer experience of a challenge
Every time policy challenges or restricts an agent, a real customer may be on the other end. The experience of that moment matters.
- Explain, do not just refuse. Tell the customer, in their own app or by message, that an automated tool tried to perform an action and what they can do next.
- Offer the sanctioned route. Point customers to the supported way for their assistant to work, such as consented open banking access.
- Keep confirmation in your channel. For sensitive actions, ask the customer to confirm in the bank's own app rather than inside a third-party interface.
- Record the outcome. Log what was challenged and how the customer responded, so policy can be tuned.
A practical starting checklist
- Measure how much automated and agent traffic reaches each channel today.
- Define the four classes and how each is identified.
- Write a tiered policy per function: allow, confirm, challenge, restrict, block.
- Build or buy detection that supports the classes, and evaluate vendors on evidence.
- Offer sanctioned routes for legitimate agents.
- Tune alerts so routine agent activity does not swamp investigators.
- Name an owner for agent policy and review it quarterly.
Where Innovify fits
Innovify is a London-based technology partner with deep expertise in banking, financial services, fintech and insurance, and we design compliance in from day one. Through our AI / ML Development teams, we help banks and payment providers build detection and classification pipelines, policy engines and feedback loops, and integrate them with sanctioned access routes so legitimate customer agents can work while threats are contained.
FAQ
How do banks detect AI agents?
Banks combine declared identity signals, such as signed credentials or user-agent strings, with behavioural patterns, device and environment signals and the access route used. Specialist tools for detecting agent traffic are emerging. Declared identity should be verified, because it can be spoofed.
What is the difference between an AI agent and a bot?
A bot is any automated software; many are malicious or irrelevant. An AI agent is automation that pursues a goal, often on behalf of a real customer, and can adapt its steps. Agents may be legitimate customer tools, so treating them all as bots risks blocking genuine customers.
Should banks block AI agents?
Blanket blocking is a common first reaction but pushes customers towards risky workarounds such as sharing credentials. A better approach is to detect and classify agent traffic, apply a tiered policy per function and offer sanctioned routes, such as consented open banking access, for legitimate agents.
What is an agent-access policy?
It is a set of decisions about how a bank or payment provider treats automated agent traffic, per function and per class of agent: allow, allow with confirmation, challenge, restrict or block. It should have a named owner and be reviewed as standards and threats change.
How can fraud teams avoid being overwhelmed by agent alerts?
Automate decisions for sanctioned agents acting within scope, alert on changes in behaviour rather than agent presence, feed confirmed outcomes back into classification and track the impact on legitimate customers alongside threats stopped.
Conclusion
AI agents are now part of the traffic at every bank's door. Blocking them all is simple but costly; letting them all in is dangerous. The workable path is the one fraud teams already know from other threats: detect, classify, decide and learn. Add sanctioned routes for legitimate agents and a clear owner for the policy, and agent traffic becomes something to manage rather than something to fear.
Related reading
- Know Your Agent: How Banks and Merchants Can Verify an AI Agent and the Accountable Owner Behind It
- Why Banks and Marketplaces Are Blocking Outside AI Agents, and the Sanctioned Routes Left Open
- The Hidden Cost of Fraud Detection: When Too Many Alerts Become the Risk
- Adaptive Fraud Detection for Online Payments: Architecting the Feedback Loop, Not Just the Model











