Suppose your bank’s terms say that when a customer authorizes an AI agent, responsibility for what that agent does moves toward the customer. Then a customer disputes a transfer and says she never set up any agent.
Open the file. What does it contain?
The signal from the largest U.S. bank by assets
On October 13, 2026, CNBC reported that JPMorgan Chase currently blocks external AI agents from operating on behalf of retail customers, a point a bank spokeswoman confirmed to CNBC. The same article reported that agents could still read some data, such as balances, through Plaid or another connector (CNBC, 2026).
The line your risk and legal teams should read is a different one. Dimon said: “If people are using them, they should understand the risk. They should understand there’s a liability transfer.” CNBC’s reading is that he “seemed to draw a line about liability when customers authorize outside AI agents to act on their behalf” (CNBC, 2026).
CNBC’s report does not cover JPMorgan’s terms, but the bank’s Digital Services Agreement does. Its November 2026 version says a customer who permits a third party, including “External AI,” to access their accounts will “assume all risks and losses associated with such access or use,” to the fullest extent permitted by law. CNBC does report that the spokeswoman said there is no timeline for the safeguards that would let agents deal with accounts directly, that the bank supports “secure, customer-authorized data sharing through open banking providers,” and that it is “committed to enabling agentic connections securely” (CNBC, 2026). What follows is Trusona’s view of the evidence problem, not a statement of law, and your counsel writes the terms.
What the file holds today
A login, a code or a tap records that something was presented. None of them records who presented it.
| Evidence in the file | What it shows | What it leaves open |
|---|---|---|
| A valid password or active session | Someone held the credential | Whether the customer was the one at the keyboard |
| A one-time code sent by text or email | Someone could read that phone or inbox | Whether that someone was the customer |
| A tap on “Allow” in the app | A signed-in device recorded a tap | Who was holding the device |
| A government-issued ID checked against the issuing authority | A person passed a document check against the authority’s record at that moment | Whether the delegation was wise, and what the agent does afterward |
An attacker who has the credential, the inbox or the unlocked phone can produce the same log entries the customer would. In Trusona’s view, a liability position that rests on the first three rows has a hole in a dispute, which is where it gets tested.
Four people who could have tapped “Allow”
Picture the dispute. The customer says she did not authorize the agent. Four accounts of that moment are plausible, and the log fits all of them:
- She did, and has forgotten or changed her mind.
- Someone in her household used her phone.
- A criminal holding stolen credentials set the agent up.
- A criminal impersonated her to your help desk and had the agent attached there.
Your fraud team can reason toward an answer by looking at device history and behavior. That is investigation after the fact. A record made at the moment of authorization changes the question from “who do we think this was?” to “who did the check say it was?”
A check that reaches the person
The check has to leave the session, because a session can only confirm what it already contains. Trusona publishes two APIs that verify a person against a government ID.
- The ATO Protect API (ATOP API): the person “gets a verification link by text or email and scans their ID,” and a request can also include a state DMV check over AAMVA (Trusona, integrations page).
- The IDV API (AAMVA+): “Confirm a real person behind a government ID at onboarding or step-up.” It checks the government ID against the authority that issued it, which in the US is the state DMV over the AAMVA network (Trusona, IDV API page).
For a customer, the scan can happen “in your own app or in the mobile browser” (Trusona, IDV API page), and the homepage lists “No pre-registration.” Asked whether it runs a selfie or liveness check, the IDV API page says: “It verifies the document against the issuing authority instead of matching a face.”
Trusona verifies the person. It does not inspect, authenticate or score the agent. Whether a customer’s authorization of an agent counts as a step-up in your flow is a design call for your team, and the right starting point is a conversation with sales.
What the record would not do
Four limits, so nobody reads this as a guarantee.
It does not settle who pays. That depends on your terms, your regulators and your counsel, and a check result is evidence, not an outcome.
It confirms one moment. A verified customer can still authorize a careless agent, so scopes, limits and revocation stay your controls.
It needs a government-issued ID. Customers without one need another path.
It adds a step. Put it where the exposure justifies it, such as granting an agent the ability to move money, and not on every login.
Questions for your next risk review
- If we move responsibility to the customer for an agent’s actions, what evidence shows the customer authorized it?
- Would that evidence look any different if a stranger held the customer’s credentials?
- Does the check run against our own data, or against the authority that issued the ID?
- Which agent permissions justify the extra step?
- Who approves the wording in our terms once the evidence exists?
For the case where a bank admits agents on a condition, see How Banks Can Let Customers’ AI Agents In: Verify the Person Who Authorized the Agent. For the help desk version of the same gap, where an agent acting for someone inherits their authority, see verifying the human behind an AI agent.
Ori Eisen, Trusona’s founder and CEO, puts the stance in one line: “Don’t block agentic AI, but do ask it to reveal the True Persona behind it.”
Before your terms move responsibility to the customer, get a record of who said yes.
Frequently asked questions
What does “liability transfer” mean for AI agents in banking? Jamie Dimon used the phrase when he said customers who use outside AI agents “should understand there’s a liability transfer” (CNBC, 2026). JPMorgan’s Digital Services Agreement, in its November 2026 version, says a customer who permits a third party, including “External AI,” to access their accounts will “assume all risks and losses associated with such access or use,” to the fullest extent permitted by law. How responsibility is allocated depends on each bank’s terms, its regulators and the law, and is a question for counsel.
How can a bank show a customer authorized an AI agent? A password, a one-time code or a tap on “Allow” shows that a credential or device was used. In Trusona’s view, a record that a person passed a government ID check against the issuing authority at that moment is stronger evidence of who authorized it. Whether it is enough is a legal question.
Does Trusona verify the AI agent? No. Trusona verifies the person. What the agent may do afterward stays with your platform.
Is there a selfie, and does the customer need an app? No selfie: the IDV API page says it “verifies the document against the issuing authority instead of matching a face.” The scan can run in your own app or in the mobile browser, and the homepage lists no app download and no pre-registration.
Which Trusona products run the person check? The IDV API (AAMVA+) and the ATO Protect API (ATOP API). See the IDV API page and the integrations page.
If your team is deciding what evidence to hold when a customer authorizes an agent, talk to sales about where the person check fits in your flow.
See how ATO Protect verifies the person behind an AI agent.