Your application queue gets a new entry. Every field is filled in, the address is already standardized, the date of birth is in the format your form wants, and the whole thing took four seconds. Your validation rules have no complaints. Behind the submission sits an AI agent with no thumbs and no driver’s license.

The biggest U.S. bank said no to outside agents acting on customer accounts

On October 13, 2026, CNBC reported that JPMorgan Chase will not let outside AI agents operate on customer accounts. CEO Jamie Dimon told analysts: “We’re not going to allow just any … outside agent to come in inside the company, because it’s just too risky for clients and for JPMorgan.” (CNBC, October 13, 2026). A JPMorgan spokeswoman told CNBC there is no timeline for when the bank would have safeguards in place to let agents interact with customer accounts directly.

That is a decision about accounts that already exist. Your team meets the same question earlier, at the front door. Someone’s assistant opens an account for them, or a customer taps “Allow” and hands an agent the keys to the one they have.

A complete form proves the data, not the person

Agents are good at having information. When a customer connects an agent to their mailbox and files, it can fill in forms from them. A fraudster holding a stolen data set can supply the same fields, just as cleanly.

So a correct application tells you the data is right. It does not tell you the person it describes was there.

Dimon also put the burden on the customer: “If people are using them, they should understand the risk. They should understand there’s a liability transfer.” (CNBC, October 13, 2026). If your platform is going to move liability onto the customer who authorizes an agent, your team needs to know it was the customer who authorized it.

Why the usual checks fall short

Check What it shows Gap when an agent is in the loop
Field validation The data is well formed Says nothing about who submitted it
Emailed or texted code Someone can read that inbox or phone An agent with access to that inbox or phone reads it too
Knowledge questions Someone knows the answers If the answers are in the data the agent was handed, the agent has them too
Selfie and liveness A face appeared on camera Trusona’s position is that deepfakes beat these checks, and it skips them
Government ID checked against the issuer The issuing authority holds a matching record The person still has to be the one who scans it

The last row is the only one that asks an authority outside your system and outside the agent’s reach.

Reach the person, not the session

The agent lives in a session. The person lives on a phone. A check that stays inside the session can only confirm what the session already contains, so the check has to leave it.

Trusona publishes the pieces for this. With the ATOP API, “the person gets a verification link by text or email and scans their ID” (trusona.com/integrations). When an agent may be reading the customer’s inbox, send it by text. The published example request does that: it sends an SMS link to a phone number and asks for a DMV verifier. Your system gets the result back through the API.

The IDV API (AAMVA+) is the lighter version. Its page describes it as a way to “confirm a real person behind a government ID at onboarding or step-up,” and lists new account opening as a use case. It checks the license against the state DMV through AAMVA’s Driver’s License Data Verification service, reads the PDF417 barcode, and validates up to 17 fields. The person can scan in the mobile browser, and the homepage lists “no app download” and “no pre-registration.” On selfies, the page is blunt: “It verifies the document against the issuing authority instead of matching a face.”

A forged license built on invented details has no DMV record to match, however good it looks.

If you have a phone number and the person’s details, IDV API MNO+ works from those: “One SMS message and you are done.” Numbers in the US, Canada and the UK work by default.

Where the call goes

Two places, and they are your team’s decision.

At application. The agent submits the form. Before the account exists, your system sends the verification to the applicant’s phone along with the details the application gives, and the result shows whether they hold up against the issuing authority’s record.

At authorization. An existing customer wants an agent to act for them. Before the agent gets scope, your system sends the verification to the customer’s phone. The agent waits while the customer scans, and the scope opens only on a verified result.

Trusona’s use-cases page covers this case under AI Agents, with the ATOP API and the IDV API (AAMVA+): ATO Protect checks the person first. Your onboarding or authorization flow can call the verification, and the check lands on the person. Trusona does not verify, authenticate or score the agent, and it does not need to. Your platform still decides what an agent may do once a verified person has approved it. The enterprise side of the same gap, where an agent inherits an employee’s authority, is covered in 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.”

Where this runs out

  • Documents. The IDV API (AAMVA+) checks a U.S. driver’s license, permit or ID card. Applicants without one need another path, and MNO+ covers phone numbers in three countries by default.
  • One moment. It confirms a real person took part now. It does not judge whether the delegation was wise or what the agent does next.
  • Friction. A text message and a scan add a step. Put it where the risk justifies it, such as account opening and new agent permissions, rather than on every login.

Questions for your next design review

  • When an agent submits an application, what confirms a person stood behind it?
  • When a customer authorizes an agent, what confirms it was the customer?
  • Does the check run against our data or against the authority that issued the ID?
  • Does it need the customer to have enrolled something first?
  • Which agent actions justify the extra step?

Frequently asked questions

Can an AI agent open a bank account for a customer? An agent can fill in and submit the application. What it cannot do is be the person on the ID, so your team needs a step that reaches the human directly.

Should banks block AI agents? JPMorgan Chase currently blocks outside AI agents from operating on retail customers’ accounts, according to CNBC, while allowing read-only data sharing through open banking providers. The alternative is to let agents in and ask for proof of the person behind them before they get access.

How do you verify the person behind an agent without a selfie? By checking a government ID against the authority that issued it. The IDV API (AAMVA+) checks the license against state DMV records through AAMVA and does not match a face.

Does the customer need to install an app or enroll first? Trusona’s homepage lists no app download and no pre-registration. The IDV API (AAMVA+) page says the scan can run in the mobile browser or in your own app.

Does Trusona verify the AI agent? No. Trusona verifies the person. What the agent is allowed to do afterward stays with your platform.

Your team can block the agent or wave it through. It can also ask who is behind it, and say yes with a reason. See the IDV API and the ATOP API, or talk to sales.

See how ATO Protect verifies the person behind an AI agent.