Banks Are Running Out of Time to Build. Here’s Ready, in 60 Seconds.

The agents are already inside. The question now is whether you can afford to spend a year building the answer.

Tim Williams — CEO and CoFounder — AstraSync AI

This week Axios reported that major US banks are finding unauthorised AI agents operating inside customer accounts. Some are acting beyond the customer’s instructions. Some are ringing call centres. One bank executive said they can’t tell one agent from another, or whether the customer authorised it or it has gone rogue.

We’ve been saying this to banks and exchanges for the past year. This isn’t a told-you-so. It’s a warning that the clock just got a lot shorter. We even published this article, warning about this impending crisis only two weeks ago.

The banks’ first move has been to lock the door. JPMorgan, Citi, Bank of America, Schwab and Morgan Stanley have blocked or restricted outside agents, and read-only access is the stopgap. The same executives told Axios they know customers will want agents, and that they need a middle ground.

The block buys time. The real question is how much, and what you do with it.

Some saw this coming

While most of the industry waited, a few players got on with it.

Walmart started shopping through ChatGPT with OpenAI in October 2025. This week, roughly a year later, it launched its own Agent Gateway and announced Agent Profiles: each agent authorised individually inside the customer’s account, with its own spending limit or approval on every purchase, and terms that get non-compliant agents blocked. Agent Profiles don’t arrive until next month.

Robinhood opened its platform to customer agents in May. More than 150,000 customers have since opened agentic trading accounts, and agents now call its tools nearly 30 million times a day.

Neither of them waited for a standard or a regulator to tell them what to do. Both built a front door, because the alternative was agents coming in through the window.

The build-or-buy clock

If your plan has been to block now and build later, look hard at that timeline.

Walmart has the resources of a not-so-small nation. It took about a year to get from first agent partnership to its own gateway, and the customer controls still aren’t live. Most banks don’t have a year. Customers are already handing agents their logins. Agents are already in authenticated sessions you can’t see. The first complaint about an agent moving money without consent is a matter of when, and the supervisor’s question won’t be about your roadmap. It will be: did you know an agent was acting on this account, and what stopped it?

The urgency of this space is turning a build decision into a buy decision. Better to make that call on your own terms than have it made for you after an incident.

We’ve spent the last 18 months building the “buy” option and you can watch it in the video here:

Watch the 60-second demo on YouTube

What you just watched

1. See every agent. Visit Intelligence shows which sessions, API calls and interactions have an agent in them, and which platform each agent comes from.

2. Give legitimate agents a front door. Through the Trusted Agent Gateway, an agent declares itself instead of impersonating the customer. The human behind it is verified and bound to the agent, along with a mandate: what it’s for, how long it may act, how much it may move, what it may touch.

3. Decide before the money moves. Every consequential action is checked against the mandate. Execute, deny, or hold for the customer to step up.

4. Keep the evidence. Every decision is written to a tamper-evident record linking the human, the agent, the mandate and the outcome. When a customer disputes an agent payment, you have the answer the regulator will ask for.

That is the same front-door model Walmart has just announced, built once and available to anyone who needs it, without the year-long build.

Proof, stated precisely

• Regulator-grade reference: with an institutional cross-asset exchange regulated in multiple tier 1 jurisdictions, we made a joint submission to MAS as a reference implementation of its SAFR framework. The submission was accepted and the deployment is in proof of concept.
• Fast: about 12 milliseconds added per decision.
• Human identity built in: Key partnerships with FrankieOne, ConnectID and Persona, verifying the human behind the agent.
• Protected: patent pending.

Ready to pilot today

We have a solution ready to pilot now, not on a roadmap.

We’ll be in the US from 30 October to 13 November to kick off customer pilots, and we’re taking meetings to plan more. If you’d rather not spend the next year building, let’s talk while we’re there.

Want to take it for a spin first? Our AWS Marketplace listing is live. You can buy it on your existing AWS bill and start today.

BOOKING LINK · AWS MARKETPLACE LINK

Further reading on AstraSync

• The Barbarians Are Already Inside The Gates
• Trusted Agent Gateway: give legitimate agents a front door
• What is Know Your Agent (KYA)?

This essay first appeared on Medium on 23 October 2026.