Support agents can now 'connect once, act for weeks'

Decagon's AI agents can reuse one bank connection for weeks without a new login, but SOC 2 Type 2 is a point-in-time check, not continuous monitoring.

Published:  

October 2, 2026
4

min  

Support agents can now 'connect once, act for weeks'

Key takeaways

  • Decagon's AI agents keep a customer's bank connection alive across weeks of separate conversations after a single login.
  • Its monitoring covers conversation quality, not whether Decagon is a safe place to hold standing bank access.
  • SOC 2 Type 2 is a point-in-time credential, not continuous monitoring.
  • The customer authorized a brand but inherited a chain of banks, aggregation rails and an intermediary.

Customer support platform Decagon announced a partnership on September 28 that lets its AI agents connect to a customer's bank account once and keep using that connection across later conversations, spanning days or weeks, without asking the customer to log in again. Decagon's agents are already live at Square, Affirm, Chime, and Valon. The pitch is straightforward: fewer customers get bounced out of a support chat to go authenticate somewhere else, so more issues get resolved in the conversation itself.

The design choice worth looking at closely isn't the bank connection. It's what happens to it after it's made.

One connection, reused indefinitely

Decagon's own description is direct about this being the point, not a side effect: “the account connection from the first conversation persists,” so when the customer's system “signals that the next step is ready, the agent picks it up and continues from there,” without starting over from authentication. A claim filed today and settled three weeks later, or a payment that fails on the 3rd and needs recovering by the 15th, can now be handled by an agent acting on a bank connection the customer authorized once, at the start, and never revisited. The standing access is the feature.

The monitoring is watching the wrong layer

Decagon describes real oversight: guardrails running before, during, and after each conversation, and a system that scores every conversation against the deploying company's own criteria. That's monitoring of conversation quality, whether the agent stayed on script and followed the rules Square or Chime wrote for it. It says nothing about whether Decagon itself, now a standing intermediary holding reusable bank access across every one of those companies' customers, is a safe place for that access to sit. Decagon's own materials point to a SOC 2 Type 2 certification and similar compliance attestations. Those are point-in-time credentials, assessed on a renewal cycle, not a continuous read on whether an entity holding live financial access this week is the same risk it was when it was certified.

The customer authorized a brand. They inherited a chain

A customer connecting their bank account inside a Chime or Affirm support chat reasonably believes they're extending trust to Chime or Affirm. What they're actually authorizing runs through at least two more parties: the bank issuing the account, the account-aggregation rail carrying the connection, and now Decagon, sitting in the middle, holding the standing credential that makes the multi-week follow-up possible. None of those intermediate parties were named to the customer at the moment they said yes. The visibility available to them afterward is a connections page where they can see what's linked and disconnect it, the same self-service model this blog has already covered elsewhere. It tells a customer what's connected. It was never built to tell them whether every party in that chain is still a safe place for their bank access to live.

To be clear about what isn't the problem

This isn't a claim that the underlying account connection is poorly built. Credentials going directly to the bank, and a customer-facing portal to review and revoke access, is the same reasonable architecture other providers in this space use. The design decision under scrutiny sits a layer up: what an AI vendor and its client brand choose to do with that access once it's granted, for how long, and under whose ongoing watch. Persisting it for weeks, unattended, invoked automatically whenever an internal system decides the moment is right, is a choice, not a default the underlying rails required.

What actually happens if this goes wrong, and who answers for it

Put the two gaps above together and the point of failure is clear. Decagon's certifications are checked on a renewal cycle, not continuously, and the customer's only window into the connection is a page showing that it exists, not whether it's still safe. So the way this goes wrong is a compromise at Decagon, or a flaw in how it stores or replays that standing credential, wouldn't announce itself. It would surface at a later point as unauthorized activity on a customer's account, long after they'd forgotten they'd connected anything at all. And the party a customer calls first, and the one regulators and dispute processes typically look to initially, is the bank whose name is on the account, not Decagon, not the brand that deployed the agent, and not the API that carried the connection. The institution furthest from the decision to let that access persist unattended is usually the first one on the hook for making the customer whole.

Invela is the infrastructure layer that makes open finance trustworthy - accrediting who's in the network, monitoring risk in real time, and ensuring liability lands in the right place.

‍

Invela INSIGHTS Blog

Open finance, explained

Insight on accreditation, risk scoring and warranty for banks, intermediaries and third parties.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
Louise Beaumont
Written by
Head of Marketing, Communications & Design

Louise Beaumont leads marketing, communications and design at Invela, with 30+ years in tech and fintech and 15+ in open banking and open finance.

Connect on LinkedIn