Banks need shared controls before they give agents more authority
Banks should make shared financial controls a condition of expanding agent pilots. Waiting until agents are embedded across departments makes the problem harder and more expensive to fix.
Banks should make shared financial controls a condition of giving AI agents more authority. I think waiting until agents are embedded across servicing, collections, and payments is a mistake. By then, customers may already depend on workflows that were never designed to respect one another's decisions.
The immediate temptation is understandable. Find a task, show that an agent can do it, measure the time saved, and expand. A pilot can look successful on those terms while leaving a fairly consequential question unanswered: what happens when another agent is working on the same account?
I would want that answered before approving the expansion. Otherwise, each successful pilot can create another place where the bank makes a commitment without knowing what it has already promised elsewhere.
The window to sort this out is while the deployments are still small
This is already becoming an industry concern. On September 22, Bank of America, Capital One, and four other banks published joint principles for agentic commerce. Their announcement explicitly raises authorization, liability, and customer protection. That work concerns agents shopping and paying on customers' behalf, but the same questions deserve attention inside the bank.
If every department connects its own agent to a different set of systems, coordinating them later means revisiting integrations, approval processes, and assumptions about which records can be trusted. The first deployment is where the bank starts making those choices, whether or not anyone calls them a strategy.
I think the urgency comes from that accumulation. A bank does not need thousands of autonomous agents before two workflows can conflict. It needs two workflows with authority over the same financial position.
A correctly handled dispute can still lead to an unlawful collection
Consider a hypothetical customer disputing $900 of a $1,200 credit-card balance. Assume a qualifying written billing-error notice arrives within the required 60-day window and more than three business days before issuer autopay, with no related charges to account for.
Regulation Z protects the disputed amount during resolution. Under those conditions, the issuer cannot deduct that amount through its autopay plan. The undisputed $300 can still be paid.
A servicing agent can open the investigation correctly while an autopay agent prepares to collect $1,200 from an older account record. The customer protection only works if the dispute changes what the collection system is allowed to execute.
That is why evaluating agents individually is insufficient. The bank needs to test whether a decision in one workflow actually constrains another. Giving both agents the regulation does not establish that connection.
A person approving the action needs the same shared picture
Company limits have the same problem. Imagine a hypothetical $200 allowance for discretionary goodwill credits over 30 days. A customer has received $50, another $80 is approved but pending, and an agent proposes $100 more.
Counting only posted credits makes the new request look acceptable. Counting commitments puts the total at $230. These discretionary credits are separate from mandatory corrections, which cannot be withheld because a goodwill budget has run out.
Putting a human in the approval loop does not resolve this if the reviewer also sees only the posted $50. The bank needs a shared record of proposed and authorized actions, including pending amounts, before either a person or an agent can make an informed decision. That record also has to prevent two workers from committing the same remaining allowance.
This should be part of the investment decision
There is a commercial cost to getting this wrong too. A bank that has to rebuild these controls for every new agent will spend more on integration and review each time it expands. An incident can also force it to pause useful automation while it works out which other workflows have the same weakness.
I would fund the ability to apply shared obligations, limits, and approvals early. I would also require evidence from the underlying financial systems before an agent declares the work complete. An accepted request to change autopay is not confirmation that the scheduled debit changed.
Banks can keep learning through narrow deployments while doing this. The decision I would push back on is expanding an agent's authority because its own pilot looks good, without testing how its actions affect the rest of the institution. That is easiest to address before the next department builds around it.