The case

Fictional case. Bank A is a composite of a large US consumer bank. Names, figures, and identifying details have been changed or invented to protect confidentiality.

What Bell remembered

Bell had been at the bank long enough to remember the old call center, back when automation there meant a phone tree and hold music. The tree made people learn the bank's categories before the bank would listen to their problem. Press one for balances, press two for payments, press three for a lost or stolen card, and if you pressed the wrong one the system sent you back to the top. A customer would sit with a bill open, trying to remember whether they had pressed two or three, wait through the hold music, and sometimes learn the office had closed only after the wait.

He had worked that floor for years, long enough to know exactly what was wrong with it. He knew what it did to customers, and he knew what it did to the people who worked the phones. Reps picked up calls that were already angry, because the system had made the customer wait before it let them reach a person. Managers measured seconds, because seconds were the only thing they could afford to improve. For years the department's answer to a December call spike was to route a little faster and put a few more people on the phones.

He led the work that changed it, first with a system that could understand plain speech, so a caller could say what was wrong instead of translating it into a menu, then with the front-line agent that grew out of it, the one customers reach first today. Later came the second agent, the one that sits with the reps. Four years, and a good part of the standing he had built at the bank, went into getting both into production. By the summer of 2026 he ran card servicing, and card servicing ran on two agents he had put there himself. He was proud of that, and the pride was part of what made the choice a hard one.

The choice on the desk

Card servicing had one build line for the coming year, meaning one engineering team, one stretch of its time, about seven million dollars, and two proposals it would not both cover. The executive committee had been plain when it handed the decision down. This was card servicing's call, on card servicing's own budget, and there was no central pool to fund this kind of build this cycle.

The calendar took away the soft options. Whatever Bell's team built, it had to build now, before the line was gone. The bank had a new consumer card coming, Cascade, with a go-live the committee had already fixed and announced for March 3, and Bell's engineers were on the hook for the servicing behind it. From the day they rolled onto Cascade they were spoken for through the rest of the fiscal year. The proposal he passed over would wait a full year for the next cycle, or it would not happen at all. His recommendation was due Friday, and it was Wednesday.

What the floor runs on now

The first agent is the one a customer reaches when they call. They can say “I lost my card,” or “when is my payment due,” or “I'm traveling next week and I don't want a hold on the account,” and the agent verifies who they are, opens the account, does the work, and closes the call. Balances, payment dates, lost cards, travel notices, the ordinary traffic of a card business, it handles end to end. It now finishes a little over seventy percent of contacts on its own, and the share climbs most quarters. Bell watched that number more closely than any other, because it was the plainest sign the thing worked. Customers stopped waiting, and reps stopped taking calls a machine could handle.

The second agent only shows up after a handoff. When the first agent cannot finish, or a caller asks for a person, the call goes to a human rep, and by the time the rep is on, the second agent has already read the account, pulled the recent transactions, and flagged what looks off. The rep starts the conversation with the whole picture in front of them instead of searching six systems while the customer listens to them type. It settles nothing on its own, but it gets the rep to the answer faster. Bell had built it for a blunt reason, because his reps were quitting. Once the first agent took the easy calls, what was left for a person was a day of nothing but complaints and edge cases, and that wears people down. The assist agent was how he kept the floor staffed.

What his own team wanted

There was one call the front-line agent still could not finish, and it was a big one, billing disputes. A customer says a charge was never theirs, or the goods never came, or they were billed twice. The agent hears them out, pulls the charge, checks it against the account, asks the questions the process requires, and gets almost all the way there before it stops and hands off to a person. Provisional credit puts real money out the door on the bank's say-so. Regulation Z, the federal rule that sets deadlines for how a bank must handle a disputed charge, puts hard clocks on the whole thing, and the bank's own guardrails keep a human on the decision. Disputes were close to a fifth of everything that still reached a rep, and they ran the same handful of steps almost every time.

Gary Pruitt, who runs the floor day to day, had been after Bell about it since the spring. He wanted to teach the front-line agent to carry the whole dispute, provisional credit and all, inside the rules the bank already used. “We built it up to the doorstep and stopped,” he told Bell. “The agent does everything except the part that matters, and then it wakes up a person to do the part it already understands. Twenty percent of what still comes to my reps is this. Give me the disputes and I give you back a chunk of the floor.”

The numbers were good, if not quite the clean sweep they looked like at first. The build ran about 2.4 million dollars against a gross benefit near 9 million a year, in handoffs deflected, handle time cut, and reps freed for the calls that need a person. It ran on the busiest system in the department, though, and the running cost that came with that volume, a couple of million a year, pulled the real number down toward seven. It still paid back fast, and it still cleared the bar the department used (Exhibit 3). It was the kind of build Bell had spent years teaching the bank to say yes to, local and measured, with the people who would live with it the ones asking for it.

The case for it was not only the money. Disputes were where the bank's Regulation Z exposure showed up most directly, and a late or unevenly handled dispute could still become a regulator's letter or a customer who closed the account. Handling them the same way, inside the guardrails, might lower that risk somewhat, even if the cleaner case was still cost and capacity. The timing pressed too, because his reps were closest to worn out right as Cascade was about to add a new card's worth of volume to the floor, and the dispute build was the one thing that would take real weight off them before that hit.

There was a catch the risk side kept raising, and Bell did not have a clean answer for it. Teaching the agent to carry the whole dispute meant letting it issue provisional credit on its own read, and moving money out the door on a model's judgment was the exact thing the human gate had been there to hold. A wrong call would stop being a mistake a rep could catch in the moment and become a payment the bank had already made, and the build leaned still more of the department's weight onto the one system that already carried the most. Bell told himself the guardrails would hold, but he had told himself that before.

What compliance wanted

The second proposal came from outside his floor, from Elaine Cho in compliance, and it was a harder thing to defend.

Cho did not want an agent that talked to customers, she wanted one that read the bank's own work before it went out. When a product team ships a new fee disclosure, or changes how a promotional rate is applied, or adjusts a model that decides who gets approved and on what terms, someone in compliance is supposed to check it against the rules before it goes live. Today that someone is a small team reading documents against a launch date, and how good the check is depends on whose desk it lands on and how much time that person has that week. Cho wanted an agent that would read the same material, check it against the bank's compliance requirements, and flag what a person needed to look at before anything shipped. She had come to Bell because his engineers were the ones with room on the calendar. Hers were not.

Cho was careful about how she put it. “I can't give you a number as clean as Gary's,” she said. “The value here is a thing that doesn't happen. A finding we catch in March instead of an enforcement letter in November.” One example still bothered her, a card-marketing disclosure that had gone out for review nine days before launch, and someone caught that it did not say what the CARD Act required it to say. They fixed it in time, but they had nearly not. “The next one we might miss,” she said. Across comparable issuers, she added, cleaning up a single disclosure or CARD Act finding had run into the tens of millions once the remediation and the penalties were counted. None of that was money card servicing would ever see. The direct return to his floor was maybe half a million a year in rework it would not have to redo, against a build of about 6.5 million (Exhibit 4).

She was honest about the part that hurt her own case. “My team is too thin to validate this thing even once it's built,” she said. “I'd need product to lend us their people, and product has a day job.” Near the end, she mentioned one more thing about the design, that the rules engine sat apart from the card-specific rules it checked against. The engine did the reading and the checking, while the rules it applied were a separate set it loaded in. Bell noted the point and set it aside.

What the choice would cost

Bell knew what his scorecard said. He had built that scorecard himself, more or less, over years of arguing for the resources that put his two agents into production. It set what a build cost against what it saved card servicing, and by that measure Gary's dispute agent was the easy yes and Cho's was a hard sell he would have to walk his own team through.

Choosing Cho's build meant telling his floor that the dispute work they had wanted for a year would slip, and with the team moving onto Cascade right after, slip for a good while. Gary would take it hard, and he would have a point, because his people had carried the agent program from the start, absorbed its failures, retrained it, taken the handoffs, and explained the errors to angry customers. Now that the next piece was in reach, the line would go to compliance instead.

Choosing Gary's build was the easy call, and that was the part that nagged him. He would again be the executive who spent the whole line on his own operation while the rest of the bank waited. He had gotten good at that, good enough that his floor ran on two agents while much of the bank was still trying to get its first one through. That had always felt like proof he was moving faster than everyone else, and it had usually been enough to decide by, but this time it was not deciding anything.

He read back over both proposals, the queue of open requests behind them (Exhibit 2), and the compliance findings from the last four quarters (Exhibit 5). The recommendation was due Friday, and after that the team was gone for the year.