Follow the Money with Claire Dubois

At 9:58 on a Tuesday evening in London, Claire Dubois stood at her kitchen counter with a laptop open next to a bowl of olives and read a note that was not written for her. It was written for Melbourne. In two minutes the London payments operations shift would end, and 214 open items would move to a team that was just then making coffee on the other side of the planet.

The note was four screens long. An agent had built it from the London queue: every returned payment still in flight, every business onboarding case waiting on a document, every reconciliation break above the threshold, every card dispute with a deadline inside 72 hours. Each item had one line on what happened, one line on what London had tried, and one line on what Melbourne should do first. Claire scrolled, stopped at item 61, and typed a comment: “This one is not a timing break. Call the partner bank before you touch it.”

Then she closed the laptop. I asked her if she does this every night. She said no, only most nights, and that the habit is the whole reason the system works. “The agent writes a better note than any of us did,” she told me. “But I want Melbourne to know that a person read it. If they know I read it, they read it properly too.”

Tradewind gives businesses global accounts, currency exchange, cross-border payouts, and cards. For the operations team, that product list turns into a long tail of things that go wrong in the space between countries. A supplier payment to Vietnam comes back because a beneficiary name has a diacritic the receiving bank stripped. A new customer in Hong Kong uploads a company registry extract that is eleven months old. A card transaction in euros settles at a different amount than it authorized. None of these are rare. At the volume the team works, each one happens hundreds of times a week.

Claire’s team is 41 people: 16 in Melbourne, 14 in Singapore, and 11 in London. Two years ago it was 58, and the work has roughly tripled since. The team does not divide by office. It divides by queue, and every queue runs through all three offices. Melbourne takes it at 7 a.m. local time, Singapore picks it up mid-afternoon, and London carries it through the European day and the opening hours of the US. Nothing waits for a person to wake up.

The handoff note was the first thing they gave to an agent, and Claire says it was the most important. Before, each shift wrote its own summary in a shared document. The summaries were good when the shift was calm and thin when it was not, which is to say they were thin exactly when they mattered. Items fell through the gap between offices. In the quarter before the change, the team traced 23 customer escalations to a handoff where the next office did not know an item existed.

Now the agent writes the note from the state of the queue itself, not from anyone’s memory. It cannot forget an item, because it reads all of them. In the two quarters since, handoff misses dropped to three. The people at the end of each shift spend their last thirty minutes on the items the agent marks as unclear, and on adding the context only a person has: a phone call that went badly, a hunch about a partner bank, a customer who sounded worried.

The London shift maps the day’s open reconciliation breaks before the handoff to Melbourne.

Tavita Faleolo runs payment exceptions from Melbourne, and she has a rule she repeats to every new hire: a returned payment is never just a return code. The code says the account was closed, or the name did not match, or the bank could not apply the funds. The story is what the customer was trying to do and what will happen to them if the money does not arrive. A payroll run that bounces is a different problem from a refund that bounces, even when the code is the same.

So the agents do the first pass on every return, about 3,900 a week across all corridors. For each one, an agent pulls the original payment instruction, the partner bank’s return message, the customer’s history of payments to that beneficiary, and any earlier returns on the same account. It sorts the item into one of eleven groups the team defined. For the simple ones, such as a name mismatch where the customer has paid the same beneficiary under a corrected name before, it drafts the message to the customer and the corrected instruction, and a person approves both with one click.

That covers about 62 percent of the volume. The rest goes to a person with the case already assembled. Tavita measured the change on her own team: the median time from return to customer contact went from 19 hours to just under 4. The time her analysts spend on each complex case went up, not down. “They used to spend twenty minutes finding things and five minutes thinking,” she said. “Now it is the other way around, and the customers notice.”

There is one thing Tavita will not let an agent do, and she was clear about it. An agent never sends money again on its own. It can prepare the corrected payment, check it against the return reason, and put it in front of a person. But the person who presses the button must read the beneficiary details with their own eyes. A wrong resend is the one mistake in her area that is hard to undo, and she wants a name next to every one.

Onboarding is the queue that worries Claire most, and Priyanka Raghunathan runs it from Singapore. Before a business can hold money with Tradewind, someone must confirm that the business is real, that the people who control it are who they say they are, and that what it does matches what it told the company. The documents arrive from more than 60 countries, in a dozen languages, from registries that each format a director list in their own way.

The agents read everything first. They extract the company number, the registered address, the directors, and the ownership chain from each document. They compare the documents against each other and against registry data. They check the website and the stated business against what the payment patterns in the first application suggest. Then they write a review memo with every mismatch flagged and every source linked. A reviewer used to spend about 47 minutes on a standard case. Now the median is 16, and Priyanka says the reviews are more careful than before, because the reviewer starts from a list of questions instead of a pile of PDFs.

The team has a firm line here. No agent approves or declines a business. An agent can recommend, and it can close a case that is missing documents after the customer goes quiet for 30 days, but every approval and every decline carries a reviewer’s name. For higher-risk cases, it carries two. Priyanka told me about an application from a logistics company in which every document was perfect. The agent found nothing wrong. A reviewer named Wen Jie noticed that the director had signed three different ways across four documents, and asked for a video call. The call never happened. “The agent was right that the documents matched,” Priyanka said. “Wen Jie was right that something was wrong.”

When a reviewer overrides an agent recommendation, the reason goes into a log the team reads every Friday. About 7 percent of cases get an override. Most of those become a new check the agents run on the next case. The signature check is one of them now.

Oluwaseun Adebayo leads reconciliation from London, and his description of the job is short: every night, the team must prove that the money Tradewind says it holds is the money that sits in its accounts at partner banks. The difficulty is that there are more than 40 of those banks, in more than 20 currencies, and each one sends its statements in its own format, at its own time, with its own idea of what a value date means.

Before agents, the team matched about 94 percent of lines automatically with rules, and four people spent their days on the other six percent. That six percent was around 11,000 lines a day. Most of it was timing: a payment that left one day and arrived the next, or a bank that batched fees into one line at the end of the month. The rest was real breaks, and the real breaks hid inside the timing noise.

Now an agent works through every unmatched line before a person sees it. It looks for the partner line on the following days, checks the bank’s fee schedule, groups lines that a bank netted together, and writes a short reason for each match it proposes. The proposals go into a review queue that a person clears in batches. The unmatched lines that remain after that pass are usually under 400 a day, and they are almost all real. Oluwaseun says his team now finds a true break in an average of 2.3 days, down from 9.

Item 61 on the handoff note that Tuesday was one of those. The agent had marked it as a probable timing break: a USD credit at a partner bank that had arrived a day late. Claire did not believe it. The amount was off by exactly the size of a correspondent bank fee the partner had never charged before. She was right. The partner had changed its fee terms without notice, and the change touched 1,280 payments over three weeks. Melbourne called the bank before breakfast. The agents now check every partner’s fees against the contract each week.

Card disputes are the queue with the hardest deadlines. When a cardholder disputes a charge on an Tradewind card, the card network gives the team a fixed number of days to respond, and a missed deadline loses the dispute no matter how strong the case was. The team handles about 1,100 disputes a month, and before the change it lost roughly 4 percent of them on time alone.

Agents now build each dispute file the day it arrives. They gather the transaction record, the merchant details, the cardholder’s history with that merchant, and any messages from the customer. They draft the response in the format the network expects and put a deadline on the item that the handoff note always shows. No dispute has been lost on time in the last five months. A person still reads and submits every response, because the wording of a dispute response is a promise to the network, and Claire wants that promise to come from her team.

The dispute work is also where Claire saw the biggest change in what her people do all day. Two analysts who used to spend most of their time copying data into response forms now spend it on patterns. One of them found a merchant category where disputes ran at nine times the average, traced it to a confusing statement descriptor, and worked with the merchant to change it. Disputes in that category fell by 70 percent the next month.

Claire has hired 14 people in the past year, and she has changed what she looks for. She used to hire for accuracy and speed, because the work was mostly volume. Now the volume goes to agents, and what is left is the case that does not fit. So she hires for what she calls the second question. Any analyst can ask why a payment failed. She wants the person who then asks why it failed in this way, for this customer, this week.

Her interview has one exercise. She gives the candidate an agent’s draft on a real, anonymized case, with one subtle error in it. She does not say there is an error. About a third of candidates find it. Of those, she hires the ones who also say what they would change so the agent does not make it again. “I am not hiring people to check the agent’s work,” she said. “I am hiring people who will make the agent better every week. Those are different people.”

Each analyst keeps a small number of agent instructions, and the change log shows their names. Tavita’s team edited the return instructions 38 times last quarter. Most changes were one or two lines: a new bank that formats names in a particular way, a corridor where a return code means something different than the network says. Claire reads the change log on Mondays the same way she reads the handoff notes, and she mentions the good changes by name in the weekly team call.

On my last day with the team, I asked Claire what she would protect if she had to give up everything else. She did not talk about the agents or the numbers. She talked about a small business in Auckland that imports furniture from Indonesia. Last year a supplier payment of about 38,000 dollars came back on a Friday afternoon, because the supplier’s bank had merged with another bank and changed its routing code. The owner had a container on the dock and a supplier who would not release it without payment.

The agent caught the return in minutes and built the case. A Melbourne analyst saw that the customer had a container deadline in the payment reference, called the customer before the customer called them, found the new routing code with the agent’s help, and had a corrected payment approved and on its way before Singapore took the queue. The owner wrote to the team on Monday. Claire printed the email and put it on the wall in the London office.

At ten o’clock that night, she opened her laptop at the kitchen counter again and read the note for Melbourne. It was three screens this time. She added one comment, on a dispute with a deadline in two days, and closed the laptop before 10:05.