Mathew·Blog

Part 6 of 6 · The one-person finance team

How do you choose a spend management system?

What I needed, why Airwallex won, and how much of the running I handed to Claude.

In this part

  • Choose your spend philosophy first: control who can pay, or give everyone a card and control the limits.
  • One card per budget line, each mapped to one Xero account and one team tag, meant every swipe arrived in the books already coded.
  • Claude ran the daily bill-pay inbox up to a saved draft. Keep a human wherever money moves, and set agents' limits in platform settings, not prompts.

New to the series? It follows one finance lead rebuilding a startup's finance function with Claude. Start with Part 1.

By mid 2026, the monthly ledger was nearly 3x the FY25 average, and growing. But spend controls were still the FY25 ones, built for a team of about 20, where three people initiated almost all payments.

Figure to come: Counts from the Xero general ledger; no amounts.

In broad strokes, there are two schools of thought on spend control.

The traditional view is to control payment methods. Only a select few authorized people get cards to spend company money. Everyone else pays out of pocket, then requests a reimbursement.

The other view is to control limits. Give everyone in the company a card, then use software to manage each card's limit. Default to zero so every swipe goes through an approval chain, or set a budget that dictates limits for every card issued under it.

We were squarely in the former camp. Only a small group held company cards. About 50 people claimed expenses through a Google Sheet. And by my third week, I'd been roped into 30 budget approval Slack threads.

So I drew up a plan to eliminate reimbursements and switch to budgets and cards.

Figure to come: Expense steps by person · reimbursement vs budget cards

1. What we needed, and how we set it up

First, I wrote down the features the company needed and why:

  1. Hold cash and earn yield. Our cash was sitting idle.
  2. Work as a control layer for whichever spend philosophy leadership chooses. That means granular card limits and approval settings, so the same platform could run tight or loose.
  3. Integrate with Xero and Claude. The books should code themselves as much as possible, and I wanted to build automations on top.

Once leadership decided they wanted a Q3 budget (Part 5), I set up budget owners who each got cards for each budget line under their care.

They could also delegate part of a budget to their team. For example, client-facing roles got sub-limits under a broader client travel and entertainment budget, meant to help hit Q3 targets.

The rules fit in one table:

Control Where it's enforced On a breach
Card limit = the budget line At checkout, by the card The swipe declines; the budget owner asks for a higher limit
Team tag and Xero account On the card, at the swipe Caught in monthly accounting close
Receipt within 14 days Airwallex app, or forwarded by email The card freezes
Reimbursement only by exception Finance approves it in Airwallex Anything else goes back on a card

The trick that made it work: one card per budget line, each mapped to one Xero account and one team tag. The card's limit is the budget, and every swipe arrives in the books already coded.

2. Why Airwallex

Seven contenders went through my three requirements in order, and Airwallex was the only one that cleared all three (though the results may be different now).

Figure to come: Platform selection · 7 options, 3 requirements

One less rational factor mattered too: CEO time, and Airwallex needed the least. The account already existed, just dormant, and the Finance Admin role gave me 80% of the access I needed, with Wise as a backup rail and wallet.

3. How it held up

By September, the ledger had a third more lines a month than in July, and more than three times the FY25 monthly average.

Volume was up, but the team was happier. One said, "Thanks, uploading receipts in the Airwallex app is way simpler [than the old expense process]."

As the transaction load grew, very little of the extra load weighed on me. An Airwallex expense policy AI pre-verified expenses for me, marking the ones that had a receipt, the merchant matched the intended budget category, and it complied with submission guidelines. I spot checked material swipes, then spent my focus on the policy exceptions that needed judgment. Another Claude-augmented workflow for bill pay ran daily, so all I had to do was review and approve (more on this below).

Figure to come: The same counts as the opening chart, with the Airwallex months shaded.

One day, I noticed one of my own cards had been frozen. More than 14 days had passed since I'd swiped it, and I hadn't submitted the receipt. At first, I was angry. How could it freeze my card when I had set it all up? But then relief kicked in, and I realized it was just doing exactly what I had asked it to.

Where it fell short. Three hurdles I wish I'd known about before the switch:

  1. Early on, about every other payment came back with a compliance request for information (RFI). So onboarding was painful. One hack helped: we kept the first transfers deliberately small to trigger the compliance review before real money moved, so KYC would never freeze working capital.
  2. Airwallex supports fewer payment rails than Wise. So for some corridors, we had to use the backup rail, Wise.
  3. Three-currency reimbursements. It's an edge case, but not uncommon for APAC companies: a merchant, the company's funds and the employee each in a different currency. Think of an employee in India buying software from a US vendor, paid back from our SGD wallet. Wise synced these to Xero cleanly; Airwallex often struggled.

How Claude helped

Most of the money that didn't go through a card left through one finance inbox. Every vendor invoice landed there, and Claude, working in Gmail and in Airwallex through Claude in Chrome, followed the same standing instructions each day:

  1. Find new invoices and draft a note asking the budget owner to approve. Give me a link to the drafts when done.
  2. Once the owner approves via email, forward the invoice to Airwallex and add label "Sent to Airwallex".
  3. In Airwallex, match the bill to the vendor, set the wallet and payment date by the rules I gave you, and save it as a draft.

All I had to do was review and approve. In a mature finance team, someone else would release the payments I had approved. But alas, it fell to me.

Where Claude fell short. It once flagged 19 receipts as "dated 2024". The receipts had no year at all: Claude had passed along the Airwallex policy agent's flags without reading them. The rule since: read the receipt, not the flag.

What I'd do differently

Claude does manual, detail-heavy work with fewer mistakes than I do, and it's a strong systems designer. What it lacks is context, so the judgment calls stay with me. So does the accountability: the agents act under my logins, so their mistakes are mine. I give each one only the permissions I'd sign for.

So if I built it again, I'd set up the agents like this:

  1. Give Claude a manifest with limits for read-only agents and read-write agents, and ask it to propose the sub-agent workflow or coordinator setup.
  2. Set each agent's limits myself and enforce them via platform configuration (tool allowlists, connector scopes, the Airwallex role), not in the prompt.
  3. Use as many agents as the trust boundaries need. Here the bare minimum was two: one that reads untrusted email and summarizes a list for my review, and one that acts in Airwallex based on an approved list I give it.
  4. Connect Airwallex and Wise through their APIs from day one, so Claude can keep one live view of cash, cards and bills. I focused on the controls first, and left the visibility for later.
Figure to come: Bill pay with two agents · one trust boundary

Takeaway: Choose your spend philosophy first, use a software platform to enforce it, and keep a human wherever money moves.

By October, costs mostly ran themselves: cards for spend, a daily Claude run for bills, and me on the exceptions. Revenue was a different story.