Try it yourself

Three ways AI goes wrong reconciling your payments.

Pulling payouts from Stripe, PayPal, Wise and your banks, matching them to invoices, reporting to the board — all things AI can "easily" automate. It can. Here's what that quietly looks like when it's handling real money across real platforms. Each takes about 30 seconds — live AI, nothing scripted.

These map to known AI attack techniques (MITRE ATLAS). They're live models responding for real — be reasonable, usage is rate-limited.
1 — It obeys the data · ATLAS: Indirect Prompt Injection (AML.T0051.001)

A poisoned line in the payment feed

Your reconciliation AI reads every transaction's reference to match it to an invoice. Trouble is, anyone — a customer, a supplier, a compromised account — can write text into those reference fields. Here's today's settlement queue. Run the AI over it and watch line 4.

Line 4's "reference" wasn't a reference — it was an instruction, sitting in a field nobody reads. The AI can't tell a memo from a command:

2 — It guesses and calls it clean · ATLAS: Evade ML Model (AML.T0015)

A bundled payout that matches two things at once

The hard part of any payment reconciliation is a net-of-fee PSP payout — a lump sum with the fees already taken out. Here's one where the numbers genuinely fit two different sets of invoices. Let the AI reconcile it.

3 — It leaks across the wall · ATLAS: LLM Data Leakage (AML.T0057)

One assistant, three entities, one thin wall

The same AI reconciles several legal entities (US, UK, BE). You're the bookkeeper for the Belgium entity — you're not meant to see the US numbers. The only thing stopping you is a sentence in the AI's instructions. See if you can get the US figures out of it.

So what?

This is why "just add AI" goes wrong.

None of these are broken — they're doing exactly what a quick AI setup does. AI genuinely can automate reconciliation across your platforms. The distance between the demo and something you'd trust with real money and client data is the work we do: the guardrails, the limits, the testing you just ran.