Your First AI Task in Treasury Should Be Unglamorous
Four questions to decide which tasks you can trust AI with, and which must stay with your team.
Use AI first on low risk tasks, where a mistake costs little and someone would spot it quickly. Give it more authority only once it has earned it.
- Allow AI to prepare, human reviewsCash commentary · Forecast versus actual notes · Bank queries · Flagging exceptions
- Allow AI to suggest, human approvesCash sweeps · Hedge execution
- No AI, human decidesPayment release
Examples only, not a full list.
Most Treasury teams are under pressure to show progress on AI. The first question is usually what AI can do. A better question is what happens when it gets something wrong.
An AI mistake can be cheap or costly, depending on the task. If AI drafts a cash commentary with a wrong number, the reader spots it and you lose ten minutes. If AI releases a payment to the wrong supplier, nobody notices until the supplier calls, and the money may not come back. Start with tasks like the first, not the second.
Four questions to ask before handing a task to AI
- What does a wrong answer cost, and can it be reversed?
- Who would notice a mistake, and how soon?
- Can a person check the result against the source data, such as a bank statement or the ledger, in a few minutes?
- Do you need AI at all? A simple rule, a better report or a cleaner process is often cheaper and easier to audit.
Which Treasury tasks can AI automate?
Apply the four questions above, and most Treasury tasks fall into three groups. Here are typical examples of each.
Allow AI to prepare, human reviews
- Cash commentary
- Forecast versus actual notes
- Bank queries
- Flagging exceptions
Why it’s safe: a person reads the output before anything is sent or posted, and the source data is close at hand. If AI gets it wrong, you lose only a few minutes.
Allow AI to suggest, human approves
- Cash sweeps
- Hedge execution
How it works: start only after the AI’s suggestions have been tested and found reliable. A person approves each action, and every approval is recorded. Set limits, such as maximum amounts, in the banking and payment systems themselves, not in the AI’s instructions. Then even a wrong suggestion cannot go past them.
No AI, human decides
- Payment release
- Anything irreversible or time critical
Before automation comes near it: agree limits, approvals and an escalation path. Make sure you can quickly remove the AI’s access to your banking and payment systems, return to the manual process, and test that fallback.
Speed is not the goal. Being able to show your auditor exactly what the AI is allowed to do is.
When to give AI more authority
Before a pilot starts, write down what results would justify giving AI more authority. Useful measures are how often reviewers change the output, how many errors are found late, and how long exceptions stay open. Set the level each must reach. If the numbers fall short, do not widen the AI’s role, and be ready to take authority back.
List every task where AI is used or proposed in your Treasury. Next to each, write what a wrong answer would cost and who would catch it. The list will tell you where to start.
This article is general discussion, not professional or investment advice. Views are the author’s own and do not represent any employer.