Can an AI agent handle foreign currency revaluation?

Yes, within tight bounds. Revaluation is rule-based: restate each foreign-currency monetary balance at the closing rate and book the unrealised gain or loss. An AI agent can do this at depth 3 once the rate source and account list are fixed. The financial controller approves the rates each period and any result outside expectations.

What makes revaluation a good fit?

Revaluation is repetitive, calculation-heavy and governed by a short list of rules. Each month the same accounts are restated at a new rate and the previous entry reverses. Errors usually come from inputs, not arithmetic: the wrong rate, an account left off the list, a non-monetary balance revalued by mistake. Those are exactly the things an agent can check every time.

What it should not do is choose the inputs. Which rate is the closing rate, and which accounts are monetary, are decisions finance makes once and maintains.

What conditions decide whether the agent posts or drafts?

The agent posts at depth 3 only when all of these hold:

• The rates come from the agreed source. Loaded from the group's rate provider or the central bank feed the policy names, as a sealed input. A rate typed into a spreadsheet or found on a web page is not acceptable.
• The account list is maintained. Only accounts flagged as monetary and foreign-currency are revalued. Stock, fixed assets and prepayments are not.
• The period is open. The agent cannot post into a locked period.
• The result is within expectation. The total gain or loss sits inside a range the controller set, based on balance sizes and rate movement.
• Auto-reversal is set. Each revaluation entry reverses on the first day of the next period.

If any fails, the agent drafts the entry at depth 2 and holds it for the controller.

What does the agent do, and what does the controller approve?

The sign-off that matters is on the rate table. Once the rates are right, the rest is mechanical. Get them wrong and every entry is wrong in the same direction.

An illustrative example

Take an exporter that invoices customers in two foreign currencies and holds a foreign-currency bank account. At month end, the agent loads closing rates from the central bank feed and compares them with the prior month. One currency has moved sharply.

It identifies the foreign-currency receivables, payables and the bank balance, restates each, and calculates a total unrealised loss of 64,000. The controller's expected range for this month, given the rate move, tops out at 40,000. So the agent does not post. It drafts the entry and breaks the loss down: 38,000 on receivables, 26,000 on a foreign-currency loan that someone added to the ledger mid-month without flagging it as monetary.

The controller confirms the loan should be revalued, updates the account list, and approves the entry. The agent posts it with auto-reversal and records both the approval and the list change.

What controls make it safe?

• Sealed rate source. One named source, loaded automatically, with the load recorded.
• Account list owned by a person. The agent cannot add or remove accounts.
• Expected range. Set by the controller, so a large move triggers review rather than posting.
• Auto-reversal on every entry. Prevents gains and losses from compounding across periods.
• A record per run. Rates, balances, calculation and verdict are filed together.

The step-by-step set-up is in how to run foreign currency revaluation with an AI agent. Revaluation measures exposure; it does not manage it. Reducing exposure is a treasury decision, covered in whether an AI agent can hedge foreign exchange exposure. Converting money for customers is a different activity again, set out in whether an AI agent can exchange currency for a customer. The wider finance controls are in AI agents for finance: books you can audit.