AI and currency risk: what it does well, what it does badly, and where it becomes dangerous

Finance departments did not wait for vendors' promises to try generative AI: most treasurers have already asked an assistant to summarise a central bank statement or rephrase an email to the bank. The question is no longer whether to use it, but what for — and above all what not to use it for.

On currency risk, the dividing line is clear. AI is excellent at anything that involves reading, sorting and writing. It is poor, and sometimes dangerous, as soon as it has to produce a market number or take a decision.

What AI cannot do: predict the rate

It is the first question a CEO asks, and the answer is disappointing: reliably predicting the direction of EUR/USD a few months out remains out of reach, with or without AI. Since Meese and Rogoff's work in 1983, studies have shown that it is very hard to beat a naive forecast — "the rate will not move" — and that the models which manage it do so fragilely, depending on the period, horizon and method chosen. A model that found a lasting edge would in any case be exploited until the edge disappeared.

What can be anticipated better is the size of moves — volatility. And the best indicator available is not an AI model: the reference review on the subject concludes that option-implied volatility forecasts future volatility better than statistical models. That is useful for sizing a hedge or a budget rate. It is not a forecast of direction, and it does not tell you whether to hedge today or wait.

The practical consequence is simple: AI changes nothing in the logic of a hedging policy. You hedge because you do not know, not because you think you know. A company that delays its hedges on the strength of a forecast, human or algorithmic, is speculating.

What it cannot do either: give a price

A general-purpose chat assistant does not, by default, have current market data. Asked for today's rate, six-month forward points or an option premium, it answers anyway — with a plausible number, stated confidently, and often wrong.

Even connected to the internet, it lacks what makes the price of a hedge: the spot rate at the minute of the quote, the forward-points curve, implied volatility by maturity and strike, and the product's exact conventions. An option premium recomputed without the right volatility is not an approximation: it is the price of a different product.

• An exchange rate is checked against a time-stamped market source, never in an assistant's answer.
• Forward points are computed from both currencies' rates at the exact maturity, not from a policy rate.
• An option premium requires the implied volatility for the relevant strike and maturity — data general-purpose assistants do not have.

Where AI really pays off

Take away forecasting and pricing and what remains is most of a mid-market treasurer's time, spent reading, consolidating and explaining. That is where AI saves hours.

1. Reading documents

Term sheets, bank confirmations, commercial contracts, account terms: AI extracts their key features — amount, maturity, barrier, strike, a contract's currency clause — and lays them out in a table. On a portfolio of twenty hedges, that is a day of re-keying saved.

2. Consolidating exposure

An export of customer and supplier invoices, a sub-ledger or a multi-currency statement lends itself to a first classification: by currency, by maturity, by counterparty. AI also spots inconsistencies — a dollar invoice attached to a German supplier, an obviously wrong due date.

3. Preparing the cash forecast

Statistical and machine-learning models bring a real gain on frequent, recurring flows: customer receipts, actual payment delays, seasonality. They bring little on rare, large flows — a down payment, an acquisition, a big contract — which remain a matter of judgement. A 13-week forecast by currency usefully combines the two.

4. Following the news and summarising

Summarising an ECB or Fed decision, preparing a review of the portfolio's currencies, translating a bank note: summarisation is a strength of language models. Dates and figures quoted are checked against the official source, because that is where errors slip in.

5. Writing

A first draft of an FX policy, a memo to the executive committee, a renegotiation request to the bank: AI produces a structured outline in minutes. The substance — hedge ratios, delegations, thresholds — remains the company's decision.

Where it becomes dangerous

1. 1
   When it decides

   An agent that places or adjusts hedges without human approval commits the company to significant amounts. Treasury governance — who decides, up to what amount, with what record — applies to a tool as it does to a person.

2. 2
   When it reassures

   Ask an assistant whether to hedge now and it will build a convincing case in the direction of the question. AI is an excellent advocate, not a judge. Ask it instead to list the arguments against your instinct.

3. 3
   When it is given sensitive data

   Bank details, hedge amounts, counterparty and customer names do not belong in a consumer tool whose terms allow data reuse. An enterprise offering that excludes it, or prior anonymisation, is required.

4. 4
   When its numbers go unchecked

   A wrong rate, premium or date copied into a committee memo becomes a decision taken on a false number. The rule: no market number comes from AI.

What to hand to AI, and with what check

The three-check rule

1. Every market number has a time-stamped source: rates, forward points and volatility come from market data, never from a generated answer.
2. Every extraction is checked against the document: barriers, amounts, maturities and doubling clauses first.
3. Every decision stays human and recorded: AI prepares the file, an authorised person decides, and the decision is logged like any other.

Where to start in a mid-sized company

No need to wait for an IT project. Three uses can be in place within a week, without integration:

• Have the features of current hedges extracted into a table, then review it: this is also the starting point for mark-to-market tracking.
• Have twelve months of invoices classified by currency and maturity, for a first measure of exposure.
• Have the outline of the monthly committee memo drafted, for the team to complete with its checked figures.

The only prerequisite is a written rule: which tools, what data may go into them, and who approves what comes out. It fits on half a page and is added to the FX policy.

Sources

• Meese, R. and Rogoff, K. (1983), "Empirical exchange rate models of the seventies: Do they fit out of sample?", Journal of International Economics, 14, pp. 3-24.
• Rossi, B. (2013), "Exchange Rate Predictability", Journal of Economic Literature, 51(4), pp. 1063-1119.
• Poon, S.-H. and Granger, C. (2003), "Forecasting Volatility in Financial Markets: A Review", Journal of Economic Literature, 41(2), pp. 478-539.
• Borio, C., McCauley, R., McGuire, P. and Sushko, V. (2016), "Covered interest parity lost: understanding the cross-currency basis", BIS Quarterly Review, September 2016.
• Data reuse by assistants: OpenAI, "Enterprise privacy" (enterprise data is not used for training by default, unlike consumer plans).