As the solution engineer at Palm, I sit in treasury meetings almost every day. Which means I spend most of my time with group treasurers, cash managers, and treasury analysts who are walking me through what their day actually looks like. Different industries, different sizes, different geographies. The pattern is remarkably consistent.

The friction is not in the decisions themselves. Treasurers are good at decisions. The friction is in the prep work that has to happen before a decision can be made. Consolidating balances from a dozen places into one view. Reconciling what one categorisation tool produced against what another tool produced. Re-spreading a flat forecast into something that matches reality. Finishing a forecast that, by the time it is ready to share, already describes a week that has passed. Re-explaining a variance because the version someone forwarded was last Tuesday's. By the time the picture is complete enough to act on, most of the day is gone, and the question that actually deserved thirty minutes of strategic thinking gets five.

This post is about where I see AI replacing that prep work first. Not a platform overhaul, not a multi-quarter implementation. Six specific places where the gap between "I have the data somewhere" and "I have a decision I can defend" is large enough that closing it changes how the team spends its week.

1. One view of cash, forecast, and a funding plan you can act on

This is the use case I would start with for almost any team. Most treasurers have decent bank connectivity now. That is not the problem. The problem is that connectivity gives you balances, and balances on their own do not tell you what to do.

The version that does is one consolidated view that combines current balances across every entity, currency, account, and account type (operational, money market, time deposit, notice account, collateral), overlaid with the forecast for the next several weeks, with the answer already pre-computed: which entity-currency positions are heading toward a shortfall, when, and what is the best way to fund them.

"Best" here is not a vague suggestion. It follows the priority that any treasurer would apply manually if they had time. First, same entity, same currency, different account. Then, different entity, same currency, via an intercompany transfer. Only after that, an FX trade. Constraints are injected automatically. Capital-control countries flagged. Counterparty and concentration limits respected. Time deposit maturities recognised as natural liquidity that may eliminate the need to act at all. The output is a ranked list of actions with amounts, source and target accounts, and deadlines.

The shift this creates is the one people feel fastest. The morning stops being about assembling the picture. It starts with the picture already in front of you, and your job is to confirm or override.

2. AI Transaction categorisation that does not need rules

Every analysis downstream of bank statement data depends on categorisation being right. Forecasting, variance, cash positioning, the funding plan above, reporting to the CFO. All of it.

The traditional approach is a rules engine. Someone writes hundreds of rules over the years, owns them, and fixes them when they break. They break constantly. A new entity gets added. A bank changes the format of a description field. A region introduces a payment type that does not match anything in the catalog. Each of those is a maintenance task nobody wants, and the result is that categorisation quality decays, along with trust in everything built on top of it.

AI categorisation works differently. It learns from the patterns in your historical data, adapts to corrections the team makes inline, handles descriptions in any language, and continues to work when a new entity is onboarded without anyone writing a single new rule. It is also forgiving of the messy reality of statement data, where the same bank uses different formats across different countries and the same counterparty appears under three slightly different names.

Categorisation is not glamorous. It is the foundation that makes the other five use cases reliable.

3. Cash forecasting that improves accuracy at the daily level

The most common forecasting pattern I see, even at sophisticated treasury teams, is some version of: take a monthly figure from FP&A, divide by the number of days, and call that the daily forecast. Or: assume 70% of receivables land on time and ignore the other 30%. Or: spread a quarterly tax payment evenly across the quarter even though everyone knows it hits on a single date.

Cash does not move like that. Payroll concentrates on specific days. Supplier runs follow weekly cycles. Customer receipts cluster at the beginning or end of the month depending on the segment. Tax obligations hit fixed deadlines. A flat forecast is a forecast that is wrong every single day of the month, and the average accuracy looks fine only because the highs and lows cancel out.

The version that works learns the actual timing and amount patterns at the category and account level. It looks at how this entity's customer receipts have historically landed across the days of each month, not just the total. It recognises that the manufacturing subsidiary pays its largest supplier every second Tuesday. When a forecast misses, it tells you which category drove the miss and whether it was a timing issue or an amount issue, so you know whether to adjust your assumptions or wait for the cash to land next week.

4. Cash Forecast variance by category, entity, and timing

This is the use case that most treasurers I meet have quietly accepted as impossible to solve. They have stopped trying, because the answer has always been "let me get back to you on that."

The CFO asks what happens if three of the largest customers pay a week late. The board wants to know why this week's forecast looks materially different from last week's. The auditor wants to understand which categories drove the variance in Q1 in two specific regions. Every one of those questions, today, triggers the same response. Open the spreadsheet, find the right tab, re-link the formulas, re-consolidate, format the output, send it over. By the time the answer is ready, the meeting where it mattered has ended.

The version that works decomposes variance automatically by category, account, entity, and the split between timing and amount. What-if scenarios run against the live forecast in seconds, not days. The CFO can ask the question and get the answer in the same conversation. That changes the role of treasury inside the company more than any single other thing on this list.

5. Continuous treasury policy monitoring and AI Alerts for FX, Investment, and covenant Compliance

Most treasury policies are checked at month-end. The investment policy is reviewed when the quarterly report is produced. FX policy compliance gets verified when the variance report is run. Covenant headroom is calculated when the lender requires it. In between, the team operates on memory and a healthy dose of caution.

What I see in real numbers is that policies drift. An FX policy that says "no more than 30% of total cash in non-functional currency" can sit in breach for months before anyone notices, because no one was looking continuously. An investment concentration limit set at 15% per counterparty can creep well past that, because the money market fund yields are better than the term deposits and nobody flagged the cumulative effect. The first time anyone sees the problem is in the compliance pack that goes to the CFO at month-end, by which point the breach has been live for weeks.

AI monitoring runs against live position data. Before an investment is placed, a pre-trade check confirms it will not put any counterparty over its limit. As FX positions drift, an alert fires when the threshold is approached, not after it is crossed. Covenant baskets are tracked continuously, so the team sees the runway shrinking before it becomes a board-level conversation. Governance moves from periodic to continuous, and the surprises stop.

6. The automated treasury analyses you know you should be doing

Every treasurer I talk to has a list of things they wish they were doing regularly but cannot fit into the week. Bank fee outliers, where the bank quietly applied a default rate card to half the accounts. Forecast bias by region, where one country is consistently 12% high and nobody has corrected the assumption. Intercompany flows that look like they should be netting but are not. Dormant accounts paying maintenance fees for no reason. Seasonality lookups for any entity, any period, on demand.

None of this is conceptually new. Treasurers know these analyses are valuable. The reason they do not run them is that pulling the data, cleaning it, and structuring the output takes longer than the week has. So the analyses get scheduled, deprioritised, and quietly dropped.

This is where AI becomes a multiplier in the most literal sense. The analysis that used to take three days to produce is generated on request. Bank fee patterns are surfaced automatically across every entity and currency. Forecast bias is computed continuously. Anomalies in intercompany flows are flagged before the monthly review. The team gets the analyses they always wanted, without having to choose between those and the next-day position.

The common thread

The teams I see getting the most out of AI in treasury are not the ones running the most ambitious projects. They are the ones who identified one workflow where prep work was eating most of the day, replaced it, and used the time it freed up to focus on the next one. The technology is the easy part. The discipline is in picking the right starting point.

If I had to pick the single highest-leverage place to start, it is the first one on this list. Everything else is faster and more reliable once the morning view is sound. The other five compound from there.