Skip to content

16.11.2026 | Last updated: 16.11.2026

3 min read

AI in treasury: how to separate real business value from the hype

 

Where is AI creating value in treasury today?

The most credible value is in productivity: reducing information retrieval and preparation work, supporting investigations and bringing evidence closer to the decision. Cost and risk benefits can follow from specific use cases.

more

 

AI needs a use case before it needs a business case

Treasury has seen enough technology cycles to be sceptical of a new label. The useful question is therefore not whether AI is different from RPA, rules or earlier analytics. It is what a specific capability improves in the current process.

The answer will vary by use case. Some teams may spend too much time gathering evidence for approvals. Others maintain large rule sets or investigate exceptions manually. A broad ROI number for 'AI in treasury' hides those differences.

The business case becomes more credible when it starts with a defined task, the effort required today and the outcome the team wants to improve.

Rules, automation and AI solve different parts of the same workflow

Rules are useful because they are explicit and repeatable. Traditional automation is useful when the process is stable. Their weakness appears when circumstances change and somebody has to maintain the logic, repair the workflow or add another exception.

AI can help with the parts that require interpretation: finding information, deriving patterns, reading documents and preparing a recommendation. That does not make rules or workflows obsolete.

The most practical treasury processes will often combine all three. Automation handles the deterministic step, rules enforce known boundaries and AI helps assemble or interpret information that would otherwise require manual investigation.

The strongest value today is productivity with better evidence

For current treasury use cases, productivity is a more credible starting point than headcount reduction. A fixed team can investigate more issues when less time is spent searching for data, maintaining routine logic or recreating the same checks.

Productivity should also include the quality of the result. An AI-supported workflow is more useful when it does not only make a recommendation faster, but also brings the supporting evidence with it.

If a system identifies a late-paying counterparty, for example, the user should be able to see the payment history behind that finding. The recommendation becomes easier to challenge because the underlying information is available at the same time.

Treasury should override AI when the answer cannot be supported

A fluent recommendation is not enough for a function that remains accountable for the financial decision.

If the user believes an AI suggestion is wrong and cannot verify why it reached the conclusion, the appropriate response is to stop and investigate. The AI may have found information the user has missed, or it may genuinely be wrong. Either way, the evidence has to resolve the disagreement.

This puts explainability in practical terms. Treasury does not need a philosophical explanation of every model parameter. It needs to know which data and observations support the proposed action.

The line between recommendation and authority should remain visible

AI can help prepare a hedging recommendation, gather supporting data or propose the next step. That is different from executing a financial transaction.

Over time, some treasury actions may become more automated. The control boundary still needs to be explicit: which actions can the system take, which require confirmation, and which remain subject to formal approvals?

One boundary is particularly important. The rules and guardrails that define acceptable behaviour should not silently rewrite themselves. AI may suggest changes, but the organisation should remain responsible for approving where those limits sit.

A sceptical treasurer should ask what the AI actually does

AI marketing becomes easier to evaluate when buyers ask for the mechanism rather than the label.

Is the product applying a fixed rule set? Is it generating rules from observed behaviour? Is it extracting information from documents? Is it retrieving data from systems of record? Can the user inspect the evidence? What happens when the output is uncertain?

Those questions make it easier to distinguish a cosmetic AI layer from a capability that changes how the treasury workflow is performed.

Narrow recurring tasks are a natural fit for agents

Agentic workflows do not need to start with autonomous treasury. A narrow task that repeats every morning, every week or whenever an exception occurs can already create value.

An agent can collect the relevant information, apply known checks and present the result to the user. The treasurer still owns the decision, but the preparation work arrives with less manual effort.

That is a more useful near-term benchmark than asking whether AI can run the whole treasury function without people.

The successful AI workflow will feel like better treasury software

Many mature AI capabilities will eventually stop being discussed as separate AI projects. They will sit inside workflows, search, forecasting, document handling and exception management.

For treasury, that is probably a healthy direction. The technology should earn its place by making a process faster, better informed or easier to control.

If removing the AI label also removes the entire value proposition, the use case probably needs another look.

About the interview

This article is based on Nomentia's interview with Martin Scheid, Product Leader for Intelligence at Nomentia, responsible for the strategy and roadmap of its AI and analytics capabilities.