Treasury holds the cash, the payments and some of the most sensitive data in a business. Getting an answer 30 seconds faster is useful. Getting the wrong answer 30 seconds faster can cost millions.
So the question that matters for treasury leaders is whether AI can work inside treasury’s controls, to treasury’s standard of accuracy. We put that question to John Keeton, VantagePoint’s Practice Lead for Kyriba in APAC, and set his answers against the latest research on how treasury teams are adopting AI.
Ambition is running ahead of practice
Treasury has moved AI up its agenda quickly. In the 2026 AFP Treasury Benchmarking Survey, 30% of practitioners placed AI and automation among their top five priorities. Cash and liquidity forecasting, the problem many hoped technology would fix, still tops the list at 63%.[1]
Practice is slower. Crisil Coalition Greenwich found that around half of large companies had not deployed AI in treasury at all, and fewer than one in ten had built it into daily work such as forecasting or fraud detection. The barriers treasurers cited most were a lack of in-house expertise and integration problems.[2] Anyone who has run a treasury system project will recognise both.
Governance is lagging too. AFP respondents rated their AI policies at 2.9 out of 5, the weakest of any policy area measured, and only 34% considered themselves effective in AI knowledge.[1]
For a treasury leader, the picture is a gap between appetite and readiness. Closing it is an operating decision as much as a technology one. Data, controls and team capability decide how far any AI tool gets.
How Kyriba’s TAI works
Kyriba launched Trusted Agentic AI (TAI) in May 2025. It runs on an embedded large language model rather than a third-party integration, and Kyriba says customer data is not used to train it.[3][4]
“The simplest way to think about TAI is as an AI agent that sits within Kyriba,” says John. “It lets you interact with your treasury data in plain language. Rather than navigating screens, reports and datasets yourself, you ask: What’s my cash position today? Which payments are waiting for approval? Have all my bank statements arrived?”
The more significant shift is from answering questions to doing work. TAI can retrieve information, analyse it, make recommendations and run multi-step workflows through what Kyriba calls Skills. A daily liquidity check and a recurring forecast review are typical examples: processes that would otherwise mean someone working through several steps by hand.
“That’s where the distinction between generative and agentic AI matters,” John says. “You move from ‘tell me something’ towards ‘help me do something’.”
Less finding, more deciding
Treasury is a reasonable fit for AI for a plain reason: it runs on structured data. Bank statements, balances, forecasts, payments, exposures, counterparties, entities and currencies all arrive in predictable shapes. Treasury teams then spend a large share of their week finding, reconciling, comparing and explaining that data.
“That’s exactly the kind of work AI can be very good at,” John says. “Where I’d be more cautious is the idea that AI replaces treasury judgement. It doesn’t, and I’m not sure that’s desirable anyway. The opportunity is to remove more of the work that happens before the judgement.”
He puts it in hours. “Don’t spend two hours finding and preparing the information you need to make a decision. Get there in seconds, and spend those two hours understanding what it means and deciding what to do about it.”
The example he returns to is the first question of the day: is there anything I need to worry about? Behind it sit dozens of smaller checks, from unusual movements to the biggest forecast variances.
Today that question means a tour of reports and screens. With an agent, it can be one request. And once a team is asking the same questions every morning, those questions can become a Skill that runs the checks the same way every day.
That consistency is the leadership point. A morning check that depends on who is in the office is a key-person risk. One that runs as a defined process is a control.
What it looks like in practice
It is early, but there are results. HelloFresh was the first company in Germany to use TAI. Its treasury supports a business generating billions in revenue with a lean team.[5]
Felix Luber, HelloFresh’s Director of Group Treasury, describes a problem every treasury team knows: tracking down a payment that won’t reconcile. A search that once took a long time now returns an answer within seconds.[5] The team also uses TAI for system questions, the kind that used to mean digging through implementation documentation.
“Neither of those examples sounds futuristic, and that’s why I like them,” John says. “Some of the best applications of AI in treasury will be quite boring. But boring things that happen hundreds of times a year add up very quickly.”
The wider market points the same way. In Coalition Greenwich’s 2026 survey of 142 major companies, treasurers saw the most value from AI in account reconciliation. More than half were already using AI there or planned to within 12 months. Nearly half had applied it, or planned to, in cash forecasting, reporting and liquidity planning.[6]
Trust must be designed in
Treasury is the business’s front line against payment fraud, and the threat is not easing. In AFP’s 2026 fraud survey, 76% of US organisations reported attempted or actual payments fraud in 2025. Business email compromise, where fraudsters pose as a supplier or executive to redirect payments, targeted 74%. Treasury was the department most likely to discover attempted fraud, in 83% of cases. Yet only 17% of organisations were using AI to fight it.[7]
In other words, treasury is already catching most of the fraud, and mostly without AI. AI could take some of that load off. But give an agent access to bank accounts and payment workflows, and you’ve opened a new door that needs locking as tightly as the rest.
John’s view on output is plain. “Treasurers should trust it appropriately. I wouldn’t recommend anybody follow an AI recommendation involving millions of dollars because the AI sounded confident. Treat it like a very capable analyst. If they tell you something that could lead to a material financial decision, you still want to know where the answer came from.”
On access, he points to how TAI is built. It respects the user’s existing role-based permissions, so asking the agent for something you can’t see in Kyriba won’t reveal it. Kyriba states that agents can only see and do what they are authorised to in the platform, that actions run only after a person approves them, and that every action can be explained, traced and audited.[8]
“It isn’t one all-powerful user sitting above your existing controls,” John says. “Asking TAI to investigate something is very different from giving an AI carte blanche to move money. For me, the test is whether AI can do this while keeping the permissions, controls and auditability we’d expect from any other treasury process. In treasury, that’s a much higher bar.”
The vendor can build the guardrails. Deciding how to use them is the treasury leader’s job: which actions need approval, who reviews AI output before it reaches a decision, and what gets logged. AFP’s 2.9-out-of-5 score for AI policy suggests most teams have yet to write that down.[1]
Why the unglamorous use cases win
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027. The reasons it gives are rising costs, unclear business value and inadequate risk controls. Its advice is to pursue agentic AI only where the value is clear.[9]
Treasury is well placed on that test. Much of its workload repeats daily or weekly, so you can count how often a task happens and how long it takes. That makes the value of automating it easy to measure.
“There will be AI functionality that saves somebody three minutes a month,” John says. “I’m not particularly interested in that. Tell me why my forecast moved. Show me where I have excess liquidity. Do that kind of thing reliably and repeatedly, and the cumulative benefit becomes significant.”
Where treasury leaders should start
Fix the data before you ask it questions
An agent can only answer from what the treasury system holds. Coalition Greenwich found about 40% of treasury departments run on several ERPs with limited integration, and roughly one in five large companies still rely on spreadsheets and manual processes. Fewer than a quarter list data quality as a top priority.[6]
Start with the questions you already ask every week
List the checks your team runs daily and weekly, and time them. Those are your first Skills.
Use AI to keep what your team knows
Treasury teams are small, and expertise often sits with one or two people. When they leave, their knowledge of why the system is set up as it is tends to leave too. Skills and AI-searchable documentation give a new joiner somewhere to start, and give the team a record of its own processes.
Keep the system current
Kyriba ships new functionality every month. Many teams go live, then stop looking. A periodic review of reporting, connectivity and new features is how an investment keeps paying back, with or without AI.
Curious, not carried away
Asked whether treasurers should be excited about AI, John chooses his words carefully. “Curious rather than carried away.”
The research supports that stance. Treasury has made AI a priority, and the early results in reconciliation and investigation are real. The governance, data and skills to use it well are still being built, and they will decide which teams get value.
“I don’t think the objective should be an autonomous treasury department,” John says. “It’s a treasury team that spends far less time finding and preparing information, and far more time understanding it and deciding what to do next.”
That is a modest ambition. It is also the one most likely to pay off.
References
- Association for Financial Professionals, 2026 AFP Treasury Benchmarking Survey: press release, 15 September 2026. Survey of 425 treasury practitioners, May 2026.
- Tobias Miarka, Crisil Coalition Greenwich, AI in corporate treasury: what are the barriers to adoption?, Deutsche Bank flow, 26 February 2026. Based on the July 2025 Treasury AI Insights Study of 119 large global corporates.
- Kyriba, Kyriba unveils agentic AI solution TAI, press release, 13 May 2025.
- Kyriba, Kyriba Trusted AI fact sheet.
- Kyriba, HelloFresh accelerates payment reconciliation in seconds with Kyriba Trusted Agentic AI, customer story.
- Tobias Miarka, Crisil Coalition Greenwich, Tech in corporate treasury: AI is not the whole story, Deutsche Bank flow, 18 June 2026. Based on the 2026 Corporate Treasury Technology and Innovation Report, a February 2026 survey of 142 major companies.
- Association for Financial Professionals, 2026 AFP Payments Fraud and Control Survey: press release, 14 April 2026. Survey of 465 US treasury practitioners, January 2026.
- Kyriba, TAI agentic AI product page.
- Gartner, Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027, press release, 25 June 2025.


