The opinions expressed here are those of the authors. They do not necessarily reflect the views or positions of UK Finance or its members.

1. AI is about to automate core Treasury functions end-to-end
2. It's overhyped and nowhere near ready. 

Both observations miss the point. The better question is where Treasury's work is actually broken, and what AI can realistically do about that.

Treasury asset-liability management (ALM) is par excellence a judgement function. AI speeds things up and spots patterns, but it doesn't necessarily replace the thinking behind a decision. The use-cases that add most value right now are narrow tools trained on specific datasets for specific jobs, not sweeping "autonomous finance" platforms. Every Treasury professional needs to understand that boundary before touching the tools.

Three areas show why.

1. Execution risk and accountability
Agentic AI, systems that don't just assist but act, sounds efficient until it meets Treasury's core constraint: outcomes are asymmetric. A suboptimal decision is recoverable, whereas an unaudited execution isn't. An agent with delegated access to systems and counterparty data is making decisions on your authority, with no judgement behind it. Scoped access and a full audit trail aren't optional; without them, a recoverable incident becomes an unsolvable one.

2. Fluency mistaken for accuracy
AI output can be dangerously convincing. Summaries appear clean and narratives feel coherent, so it's easy to miss what's absent. Practitioners who've tested these tools on real data report the same problem: the write-up looks strong, but the figures underneath don't hold up. If an AI summary of a regulator’s Policy Statement is “97% accurate”, for a bank that’s still not good enough. In banking, 97% accuracy equates to 100% liability.

3. The productivity illusion
An hour saved feels like progress. But research from a well-known consultancy tells a different story: nearly 90% of companies have invested in AI, yet fewer than 40% report measurable gains, usually because a team automated one task and left the rest of the process broken. If we speed up a bad workflow we simply reach the wrong answer faster!

All three come back to the same gap…

AI has been introduced before organisations have defined where it sits in the control framework.

Treasury teams know how to govern payments, liquidity management and risk exposures, but haven't necessarily applied the same discipline to AI use. There may be no clear line between what an agent can execute and what requires approval, no documented checks on whether outputs are reliable, and no view of how AI fits into the end-to-end process.

Until that changes, AI creates the illusion of control while potentially moving accountability into a grey area. So for a bank Treasury, the first step must be to define the rules: what AI can access, what it can decide, where humans intervene, and who owns (and is responsible and accountable for) the outcome.

Where a written framework earns its keep

A D-SIFI bank recently asked me for exactly this: a guide to using AI to screen legal contracts, not a general chat about the idea. That tells you AI is already inside contract review at banks, whether or not anyone has written down what it's allowed to touch.

So I wrote it down against those same three challenges. The first is scope: defining exactly what AI is being asked to do. In this case, that means extracting clauses, flagging deviations and identifying what is missing. The second is control: keeping legal judgement and final sign-off with a human, so a confident-sounding answer is still subject to review. The third is workflow: showing how AI fits into the full process, from initial contract intake through to the final memo, rather than treating review as a single isolated task.

There is also a case for specialist AI. General contract tools can identify clauses and compare language, but a model trained on banking and regulatory material (including, for example, MCAI!), can add the context needed to understand why a clause matters and where the risk sits.

None of this is complicated

Good AI guidance doesn't need dozens of pages explaining every possible future use case. It needs to make the practical decisions clear: what the system can do, where a person needs to step in, and who owns the outcome.

That is the standard AI training in Treasury should be judged against: the goal is not to create people who can talk about AI; rather, it is to help teams realise where and how to use it with enough confidence and control that it improves the work without weakening accountability.

Professor Moorad Choudhry is an INED and Chair of Risk at SBUK Ltd, in London, and author of The Principles of Banking (Wiley 2012, 2023)