Artificial Intelligence in Treasury: Signal, Noise and Practical Value

The signal is real, but the numbers require care

AI is clearly moving into the treasury agenda. HSBC cites a survey in which 51% of corporate respondents were using AI or actively planning to do so, while 77% of identified use cases focused on automating routine tasks. A separate HSBC treasury survey reported that fewer than 30% had implemented AI in finance.

The Cambridge Centre for Alternative Finance found that more than 80% of financial-services firms were adopting AI in some form. It also found that 52% were experimenting with agentic AI. But only 40% reported increased profitability, while 43% reported no change. Use is not the same as value.

These findings are not contradictory. ‘Using or planning’ is not the same as controlled production deployment. They point to genuine interest, a growing pipeline of experiments and a smaller base of implemented use cases. The strongest near-term signal is automation of work around the decision—not replacement of the treasurer.

Forecasting is promising—and easy to overstate

Forecasting is one of the most credible treasury uses because it involves recognising patterns, processing large amounts of data and repeatedly comparing forecasts with actual cash movements. AI can help classify cash flows, identify unusual items, propose assumptions and generate scenarios.

It cannot repair missing ownership, inconsistent data from core finance systems or an undefined forecasting process. HSBC reports that spreadsheets remain the dominant forecasting and planning tool for 32% of respondents. A new model placed over the same fragmented data may create faster output without creating better control. Every pilot therefore needs a baseline—the measured result of the current process—covering accuracy, bias, forecast period, coverage, manual effort and the cost of error.

Practical value sits around the workflow

The most useful early applications are often unglamorous: mapping bank data, classifying transactions, drafting variance commentary, identifying exceptions, summarising policies, preparing scenarios and helping users find the right information.

Economist Impact, in sponsored research covering more than 400 supply-chain executives, found that only one in ten firms had fully integrated AI agents across operations. AI performed best where rules were clear and data was standardised, while only four in ten respondents said AI was aligned with existing platforms. The treasury lesson is straightforward: define the process and data before adding the technology.

Higher-impact uses require tighter control. A forecast recommendation can inform judgement. A payment, investment or hedge creates an external consequence. Treasury should preserve explicit human authority for decisions that move cash, create exposure or alter funding.

Oxford Saïd makes a useful distinction. With agentic AI, a final human check is not enough. Responsibility must be built into the use case, data, permitted actions and escalation points from the start. Some decisions should remain outside autonomous AI. In treasury, that includes decisions that move cash or create financial exposure.

• Define the user and decision before selecting the model.
• Measure performance against the current process, not a demonstration.
• Log inputs, outputs, overrides and material changes.
• Keep sensitive data and third-party model dependencies within the control assessment.

What implementation looks like in practice

Brompton Bicycle uses AI for a narrow production task: helping factory workers choose which bicycle frame to assemble next. The technology supports the manufacturing process. It does not replace the people responsible for building the bicycle.

In marketing, generative AI is replacing parts of the photography and editing process. Amazon's advertising tools can turn an existing product photograph into different lifestyle images. Amazon reports more creative testing and lower production effort. These are company-reported results, not independent evidence.

Lloyds Banking Group provides a regulated-industry example. Its Athena knowledge assistant is used by 20,000 colleagues and has reduced average information-search times by 66%. Lloyds says it deployed more than 50 generative-AI solutions during 2025. The useful starting point was helping people find information and complete defined work more quickly.

PayPal provides the closest treasury example. Its treasury team already uses an AI assistant to ask questions of treasury data in ordinary language and create dashboards. Agents for counterparty analysis and bank-account opening and closing are still being developed. Live use and future plans should not be presented as the same thing.

The pattern is consistent. Adoption starts with a bounded task, an identifiable user and an output that can be checked. It does not start with an unsupervised system moving cash or taking financial risk.

Governance is part of the product

The Bank of England and Financial Conduct Authority found that 46% of surveyed firms using or planning AI reported only partial understanding of the technologies they use, with externally supplied models contributing to the gap. The same survey found that named accountability, AI policies and data controls were among the most common safeguards.

The US National Institute of Standards and Technology groups AI risk work into four activities: establish governance, understand the use, measure performance and manage the risks. That is a useful treasury discipline: name the accountable person, define the use and possible harm, test performance, monitor model drift—changes in performance over time—and control changes to the system. Data, workflow, supplier dependency and human review determine whether the use is trustworthy.

A sensible 2026 position

Treasury teams should neither dismiss AI nor buy it as a generic capability. Select two or three clearly defined uses, measure the current result, run the new approach with real data and retain human review. Expand only where the evidence shows better accuracy, faster processing or stronger control.

Agentic AI does not remove controls. It makes them more important. Record what the system may observe, recommend and perform; which actions are prohibited; who owns the outcome; and how the organisation can stop, investigate and change the service.

The question is not whether the product contains AI. It is whether a defined treasury outcome improves—and whether the organisation can explain, control and rely on the result.