From copilot to agent-to-agent: The evolution of AI in treasury
Published: August 18, 2026
Can or should treasury teams progress from using AI as a copilot to deploying agents with controlled autonomy? Zafeer Ahmed, Chief Revenue Officer at global financial solutions firm Ebury, explores what each stage would look like in practice, and how the treasurer’s role changes should AI assume greater operational responsibility.
I have spent nearly 15 years working with CFOs and treasurers, and almost every treasury I have known runs on a calendar – month-end forecasts, daily cash positions, quarterly hedging reviews, and annual budgets. It took me years to see these cadences for what they are. None of them are optimal. They are bandwidth constraints dressed up as policy – the maximum frequency at which a human team can realistically gather the data, run the numbers, and act.
That is why AI in treasury is approaching a fundamental turning point. As it moves from tools that support human decisions to agents capable of making and executing decisions themselves, the constraints that shaped every treasury process quietly disappears. Work that was batched into monthly and quarterly cycles because people were busy can happen when the business changes, not when the calendar says so.
Most treasury teams remain in the ‘copilot’ era, using AI to analyse information, support existing workflows and make recommendations while humans retain control. The next stages will see autopilot systems act within human-defined policies, followed by an agent-to-agent environment in which corporate treasury agents interact directly with those operated by banks and financial providers. As AI takes on greater operational responsibility, the role of treasury professionals will evolve alongside it. Understanding that changing relationship, and the data, technology, and governance foundations needed to support it, is the starting point for progressing towards greater autonomy.
Copilot: AI supports, but humans decide
Most organisations are now moving into the copilot era. Treasurers and CFOs use AI within their existing workflows to improve efficiency, analyse data, and produce recommendations, including AI capabilities embedded within spreadsheets or treasury tools. The defining feature is that the human remains firmly in control. AI can draft, analyse, and recommend, but the treasurer still decides and executes.
Agentic AI introduces something different. Traditional automation follows predetermined ‘if this, then that’ rules. Generative AI gave software the ability to reason, whereas agentic AI adds the ability to act. Systems can therefore begin to assume responsibility not simply for producing an output, but for achieving an outcome within a defined mandate.
Autopilot: agents act within human policies
The next stage is the autopilot era. Here, agents are trusted to make and execute certain decisions within policies and limits established by people.
Cash visibility is a good example. Businesses increasingly operate through multiple entities, currencies, and banking relationships. Instead of relying on a snapshot taken at 9am each day, an agent could monitor cash positions continuously. If a balance crosses an agreed threshold, it could sweep funds into another account where additional liquidity is needed.
Forecasting is another clear application. Building forecasts from AP/AR and ERP data often remains manual. Agentic AI could update these forecasts in real-time, enabling the treasurer to validate and challenge the results instead of constructing them from scratch.
Currency risk management is where the shift is most striking, because it is where the calendar has done the most damage. Hedging programmes have traditionally been reviewed quarterly, with hedge ratios set by tenor: more cover for near dates, less for far ones. Tenor was only ever a proxy for how certain each cash flow was. An agent that monitors exposure continuously can track certainty directly, adjusting cover the moment a forecast becomes an order or a contract is signed, within limits the treasurer has set. Hedging becomes event-driven rather than calendar-driven, protecting margins instead of anniversaries.
However, even when an aircraft is operating on autopilot, the pilot remains in the cockpit. Treasury professionals remain responsible for setting policies, governing systems, managing exceptions, and intervening during periods of volatility. The system manages routine activity, but the pilot intervenes during turbulence and at the moments when human judgment is most valuable.
Agent-to-agent: treasury's longer-term destination
The final step is an agent-to-agent environment, and it will arrive sooner than many expect. Corporate treasury AI agents will interact directly with AI agents operated by banks, payment companies, and other financial providers.
A corporate agent might identify an emerging currency exposure, assess the business' requirements and communicate with provider-side agents before acting within an approved policy. The human defines the mandate, risk appetite, and limits, while agents manage more of the operational process.
This represents a significant change in how businesses interact with their financial providers, but it will be a multi-year progression. This is the world of money, payments, and cash: trust will be critical, and it will have to be engineered, through verified identity between agents, shared standards, and mutual auditability, not assumed.
Building the foundations for progression
Before progressing between these stages, it is critical that treasury teams have the right data, technology, and governance foundations in place.
Preparation starts with connected systems of record, including accounting systems, TMSs, and other relevant platforms. However, simply pointing an LLM at those systems will not produce reliable outcomes. It can be expensive, difficult to verify, and vulnerable to hallucinations.
Businesses also need semantic consistency: common definitions for metrics such as balances, volumes, and transaction statuses across different systems. Without that shared meaning, an agent may reach a technically logical conclusion using inconsistent information.
Predictability must then be tested. Organisations should establish ‘golden datasets’ showing how they expect an agent to behave and test the system against them whenever changes are made. They should also back-test agents against historical decisions to assess how they would have performed in practice.
Finally, decisions must be explainable. There should be a clear line between an agent's action and the information behind it, tracing the decision through the semantic layer to the original systems of record. Clear accountability is equally essential: organisations must know who is responsible for setting the agent's authority, monitoring its performance, and responding when something goes wrong.
Agent autonomy must be earned
Organisations should not begin by allowing agents to make critical decisions or override compliance processes. Their responsibilities should expand incrementally. An agent might initially operate in observation mode, before progressing to recommendations and then executing low-risk actions within narrow limits. Its remit can grow as the business builds evidence that it behaves predictably and reliably.
Moving too quickly, before the necessary data, testing, and governance are in place, risks undermining confidence in the technology. The organisations that benefit most will be those that allow agentic autonomy to be earned, progressing gradually and deliberately from copilot to autopilot and, eventually, agent-to-agent. Governance, viewed properly, is not the brake on autonomy. It is what makes autonomy possible.
Leading the transition
Treasury teams may be adopting AI at different speeds, but they are heading towards the same destination. As AI evolves from a tool that assists people into agents capable of reasoning, acting, and interacting with one another, it will reshape how treasury decisions are made and executed. The six questions treasurers have always asked, about exposure, appetite, strategy, execution, monitoring, and review, all survive. What disappears is the calendar to which they were chained.
Those that begin building the right foundations now, codifying policies, connecting data, and designing governance early, will be best placed to lead that transition rather than having to catch up with it.