The adoption of artificial intelligence in corporate treasury is moving fast, but there is a clear distinction between basic tool usage and full operational transformation. During the companion webinar to the AFP Treasury in Practice Guide: AI-Ready Treasury, underwritten by Wells Fargo, treasury experts outlined a five-stage framework for AI adoption.

The five stages should not be treated as a maturity ladder that every treasury function must climb uniformly. A more useful approach is to assess maturity by process. For example, a treasury team could be redesigning the account reconciliation process to integrate AI (stage three), while still experimenting with use cases to support risk analysis (stage one).

Taking a process-level view of AI also changes how teams should prioritize investment: Rather than asking where AI can be added, treasury leaders can ask which workflows should be redesigned, which decisions can be delegated within established controls and where human judgment remains essential.

The 5 Stages of AI Adoption in Treasury

1. Learn to use AI tools through experimentation.
2. Make AI part of your daily work routine.
3. Redesign existing processes with AI in mind.
4. Let AI execute defined tasks within the guardrails set by your company.
5. Scale an operating model where AI is baked into every process.

Stage 1: Learn to use AI tools through experimentation.

In stage one, treasury professionals learn what AI tools can do and how they can be applied to existing responsibilities.

This stage does not require formal training. In fact, Kathy Brustad, Director, Treasury, Microsoft, said a lot of the learning at Microsoft began the same way as anyone in any industry: with experimentation.

While early adoption requires basic experimentation, most treasury teams are ready to move past this phase Brustad’s team helped speed experimentation by sharing prompting guidance and examples, establishing AI champions and generating opportunities for the early adopters to share what they were building with their colleagues. Taken together, this peer-level assistance helped spread knowledge throughout the company at a much faster rate.

Having the support of leadership is necessary to this process too, Brustad advised. Your team might be interested in experimenting, but if they don’t have the time to do so and meet their daily responsibilities, it becomes a moot point. Learning needs to become part of the work rather than something employees are expected to squeeze in around it.

“The biggest barrier that we’re hearing is ‘I want to try it, learn it, but my day is filled with things I must do operationally as part of my job and I don't have capacity for that,’” she said. “So, we pushed our leadership very hard to protect capacity for people to do AI learning.”

Stage 2: Make AI part of your daily work routine.

Teams enter the second stage when AI stops being a novelty to test and becomes an integral part of everyday operations.

Brustad described how she used to begin her workday by opening Outlook and Teams, and now she starts her workday by reviewing the status reports of her AI agents.

“I have so many personal agents working for me,” she said. “What I first need to know is whether all those agents executed correctly to do all the things that I have them do while I'm asleep.”

Her drafting habits have changed as well. “If there's anything I need to draft, I go to my AI window first,” said Brustad. The outputs are just as she described: a draft. Brustad reads and edits all AI output before sending it.

Describing the same principle in more explicit terms, Bojan Belejkovski, Vice President, Finance & Treasurer, Voltava, put it this way: “I am the judgment layer.”

Stage 3: Redesign existing processes with AI in mind.

In stage three, with AI now a daily habit and tasks running faster, practitioners step back to redesign the processes.

In the case of the Microsoft Treasury team, they first looked at areas where large amounts of repetitive manual work were being performed, such as financial analysis and account reconciliation.

“What we always look at is the low-hanging fruit, the areas where you have highly repetitive, manual activities, and that's the space where we are looking to automate or integrate AI. That’s where we started: identify the pain points and redesign the process to remove those frictions in the process with AI,” said Brustad.

Simply adding AI into an inefficient process may improve individual tasks, but it doesn’t address the underlying workflow. At this stage in AI adoption, you need to ask yourself whether you can remove steps, reassign responsibilities or structure processes differently with the assistance of AI.

And you don’t need to stay within the parameters of the work being performed today. “We look at areas that we haven't even thought about — places where AI can help us or where we know there are things that need to be done, but we never get to them because we're busy with daily grinds,” said Brustad.

That broader thinking has led her team toward proactive insights and risk mitigation, areas where AI speeds up their response to global developments and potential impacts.

Stage 4: Let AI execute defined tasks within the guardrails set by your company.

In stage four, teams delegate specific tasks to AI within human-defined parameters. Humans continue to oversee the outcomes, applying judgment to verify accuracy and correctness.

Brustad described three increasingly advanced relationships between people and AI:

1. AI assistant: A human working alongside AI.
2. AI execution: Human-led work that AI executes.
3. AI workflow operation: Human-led processes in which AI operates an entire workflow. This is the entry point for agentic AI that plans, reasons and acts to achieve a goal with little human assistance.

Subotic illustrated this progression using a payments scenario:

• Machine learning recognizes a payment that doesn’t match an expected pattern and flags it.
• Generative AI (GenAI) goes a step further by examining relevant information to determine why the anomaly occurred.
• Agentic AI takes automated action based on the analysis.

“We don't want something to just flag and stop with explanation,” said Belejkovski. “We want something that decides within guardrails, within governance structure and within limits that we've set. It is going to notice the unexpected, explain it and then act on judgment that you built based on your fixed rules for the company.”

The treasury use cases for such a system are already there. Brustad pointed to know-your-customer (KYC) processes, which require companies and banking partners to exchange documents and information.

“We finally got to a point where I believe technology will allow us to do agentic interchanges, meaning agent to agent, probably an agent working with another agent to handle this process,” she said. However, she also noted that this work is still in the experimental stage.

Stage 5: Scale an operating model where AI is baked into every process.

Stage five represents an AI-native operating model where AI underpins every process from the start. Sustaining this model requires balancing when experimentation is called for and when stronger institutional controls are necessary.

Brustad described Microsoft's approach as “disciplined core and flexible edge.” Core finance data and processes require IT involvement and controls, but with newer tools, treasury analysts and managers can build some dashboards, agents and other capabilities themselves.

“There is a lot more you can do at the flexible edge,” she said. “But then that goes back to that governance question, that risk management question of what should be built at the flexible edge and what should be kept at the core.”

This is the point at which governance becomes increasingly important as the organization has moved from the realm of personal productivity into shared processes.

“We [Microsoft] have guardrails around every agent that's built,” said Brustad. “If it's for more than your personal use, we ask to go through all of these responsible AI reviews and security reviews to make sure those agents are built properly.”

Belejkovski’s organization formed an AI steering committee in which the AI SMEs, Chief Legal Officer, Chief Technology Officer and governance leads meet to monitor ongoing technological changes.

“We meet and discuss: What is changing? What do we need to make sure that we capture?” Belejkovski said. The group acts as a formal reviewer when new AI capabilities roll out, ensuring potential risks are evaluated before workflows shift. “If someone is asking a question, it’s probably triggered by a change in a process or a function, and it needs to be reviewed.”

Establishing an AI-native treasury function requires governance mechanisms that determine what AI can do in your company, what data it can access, what humans remain responsible for and how emerging risks are addressed.

Example: What Would the Five Stages of AI Adoption Look Like with FX?

Stage 1: Ask the agent.

A treasury analyst asks an AI agent, “Summarize our EUR exposure and explain what changed since yesterday.” The agent responds, but the analyst still gathers the underlying data, decides what to ask and determines what to do with the answer.

Stage 2: Make the agent part of the routine.

Every morning, the analyst automatically receives an agent-generated summary showing current FX exposures, significant changes and items requiring attention. Instead of needing to remember to use AI, the agent is now part of the daily FX workflow.

Stage 3: Give the agent a workflow.

The agent pulls exposure data from treasury systems and business units, compares it with existing hedges and flags positions outside established parameters. It may also gather the information needed to investigate an exception. The practitioner is no longer manually assembling the analysis before consulting AI; the process itself has been redesigned around the agent.

Stage 4: Let the agent execute within guardrails.

The agent identifies an unhedged exposure, pulls the relevant data, evaluates it against treasury’s hedge policy and established limits, and recommends a specific action — or executes approved steps within predefined parameters. The treasury professional reviews exceptions and decisions outside those guardrails.

Stage 5: Let agents run the FX workflow.

Multiple agents work together continuously: One agent gathers exposures, another monitors market conditions, another checks positions against policy and limits, and an orchestration agent coordinates the workflow. Routine activity proceeds within established controls, while treasury professionals oversee the system, set the strategy, revise guardrails and intervene when judgment is required.

Curiosity about AI is a critical treasury competency today

Subotic noted that hiring priorities for treasury professionals have fundamentally shifted. Where technical mastery of treasury management systems once drove hiring decisions, candidate curiosity around AI now takes precedence.

“I would say the number one thing I look for is someone who says, ‘How can I use AI?’” Belejkovski explained, adding that watching team members take the initiative to build their own use cases is “the fastest way to close the gap” as AI capabilities accelerate.

A team's level of curiosity often dictates how quickly it advances across the five stages of AI adoption, going from basic tool exploration in stage one to complete workflow redesign in stage five.