Finance teams are seeking to integrate AI into their workflows. Doing so effectively requires deliberate workflow design that defines where human judgment is needed, where deterministic rules should govern outcomes, where probabilistic reasoning can provide insight and how controls remain embedded throughout the process.

When these elements work together, each component can play to its strengths. AI can process large volumes of data, identify patterns and generate potential explanations. Rules and controls can provide consistency, oversight and compliance. Finance professionals contribute business context, domain expertise and the judgment needed to challenge assumptions and assess whether an output aligns with organizational reality.

During the 2026 AFP FP&A Series, Building Financial Intelligence: Deliberate Practice in an AI Era, sponsored by Ramp, finance professionals discussed how to build financial intelligence (the ability to combine financial knowledge, business understanding and judgment to make better decisions) into workflows. The four key insights and framework below summarize the speakers' collective guidance.

Key takeaways

• Use AI to deepen your understanding of the business, not simply accelerate the work you already do.
• The right way to validate AI output depends on how you are using AI — as automation, an accelerator or a thought partner.
• You can build a perfect model, but judgment still has to come from the people who know your business, competitive position and long-term strategy — the details no spreadsheet captures.
• AI agents become much more useful to finance when deterministic controls, probabilistic reasoning and human review are deliberately built into the workflow.

AI in Finance Certificate

The No Code AI for Finance Certificate program is built for finance professionals who want real, usable AI skills today. Learn to implement AI responsibly, transparently and in a way that strengthens decision-making across your organization.

Learn More

AI governance framework for finance teams

While the Association for Financial Professionals FP&A Series speakers approached AI from different angles, they collectively described a common operating model for finance teams: determine the role AI is playing, establish the appropriate controls and define where human accountability remains essential. The framework below summarizes how AI can be applied across common finance workflows.

Insight #1: AI delivers better outcomes when judgment is built into the process.

Many finance professionals worry that AI will reduce the need for human analysis. According to Carolina Lago, Managing Director of Tactic Financial, what's actually happening is different: The finance professionals pulling ahead aren't using AI simply to move faster. They're using it to understand the business more deeply, test assumptions they didn't have time to explore before and develop expertise in the process.

That works because the machine and the human bring different strengths. AI can hold more context at once, analyze the same data from multiple angles and surface patterns a summary number might hide. What it cannot do is decide which question matters. Framing the problem and understanding the context a system can't capture remain human responsibilities.

Lago explained that effective AI workflows contain three layers: deterministic steps, probabilistic tasks and human judgment:

• Deterministic steps, such as data pulls, reconciliations and calculations, can be standardized and automated.
• Probabilistic tasks, such as pattern-finding and alternative interpretations, are better suited to AI.
• Human judgment serves as the checkpoint that turns analysis into a decision.

To illustrate her point, Lago described a client whose team was using AI in its cash reconciliation process. AI surfaced potential ways to improve cash, including techniques the team had not previously used. Rather than accepting the suggestions at face value, the team evaluated them, learned from them and incorporated those lessons back into the process.

That learning loop is where the long-term value emerges. Done well, the same workflow improves everyone: The analyst develops stronger data judgment, the manager improves prioritization and the CFO sharpens strategic thinking.

As Lago observed, "The business acumen of the team was increasing because they were learning new techniques. And as they were doing this exchange, the process was improving as well."

In that sense, AI is not replacing expertise. It is helping finance professionals, in Lago's words, "become more senior faster."

Insight #2: How you verify AI output depends on what you asked the AI to do.

If AI can make mistakes, how much verification is enough? Anna Tiomina, Founder of Blend2Balance, argued that the answer depends on what you asked AI to do in the first place.

She divided AI work in FP&A into three categories: automation, accelerator and thought partner. Because each one can fail differently, they require different controls.

1. Automation handles routine work, such as loading data, flagging variances and processing invoices. It runs in the background until something changes: a new vendor format or account, an exception outside the rules. That calls for ownership, exception reporting, change controls and periodic spot checks to catch drift.
2. Accelerator is when you and AI work together. Activities include drafting analyses, building a model and creating commentary. You're actively involved and have established checkpoints. The risk is that AI might misread your sources, fabricate figures or write something that contradicts the data. To mitigate that risk, you need to verify footing, check causality, validate data integrity and make sure you can defend every conclusion.
3. Thought Partner is when you're stress-testing assumptions or preparing for a board meeting. AI doesn't know your politics, stakeholders or context. It will likely simply agree with you. Tiomina suggested asking two questions: Did you give AI the context that determines the answer? And would you reach the same conclusion without it?

You can also build checks into the workflow itself, such as having AI cross-check calculations, stress-test assumptions and even argue against a conclusion before a human performs the final review. But no control transfers accountability to the tool.

Blaming AI for a mistake doesn’t change who is responsible for what ultimately reaches management, the board or another stakeholder. “The ownership of the outputs is yours,” said Tiomina. 

Insight #3: Context dramatically improves AI output, but business decisions require human judgment about factors models can't see.

What happens when you hand AI a real FP&A problem instead of a hypothetical scenario? Paul Barnhurst, FPAC, Founder of The FP&A Guy, and Mario Vasquez, FPAC, Senior Director of Finance at The E.W. Scripps Company, walked attendees through a buy-versus-lease decision on a broadcast tower.

Barnhurst started with a basic prompt. The math was solid, but the results were inconsistent: The AI produced different layouts, assumptions and interpretations of the instructions every time he ran it. "If you're doing prompt only, you're going to end up with problems," he said. Even a detailed best-practice prompt only marginally improved it.

The shift came when Barnhurst stopped relying on prompts alone and started giving the AI reference materials. Four components did the heavy lifting:

• Personalization rules captured his modeling standards, including formatting conventions and formula discipline.
• Workbook rules established a consistent structure for how models should be organized.
• A sample file showed the AI what "good" looked like by providing a completed model as a reference.
• Plan mode required the AI to validate assumptions and present its approach for review before building anything.

In other words, guardrails, examples and structure. Crucially, those components lived outside the chat window, making them auditable, shareable and repeatable. Barnhurst described this as a shift from prompt engineering to context engineering.

The AI-produced financial model said the decision was close: Buying or leasing would cost roughly the same over 10 years, about a $100,000 difference. E.W. Scripps Company bought the tower anyway because the spreadsheet was only half the story.

The purchase made sense because of everything the model wasn't built to see: The tower offered better household reach, which meant higher Nielsen ratings and more revenue. The company needed negotiating leverage with its largest tower owner, and owning one tower improved that position. Year 11 tilted decisively toward buying. And then there was the farmland surrounding the tower — the company leased it to a farmer and split the revenue.

A mathematically sound financial model and a sound business decision aren't always the same thing. You can build a perfect model, but judgment still has to come from the people who know your business, competitive position and long-term strategy — the details no spreadsheet captures. Understanding that distinction may be one of the most important skills finance professionals need as AI gets better at modeling.

Insight #4: Agent harnesses address the AI control problem by dividing where it runs into three lanes: deterministic, probabilistic and human.

Finance runs on certainty, but LLMs are probabilistic — run the same prompt twice and you might get different results. That mismatch, said Glenn Hopper, Founder of RoboCFO, is why AI agents initially struggled in finance.

“Up until this year, that wasn’t stubbornness; it was a control problem,” he said. “In finance, we need a number that is repeatable and explainable. A number that changes on every run can't face a board or a lender. And if there's no audit trail, there's no adoption.”

An agent harness — a structure around the AI model that can provide orchestration, access to tools and data, memory, guardrails, human review and an audit trail — enforces the decisions people make about which kind of intelligence belongs where.

Think of it as three lanes: deterministic, probabilistic and human. “The edge isn't just using AI; it's designing where it runs. If we're all using AI, the question is how can we best apply it where we can trust it? You have to know what to keep deterministic, what to leave probabilistic and where humans sit in that whole mix. And that judgment is the differentiator.”

Hopper illustrated his point with a 13-week cash flow forecast. The architecture is simple: deterministic for what's known, probabilistic for what's uncertain and human for what requires judgment:

• Payroll, debt service and rent are known. Let deterministic controls handle them; there's no upside to guessing.
• Other parts are inherently uncertain: when customers actually pay. Some pay early, some late, some not at all. That's where probabilistic methods excel; an AI agent can run a statistical model on historical payment patterns to build a range.
• The judgment calls — the ones that carry liability and require sign-off — stay with the humans. 

When evaluating potential AI use cases in finance, start by considering what happens if AI gets it wrong. High-stakes, contractual and board-facing work needs deterministic controls and human approval, whereas more uncertain and directional work is where AI’s ability to reason across possibilities adds value.

Financial intelligence remains the differentiator

Across all four sessions at the AFP FP&A Series, a common theme emerged: AI may accelerate analysis, but financial intelligence remains the differentiator. As AI takes on more of the data gathering, modeling and scenario generation, finance professionals create value by framing the right questions, providing business context, evaluating tradeoffs and determining what happens next. In the AI era, competitive advantage comes from combining technology with sound judgment.