AI-Native Finance Is About Redesigning Workflows and Decisions

Sarah Friar, CFO of OpenAI, published a piece recently on what building an AI-native finance function taught her, and the clearest point in it is about the unit of redesign, more than the tools themselves. Most companies bolt AI onto an existing step: a chatbot that drafts the variance commentary, or an agent that reads the ledger faster. Friar’s team took the whole path from source data to decision, the close, the forecast, the capital allocation call, and redesigned the workflow around the decision itself rather than the artifacts that used to feed it. What disappears is the week-long scramble to piece the business back together before anyone can explain a number.

That distinction is worth sitting with before anything else, because many (if not most) AI programs start with reporting. They automate the report and stop there, leaving the decision exactly where it was, sitting downstream of five disconnected systems and someone’s judgment call about which spreadsheet is current.

Finance is the one function that touches every other function. Procurement, sales, supply chain, HR headcount, capital projects, all of it eventually runs through a finance workflow: a budget, a forecast, a variance, an approval. Redesigning finance means you are redefining workflows across departments, rather than fixing one department in isolation.

That makes finance the highest-leverage place to run this first, and the reason has more to do with what gets built there than with finance’s problems being unique. The data foundation and workflow discipline it forces become what every function that comes after it inherits.

Microsoft's own finance team makes the same case with numbers attached. Its Global Treasury and Financial Services group ran more than a thousand collectors across systems that stayed siloed from each other, invoices in one place, customer conversations in another, leaving only a fragmented view of what any customer actually owed. The team fixed that first, consolidating everything into a single SAP and Dynamics 365 environment before layering a Copilot-assisted agent on top. Payment-matching accuracy improved from 40% to 90%, 98% of payments are now applied within 48 hours, and customer inquiry resolution became 2.5 times faster. As Kathy Brustad, the director who led the effort, put it, the real lesson was understanding the existing process well enough to redesign it, rather than dropping an agent into whatever already existed.

Getting there is a sequence, and it is tempting to skip straight to building after a hackathon shows what is possible in a day. Taking that tool to production takes more than governed data. It takes thinking through reliability, security and a handful of other aspects a demo never has to survive, because finance is the backbone every other department leans on.

The sequence that holds up looks something like this.

Assess. Before any tool gets chosen, map the actual decision workflows. Where does the close bottleneck. Where does the forecast diverge from what sales already knows. Which system holds the source of truth for a given number, and which ones are already three versions removed from it. This step alone usually surfaces more waste than any AI pilot will fix in its first year.

Strategy. Decide which decision cycle to redesign first, and hold the line on cross-functional impact rather than departmental convenience. Set the target state as a continuously reconciled finance function where every variance is traceable back to source. Measure success by the value created from better decisions and faster execution, rather than by seats deployed, models built or tokens consumed. This is where CFO and CTO co-ownership has to begin, because neither function can deliver the outcome alone.

Design. This is the unglamorous part and the one most programs underinvest in. It means building or extending a governed data foundation, your own data and AI foundation in the cloud, a semantic layer that agrees on what a “customer” or a “cost center” means across every system, so that frontier models can be pointed at that data directly instead of three teams manually reconciling exports first.

Implement. Start with one workflow instead of a suite. Ground it in approved sources, with a human owning the final answer. Ship it to the people who live in that workflow every day, and correlate what they report with the telemetry coming off the underlying systems, so problems get diagnosed faster than either signal alone would allow.

Scale. Take what worked in the first workflow, the data model, the guardrails, the access pattern, and point it at the next one. This is where the foundation built for finance starts paying for itself outside finance, because the next function inherits it instead of building its own version from scratch.

One ingredient determines whether all five steps succeed: confidence. Most transformation plans treat adoption as a communication exercise that happens after the technology is built. In practice, confidence is built much earlier. People trust a system through repeated exposure to what it can do, where it falls short, and how its outputs connect back to the underlying business reality. That trust cannot be mandated. It has to be earned through participation in the redesign itself.

My team saw this firsthand while helping a customer implement AI for fraud detection. The work did not begin with model selection. It began with understanding the transactions, the behavioural patterns and the risk indicators that investigators were already using. Much of the value came from the data exploration phase, which established a strong data foundation and identified where AI could genuinely improve decision-making. The analytics and models came later. Even then, the implementation was not complete when the reports started being generated. Investigation officers had to learn how to interpret the outputs, understand the confidence levels, challenge the recommendations and connect them back to the underlying transactions. Confidence grew because they were involved in the process and could see how the system reached its conclusions.

The finance teams that make the transition fastest are usually the ones closest to the work. They help define the workflows, test the assumptions, challenge the outputs and shape the operating model that eventually goes into production. By the time the organisation begins scaling, the people responsible for the decisions already understand and trust the system because they helped build it. Adoption becomes a consequence of ownership rather than a separate phase of the transformation.

An important broader lesson from the article as well: AI-native finance is not about accelerating existing processes or generating reports faster. It is about redesigning workflows around decisions and creating a system where data, intelligence and action remain continuously connected. Because finance sits at the intersection of every major business function, the foundations built there extend far beyond finance itself.

The organisations that gain the most from AI will not be the ones that automate the most tasks. They will be the ones that redesign the workflows that matter most.