How AI Is Actually Used in the Office of the CFO

The office of the CFO runs on two things that do not bend: confidential data, and numbers you can defend to a board. That makes it one of the harder places in a company to adopt AI, and one of the places under the most pressure to do it anyway. Investors ask about it. Boards ask about it. And “we’re experimenting” has stopped being a satisfying answer.

Over four sessions of Transforming the Office of the CFO with AI-Powered Workflows, Maxio CFO Jon Cochrane pulls in the vantage points that actually shape this call: a product leader building the tools, finance operators using them on live books, an outsourced-finance partner who evaluates them across a portfolio, and a private equity investor who judges whether any of it moved the business.

Here is what we learned from watching them put AI to work on real finance workflows, and from their hard-won judgment about where it belongs, so you can skip the experiments that go nowhere and start where the value is.

Start where a wrong answer is cheap

Pointing AI at the general ledger on day one is the first mistake. Start with work that eats time but carries little risk, where a bad output costs a few minutes instead of a restatement.

Three good entry points:

• Drafting and summarizing. Draft the Slack message or the email. Summarize the thread you never read or the meeting you missed. Create an automation that runs every morning to scan yesterday’s email and return a prioritized list of what needs a reply.
• Document review. Put a counterparty’s NDA and your own template side by side and ask for a redline against your standard terms. One finance leader did this and found the other side’s terms were more generous than his own, with a single cap worth adding, and signed the same afternoon. The same move works on an audit draft: give it the current draft as well as last year’s, and ask it to tie out the trial balance, check the footnotes, and list what looks wrong. The output is a punch list you send back to the auditors.
• Ad hoc analysis. The “what does this look like if” questions that normally wait in a queue because nobody has time to build the model.

The frame that keeps this safe is to treat AI like a new hire. You would not hand a first-week analyst the board deck and say ship it. You give contained work, check it, and hand over more as it earns trust. Skip the checking and you get what any unmanaged new hire produces. That mindset carries through everything below.

Learn what AI should calculate, and what it should not

The next question is whether you can trust it near the numbers. The answer depends entirely on which kind of work you mean.

An LLM is probabilistic. Ask it the same thing a hundred times and the answer can drift, which is fine for interpretation and fatal for arithmetic. The trap is treating a chat window like a database: load it with data, ask it to total something, and the number can change on each pass.

Deterministic work needs the same answer every time. Revenue schedules, commission calculations, reconciliations. Do not ask the model to be the calculator. Ask it to build the calculator, the SQL query or the Excel formula or a small app, and let a real system run the math. Tell it to keep the formulas simple and traceable, for example: use SUMIF and XLOOKUP, no nested index-match, so I can click any cell and read the logic. That traceability is what lets you stand behind the result.

Judgment work is where the probabilistic nature helps. Framing a board narrative, pressure-testing your assumptions, surfacing the insight a sharp analyst would raise unprompted. A useful prompt here is to make it argue with you: review this the way a skeptical board member would, and list the three hardest questions they’d ask.

Underneath both sits a rule that matters even more once you connect systems: your data lives in your system, not in the chat. Point AI at a source you already trust, teach it how to read what it pulls, and keep the numbers where they can be audited.

Connect it to your real systems

Chatting with a model is the shallow end. What changes its usefulness is connecting it to where your data lives: your ERP, your billing platform, your spend tool. Most teams do this through MCP, a standard that lets a tool like Claude read from another system, and act inside it when you allow that.

That reach is why setup matters. Have these in place before you turn on anything that can write:

• Per-user authentication. Each person connects with their own credentials, so AI can only reach what that person already has permission to reach. Enabling a connector for the org does not hand everyone a master key.
• Graduated permissions. Leave read-only tools open. Set anything that writes or deletes to require approval, or turn it off, until you have a reason to loosen it. Keeping an ERP connector on “needs approval” for months is normal and strongly recommended.
• A clean audit trail. Actions taken through the connector should log as the user, with timestamps, and be visible in the system after. If you would log in to check a change you delegated to a person, you want to do the same here.
• Risk previews. The better connectors show the consequence before it commits. A subscription change through the Maxio MCP returns a preview inside Claude, with what changes and the amount due, and waits for confirmation. Most tools are not there yet.

This is not a new security discipline. You had IT and data policies before AI. What changed is how easy it is to copy an output into another tool and keep going, so the job is knowing where your data is allowed to travel. Set the policy with whoever owns your systems, and keep it loose enough that the team can still learn.

Automate the work you repeat

With a trusted connection, the payoff is turning repeatable work into something that runs itself. Two building blocks:

Skills are saved instructions for a task you do over and over, so you stop re-explaining it every month. Build one by finishing a task you like the result of and asking the model to save how it did it, then refine the saved version over time. One team taught a skill to read their ARR report using their billing platform’s own support docs, because the model had to learn what the fields meant before it could interpret them.

Skill stacking chains those into a workflow. A monthly close routine might call one skill to reconcile cash, another to email the AR aging to account management, another to interpret the ARR walk, each trusted on its own.

Then scheduled automation gets interesting. One finance team connected their spend tool and built a task that checks every unsynced transaction against a coding checklist, catches miscodes before they reach the ERP, and autocorrects the recurring ones with a link back to each transaction to verify. It runs hourly during business hours. Finding one miscode across thousands of monthly transactions was always a needle in a haystack, and there was never a good way to automate the search.

Notice the control on that same team: AI proposes journal entries, a human approves them, nothing posts on its own. Copy that pattern as you scale. Widen scope one narrow permission at a time, keep volume low while confidence builds, and keep a person in the loop anywhere the work touches the books.

Turn a faster close into faster decisions

Automation buys time, and what you do with it decides whether finance steps forward or stalls. Cutting the close from twenty days to ten, or to five, is real, and AI helps get there. But the close is not the finish line.

The shift that matters is speed to decision. Businesses now make meaningful calls weekly and daily. Show up only at month-end and decisions get made without you, sometimes off a number someone pulled straight from the CRM. Teams that stay in the room build weekly dashboards carrying the critical numbers, forecasted bookings and churn included, so leadership and the board stay aligned between meetings.

Forecasting weekly also makes the forecast better. The most accurate forecasters set assumptions and adjust them continuously as reality lands; forecast only monthly or quarterly and you are less accurate because you are out of practice. A weekly cadence only works because the automation underneath handles the assembly.

Do the analysis that used to be impossible

This is what the whole path builds toward. The most valuable use of AI in finance is not old work done faster. It is work that was off the table because it would have taken months by hand.

• Enrichment. The industry labels you buy from data providers are usually too coarse. “Tech company” covers thousands of accounts and tells you nothing. AI can classify what companies actually do, marketing tech versus cyber versus defense versus ed tech, at a scale no person could hand-tag, and that detail changes where you invest.
• Cohorting. One investor needed thousands of a portfolio company’s customers tagged and cohorted by segment and industry to make an allocation call. That used to mean people visiting websites and tagging accounts over weeks. With AI it took a few days to tag, check, and act, and the decision landed on time instead of late.

Both depend on the rule from earlier: a data source you trust. Get comfortable with the underlying population first, the total, the count, the definitions, then let AI slice and enrich it. The trustworthy base is what makes the ambitious analysis safe, which is why it sits at the end of the path rather than the start.

Telling real progress from theater

As you move along, two habits separate teams getting value from teams checking a box.

They test tools with their own examples, live. You already do this with an ERP. Refuse the canned demo, hand over your real and messy cases, and watch what happens with the ones outside the normal workflow, because that is where tools break. One firm was pitched a pure-AI revenue recognition tool that promised to build schedules straight from contracts. They sent three contracts. All three came back wrong. The category is not fake; the point is that “show me live, with my data” is the only test that counts.

They tie the work to enterprise value. “How are you using AI” is usually a stand-in for a harder question about whether any of it is creating value. Ten days off the close is an accomplishment, but the question is what you did with the ten days. A “100% adoption” number or a token leaderboard does not move enterprise value by itself. The metrics that matter connect to growth, or to faster and better decisions. Draw a straight line from a company priority to what AI is doing to advance it and it is real. If you cannot draw that line, it is theater, however good the demo looked.

Where to start

If you are early, pick one low-risk, repetitive task and do it this week. Draft with it, summarize with it, review a document with it. Learn where it is strong and where it needs checking, the way you would with a new hire.

From there the path is the one above. Move from drafting to connected data once you trust the outputs. Keep deterministic work in systems you can stand behind, and aim AI at the judgment and the analysis. Automate what you repeat. Spend the reclaimed time in the decisions, not just the close. And measure the work against what your business values, not how many prompts your team ran.

The tools improve fast and the product names keep changing. The finance fundamentals do not. Hold your source of truth, stand behind your numbers, and spend the time AI gives back on the work that was always too slow to attempt.

All four episodes of Transforming the Office of the CFO with AI-Powered Workflows are available on demand.

• Session 1: Connect AI Tools with Confidence with Chris Weber, CPO at Maxio
• Session 2: Spot Anomalies Before They Happen with Nick Blanchard, VP Finance & Accounting at Curve Dental, and Hillary Frost, Director of Partner Operations at Maxio
• Session 3: Use Cases That Actually Work with Alex Diaz-Asper, Partner at Cherry Bekaert
• Session 4: What Investors Really Want to See in Your AI Strategy with Casey Rihn, Senior Director of Operations at K1 Investment Management