Trust, but verify: what Carl Seidman showed us about generative AI for FP&A

By Natasha Sabelli, CPA: Not Your Typical CPA | Big 4 Audit & Controller Experience | Professional Yapper About Finance, Tax, Tech & Everything in Between | Finance Evangelist @Velixo & Founder @NatSab

There’s one day every month-end that finance people dread more than any other.

It’s not close day.

It’s the day after, when you have to sit across from your Director of Finance, CFO or board and explain every number that moved.

I lived that day for years. Five days closing the books. One day drilling into every variance against last month and budget. One more turning all of it into a board package anyone could follow.

So when Carl Seidman opened our recent webinar by asking how many people use generative AI at work, I already knew the chat would light up.

And it did.

“Claude. Copilot. Every day.”

Then came the harder question:

How many of you would trust what it gives you 100% of the time?

“No. Never. Absolutely not. Not with finance.”

One attendee, David, put it in three words: trust, but verify. And that tension, between what AI can do and what finance professionals are actually willing to trust it to do, is exactly what Carl spent the next hour putting to the test.

Carl is one of only 35 US-based Excel MVPs and has taught more than 250,000 people how to build better financial models. He was joined by Teague Sanders, ERP data analyst at Velixo who works with our customers across nonprofit and for-profit accounting every day. Together they walked through what AI can really do for FP&A right now, where it still falls short, and how to work with it without losing your CFO’s trust.

One dataset, 30,000 rows

Carl built the whole session around a demo company, Ace Construction, with a chart of accounts and a transaction dataset of just over 30,000 rows exported from the accounting system. Then he did something an accounting team might do without telling anyone: he deleted three accounts from the current data and asked whether AI could find what changed.

He could have filtered 30,000 rows by hand. Instead, he opened Microsoft’s new Finance Agents add-in for Excel, which most of the audience had never seen. Before he asked it anything, it had already profiled the sheet: 30,010 rows in the current table, 30,149 in the prior, top accounts by total value. From there he ran its financial reconciliation, which mapped the two tables on suggested keys, matched what it could, and sorted every row into matched, potentially matched, and unmatched. The three deleted accounts sat in the unmatched bucket. No filtering, no manual tie-out. The add-in did freeze partway through, live in front of a few hundred people, and Carl turned it into a teaching moment instead of a stumble: when he asked how many attendees had ever had AI freeze on them mid-task, the answer was pretty much everyone.

The rules that turn AI from a tool into a colleague

The most practical part of the session was something Copilot in Excel calls workbook “rules” and “skills”, and here’s what that means before going further. Think of rules as the standing instructions you’d give a new hire on day one: how we name our tabs, what our color-coding means, which shortcuts we never use. Skills are more like a saved playbook, a set of instructions the tool remembers and reuses every time, instead of you retyping the same guidance into every new file. By default, Copilot in Excel offers five generic example rules. Carl replaced them with his own standards: tab names use underscores, never spaces. The readme comes first, then macro assumptions, then checks, then model schedules. Blue font means hardcoded, black means formula, green means a link to another tab. Never embed a hardcoded number inside a formula. Never use nested IF statements. Avoid volatile functions like OFFSET and INDIRECT when a stable alternative exists. Use XLOOKUP, not VLOOKUP (seriously, take his advice on this one).

Then he showed a file he had not built. AI had built it, following every one of those rules, down to the tab order and the font colors. His advice on writing the rules themselves: have AI draft them in the language AI understands best, then ask it where the rules are still ambiguous. Talk to it like an intern seeing your model for the first time.

But Carl isn’t starting from scratch every time he opens a new workbook. That’s where skills come in. He’s already built a few for himself: one that assembles a 13-week cash flow model, another that audits a financial model for errors. These aren’t generic templates he downloaded, they are standards he wrote once, for his own work, and now every workbook he opens starts with them already loaded, so he never has to re-explain his rules to a new file again. That’s the real time savings. Not that AI can build a model, but that it can build the same reliable version of your model every single time.

Building the P&L was easy. Defending it wasn’t.

Then Carl showed where the rules stop helping. He asked AI to build a monthly P&L from the transaction data, and it did, following his rules, in about 17 minutes. The output looked right. But when he imagined his CFO pointing at other operating expenses and asking what happened there, he had no answer. The numbers came from a formula pointed at a pasted dataset, and there was nowhere to drill down.

This is the exact wall I used to hit as a controller.

The math was rarely the hard part. The hard part was explaining why the number moved.

Every variance needed a story I could actually defend in the room. Where did it come from? Was it timing? A coding issue? A real change in the business? Is it going to happen again?

A number without a path back to the transaction behind it isn’t useful to the Director of Finance, CFO or board. It’s just a number.

It got worse with forecasting. He asked AI to project the next year, and it spit out November and December numbers with no visible assumptions explaining where they came from. He asked it to reconcile the current and prior P&Ls, and it flagged the variances without explaining any of them. As Carl put it, that’s spotlighting, and FP&A needs more than a spotlight. Teague was just as blunt: if he couldn’t explain what was behind every number on a P&L, he wouldn’t submit it to his boss.

The missing piece: live ERP data

Teague then ran the same exercise with Velixo Intelligence, working from the same Ace Construction data. The difference was where the numbers came from: the ERP itself, not a pasted export. He typed one plain request, to break out the income statement by account class, month by month, for the year Carl had been working with. Velixo Intelligence consulted the documentation, wrote its own queries, and built the report with live formulas, formatted like a Velixo template. Build it once and it stays current, because nothing was copied out of the system to begin with.

I built the financial reporting model at a previous startup from scratch, and variance analysis always ate my week. I would have wanted this then: something that traces a number back to the source system on its own, instead of me rebuilding the same pivot table for the hundredth time.

The workbook needed no priming. Velixo Intelligence knows how the data is structured in the ERP out of the box, which removes the part of AI work that usually falls on the user. Prompting is a skill, and writing prompts that hold up is its own tedious project. Models change, and a prompt that worked last month starts throwing errors. Velixo Intelligence takes that on so the user can type what they want and get the report.

The difference showed up when he asked it to point out variances. It flagged material sales volatility, noticed the missing November and December data as a possible oversight, and suggested where to look. So, he asked it to drill down into other expenses. It pulled the underlying transactions directly from the ERP and surfaced the driver: subcontractor charges sitting in other expenses, exactly where a construction company wouldn’t want them. The question Carl couldn’t answer for his CFO took Teague one prompt, and every number traced back to the source system.

The takeaway

Let AI find the number. Let it build the model. Let it save you hours. But when your CFO asks “why?”, make sure you can answer.

Watch the full on-demand webinar to see how Carl and Teague did exactly that.