Three Months Ago, I Wouldn't Have Recommended an AI-Native Budget

Three months ago, I would not have recommended running a budget process through AI.

Not because I'm skeptical of AI.
Because it would have been irresponsible.

At the time, the technology wasn't capable of supporting a real budget process end-to-end — not just analysis, but implementation, iteration, and discipline around change. Recommending otherwise would have created more risk than value.

That assessment changed recently.
And not because AI suddenly got better at formulas.

What Actually Happened

I didn't start by asking Claude to "build a budget model."

I asked it to build a real consulting-firm budget — the kind that would fall apart immediately if the logic was wrong.

The first version did exactly what most generic models do. It treated revenue as a function of headcount, utilization, and rates.

That's backwards.

Real consulting firms forecast revenue first — by client, by timing, by confidence — and then back into required capacity and hiring.

When I pointed this out, the model wasn't tweaked.

It was rebuilt.

The Rebuild

The second version flipped the logic entirely.

Revenue became the anchor: contracted revenue from existing clients, probability-weighted pipeline, explicit new business targets with a realistic quarterly ramp, and realization applied after gross revenue rather than buried in utilization.

From there, everything else flowed correctly.

Required billable hours were derived from revenue divided by blended billing rates. Hours were distributed across levels using an explicit delivery mix. Hiring was driven by capacity gaps rather than wishful thinking, and attrition was calculated in a way that avoided circular references.

This is all standard FP&A logic.

What was different was that every structural change was explicit, deliberate, and captured as it happened.

The Moment It Clicked

The moment this stopped feeling theoretical was during scenario planning.

Claude built three demand scenarios — Downside, Base, and Upside — each with different assumptions for client extensions, pipeline conversion, new business targets, and realization.

The revenue logic was right.

But when I looked at the hiring plan, something felt off.

The hiring plan was the same across all three scenarios.

Anyone who has sat in a real budget review knows this moment. This is exactly where a CFO would say, "You can't hire the same way in a downside and an upside. That's not how we actually operate."

So I said exactly that.

Claude didn't argue.
It didn't defend the original approach.
It implemented the feedback.

It created three distinct hiring plans — one for each scenario — and rewired the model so the active hiring plan switched automatically with the scenario selector. Downstream headcount, capacity, compensation, operating expenses, and the P&L all flowed correctly. And the change was logged explicitly, with rationale.

No silent overrides.
No forgotten context.

Decision → implementation → memory.

That's when this started to feel less like "AI helping with a model" and more like how a CFO actually works with an FP&A team.

Why I Couldn't Say This Three Months Ago

Until very recently, AI tools could help around budgeting. They could explain models, stress-test outputs, or answer questions about work someone else had done.

They could not reliably rebuild a model, manage cascading impacts, preserve decision history, or enforce discipline without taking control away from humans.

That distinction matters.

Budgeting isn't hard because the math is hard. It's hard because decisions get made, revised, half-reversed, and forgotten — all while the model keeps moving.

Three months ago, AI couldn't handle that reality responsibly.

Now, it can.

A Necessary Clarification

The human side of budgeting has always existed.

The best FP&A teams already exercised good judgment, debated tradeoffs, and built strong narratives. The problem was doing that while also living deep in the mechanics of the model.

You either lived in Excel and lost the bigger picture, or you led the process and relied on fragile implementation.

What changed is not the need for judgment. What changed is the ability to separate judgment from mechanical execution without losing rigor.

Humans still decide.
AI implements those decisions cleanly and consistently — and remembers why they were made.

Why This Matters

This isn't "AI runs the budget."
And it isn't "AI replaces finance."

It's more practical than that.

AI becomes the layer that translates decisions into precise model changes, propagates those changes correctly, documents what happened and why, and preserves context so it doesn't disappear mid-year.

There's a useful parallel here. The general ledger works because it's auditable and immutable. Budgeting historically hasn't been.

AI makes that kind of discipline possible for planning — without turning budgeting into accounting.

The Good News on Timing

If you just finished — or are still in the middle of — budget season, this might feel like an ouch moment.

It shouldn't.

This isn't about redoing this year's budget. It's about recognizing that we now have nine months to design a better process for next year — thoughtfully, deliberately, without rushing.

That's a luxury most process changes don't offer.

Where This Is Headed

By the end of this exercise, the model itself was impressive in a very ordinary way: a full consulting-firm budget, month by month, with well over a thousand formulas tying revenue, capacity, hiring, compensation, and expenses together — and no errors.

But that wasn't the breakthrough.

The breakthrough was that the model didn't just update.
It remembered.

Three months ago, I wouldn't have recommended an AI-native budget process.

Having watched this happen step by step, now I can.

That's not hype.
That's a threshold being crossed.

Next issue: why experimenting on a firm that isn't yours may be the safest way to innovate on the one that is.