It's Not Excel

He’s using the wrong yardstick.

CPA firm owner I work with spent three weeks trying to get his engagement letters to come out the same way every time. He’d run the same prompt through Gemini once — “Draft an engagement letter for a tax audit client, $6,000 scope, three-month timeline.” Gemini would draft it. He’d review, tweak tone, send it.

Three days later — same scope, similar client — he ran the prompt again. Different engagement letter. Similar structure, but the tone shifted, the opening changed, the emphasis moved. He ran it a third time two weeks after that. Another version.

After the third one he called me frustrated. He’d been literally expecting Excel to show up wearing an AI wrapper.

Yeah. Problem wasn’t the tool. Problem was the standard he was using to judge it.

Here’s the thing. You’ve spent twenty, thirty, maybe forty years with spreadsheets. Type a formula once, it calculates the same way forever. Consistency is the whole design. That muscle is burned so deep you don’t think about it anymore — it’s just how systems work. You built a business on it. Tax time comes, you pull the same client files, run the same procedures, get the same output structure every single time. That’s predictability. That’s how you know the system is working.

You brought that instinct to AI. Same input should produce the same output. If it doesn’t, the system is broken.

Right? Except the system isn’t broken. It’s just not spreadsheet logic.

Excel is mathematical. Same input, same output, always — that’s the entire architecture. Precision. Consistency. Determinism. You feed it numbers, it feeds back the answer.

AI is different. AI produces different outputs from the same input. Not wildly different every time — the structure stays similar, the shape holds — but different. Same engagement letter for the same audit client might open warmer one day, more formal another day. Compliance language sits in the same spot. Scope clarity is there. But the approach varies.

Does that make sense? Think about how you’d actually train a junior accountant to draft engagement letters. You wouldn’t hand them a checklist — step one, state the client name and engagement date. Step two, list three services. Step three, include fee language. Step four, close with timeline. They’d produce checklist-shaped letters.

Instead, you’d have them sit in on calls with clients. You’d give them three engagement letters you’re proud of and three you’re not and say “figure out the difference. What lands? What doesn’t?” Some days your junior nails it — clear, warm, client-friendly. Friday they draft another one. Still professional, still compliant, but the tone shifted. Maybe they’re overthinking it. Maybe they nailed a different angle. Same person. Same task. Different output every time.

That junior isn’t broken. That’s just how thinking works.

AI operates the same way, except it happens in seconds instead of over weeks. Same prompt, slightly different take every single time. Not because it’s malfunctioning — because that’s how it’s engineered to work.

So when the CPA owner got three engagement letter drafts back, what he actually had was three different approaches to the same task. One opened warmer and built trust. One was more formal, corporate-sounding. One landed somewhere in between. His instinct was to conclude the tool was unreliable. Wrong conclusion.

He had information. He had a menu. And he was looking for the one “correct” answer when what he should’ve been doing was noticing which version served the client relationship better. That one warm opening — the one that built trust. He’d pull that direction. Refine it. Build from there.

That’s not a system failure. That’s the system working.

Here’s the thing — the frustration isn’t coming from the tool. It’s coming from you judging a jazz musician by sheet-music accuracy.

I ran audits on five mid-size CPA firms last month. Every single one had deployed the same AI for something — engagement letters, client tax summaries, follow-up communications. Every single one hit the same wall. They ran the same prompt twice, got slightly different outputs, concluded the tool wasn’t reliable. None of them noticed the different outputs were showing them different angles they hadn’t considered. None of them picked the better version. They were all looking for Excel. They got something else. Something more useful, actually, if you stop fighting it.

So here’s the actual move. Pick one thing you’ve been drafting with AI. Engagement letter. Client tax summary. Follow-up communication. Anything you draft more than once. Run the same prompt three times. Don’t change anything between runs. Just get three outputs sitting side by side.

Compare them. Not looking for the “correct” one — looking for the one that serves your client relationship best. The one that’s clearest. The one that opens the right door. One of them will.

When you find it, keep that version. Or note what you liked about it so you can guide the next round better. Point isn’t to get the same answer every time. Point is to get enough options that you can pick the best direction.

Here’s the thing. You’ve been running a tool like it’s a spreadsheet — every input should produce identical output. Control. Predictability. Consistency. You spent your entire career mastering consistency. It’s exactly the muscle that’s tripping you up now.

For twenty years you’ve been judging everything by spreadsheet standards. That was the right standard for spreadsheets.

So. New tool, new yardstick. Put the old one down.

Framework: The Professional Recipe — Operating Principle #1 (It’s Not Excel), which explains how AI systems differ from traditional spreadsheet-based logic.

Companion pieces: Stop Hiring AI. Start Building It. — the parent framework introducing the three operating principles that govern AI work. The Process Cage — the sister piece on misapplied instincts, where old muscle memory built one kind of cage, here it’s built another.