← All Essays

The Four Stages of AI in Business

Coding lived through this in five years. Every other white-collar industry is next.

I've been thinking a lot recently about how AI disruption is going to play out across businesses, and I've landed on a framework I wanted to share. Here's how I see the next decade unfolding.


To keep things simple, I'll use tax accounting as the example throughout. But keep in mind, this same arc will play out in legal, healthcare, radiology, bookkeeping, recruiting, financial planning, and just about every other white-collar industry.

Stage 1: No AI.

A human does the work, start to finish. You walk into a CPA's office with a shoebox of receipts. They sort through it, ask you questions, prepare your return by hand (or with basic tax software), and bill you for their time. The pricing model, staffing model, and customer expectation are all built around human hours. This is where most of the world lived until very recently.

Stage 2: Human-first, AI-assisted.

The CPA still does the work, but uses AI to go faster. They use AI to categorize transactions, summarize receipts, draft the first pass of the return, and answer client questions in plain English. Then they review everything and file it.

Stage 3: AI-first, human-reviewed.

This is the inversion. The AI does the work. The human reviews it.

The role of the CPA has changed. They're no longer producing returns. They're checking them. One reviewer can supervise the output of what used to require ten or twenty preparers.

This isn't theoretical. Look at Deduction.com, an AI-first tax service. An AI named Taylor handles your documents, advises you year-round, and drafts your return. As they say on their site: every return is verified by a licensed CPA before it's filed. The AI does the work. The CPA reviews and signs. One CPA can now cover what used to require an entire team of preparers, which is why Deduction can charge $499/year for unlimited advice plus your annual filing while a traditional CPA has to bill thousands for the same work.

And it's not just tax. The same model is already shipping in legal, healthcare, radiology, bookkeeping, recruiting, and dozens of other industries you wouldn't expect.

Stage 4: AI only.

The human review layer goes away. The AI prepares the return and files it directly with the IRS. No CPA in the loop. The whole interaction is: you connect your accounts, the AI does everything else.

Stage 5: The trillion-dollar question I don't have an answer to.

Once an industry reaches Stage 4, what happens next? I see two very different futures.

Path A: A thousand vertical AI companies.

Deduction owns tax, another ai first service owns legal, another owns bookkeeping, while another owns radiology, and another recruiting. A hundred more own everything else. Each industry gets its own AI-native winner, defended by licensing, regulatory relationships, integrations, and trust.

Path B: The foundation models eat everything.

I open Claude or ChatGPT and say "file my taxes." It does it. "Handle this legal contract." Done. "Close the books this month. Find me a new hire for my open position. Review this MRI." Done, done, done, done, done. The vertical players get squeezed the way standalone weather and flashlight apps got squeezed when the phone OS absorbed them.

What makes me lean uncomfortably toward Path B is that coding already went this way. Five years ago, you might have predicted a hundred specialized vertical coding tools. Instead, the dominant coding agents today are either the foundation model labs themselves (Claude Code) or thin wrappers on top of them (Cursor, Copilot). The vertical didn't survive. The models ate it.

And it's not just coding anymore. In May 2026, Anthropic launched Claude for Legal, having chosen path b. The foundation model didn't even wait for the vertical winner to mature. It's already eating into the lunch.

But maybe the moats hold somewhere. Tax, healthcare, finance, etc. carry licensing, liability, and proprietary data a general-purpose model maybe can’t absorb. Or maybe these vertical products with their trained workflows, customer relationships, and users are stickier and better than the foundation model will ever be in this vertical niches. Or maybe humans will always want the human touch. The honest answer: I don't know.

What do you think? Click here to let me know.