What the Software Engineering Canary Is Telling Accounting

Editor’s Note: This article is part 1 of a 4-part series. View all of Rob Hamilton’s articles here.

In June, ahead of Scaling New Heights, I wrote here that AI would force every firm to decide what business it is in, and I promised a series that went deeper. This is the first of four. Each piece pulls one thread from Zero Entry, the book Jeff Seibert and I published this spring. The full discussion lives in the book, but I am hopeful this series covers the core themes.

Let's start with a (very brief) history of AI

March 2024: The most advanced AI tool a software engineer could get was a very good autocomplete. The cursor guessed the next line, and you pressed Tab.

May 2025: Anthropic shipped Claude Code. You described what you wanted, and it wrote the code for you. Simple things worked. Anything complicated fell apart.

November 2025: Opus 4.5 arrived with almost no fanfare, and by the holidays engineers were comparing notes on something strange. The complicated things worked now. Senior engineers stopped writing code and started reviewing it.

March 2026: At prominent tech companies, the senior people architect systems and run fleets of agents that do the building. Several of them haven’t typed a line of production code since winter.

That whole arc took two years, and the back half took six months. Jaana Dogan, a Principal Engineer at Google, described handing a coding tool the spec for a distributed orchestrator her team had spent a year building, then watching it produce the equivalent in an hour. Andrej Karpathy was blunter. Anyone who hadn’t kept up over the previous 30 days, he said, had a deprecated worldview.

I’ve spent the better part of a decade talking with firm owners about technology. I have never watched a profession move this fast, and I think accounting is next.

Software went first because it was built to go first

Bret Taylor, who co-created Google Maps and now chairs OpenAI’s board, explained it in one line. A code repository is almost designed for a robot. All the context sits in one place. Code is text, which to a model is just another language. There are compilers that catch mistakes, tests that catch worse ones, and a pull request history recording every decision anyone ever made and why. Roughly 22% of new production code at large companies is now written by something nobody fully understands, which is about the same relationship most of us have with Section 174.

Accounting has none of that scaffolding. The data is scattered across a ledger, a bank portal, a receipt app, and whatever the client emailed last Tuesday. Bank feeds arrive with vendor names that could mean anything. There is nothing resembling a compiler. The description field on a bank transaction was sized for a 1970s mainframe, and I’ll come back to that in the next piece.

It is tempting to read that list as proof the software story can’t repeat here. Read it again. Software went first because its structure already existed and was sitting there waiting for the models to get good enough. Accounting’s structure has to be built, and it is being built right now inside ledgers designed for machines to read. The firms running on them describe their close the way engineers described their winter. A practitioner at PKN ran a bank download on a brand-new client and finished two years of books in four hours.

The canary did not die

The uncomfortable part of this story is that nobody at those companies got a vote. The tech sector cut around 73,000 jobs in the first quarter of 2026, per Crunchbase’s tracker, and the companies doing the cutting were posting record revenue while they did it. The work changed whether the people were ready or not.

And yet the engineers who stayed describe the work as better. Taylor, on the first time he one-shotted something good, called it an emotional experience. The last three months, he said, felt materially different from the prior twenty years. The skill stopped being typing. It became knowing what to build, why it matters, and how the pieces connect. Judgment, in other words, which is what accountants were hired for before the prep work ate the calendar.

Google’s 2025 DORA report added the detail that matters most for your firm. AI did not improve delivery on its own. It amplified whatever was already there. Teams with clean, well-structured codebases flew. Teams sitting on twenty years of workarounds saw marginal gains, sometimes went backward, and many of them blamed the model instead of their own data.

Accounting will split along the same line, and most firms will land on the spaghetti side without knowing it.

What to steal from the engineers

Three things, in order of how much they sting.

Stop measuring AI adoption by who on the team uses ChatGPT. Autocomplete gave engineers a speed boost. Agents gave them a different job. If AI in your firm is drafting client emails and summarizing tax guidance, you are living in March 2024.

Treat your data structure as the ceiling. The DORA finding is the whole next article compressed into one sentence. The model you buy matters far less than what it can grip.

Review at the level of the pattern. The senior engineer no longer reads every line. She reads the design and asks whether the system makes sense. The senior accountant of 2027 reads a month rather than a transaction, and asks whether this picture fits what they know about this business, this industry, and this owner.

Two diagnostics for Monday:

Can AI finish any single workflow in your firm, a close, an onboarding, a tax prep step, end to end, without a person carrying data between systems? And if your most AI-fluent team member resigned on Friday, what survives?

When a senior accountant at your firm goes six months without categorizing a transaction, what will you have them doing instead?

I can’t tell you the exact month accounting’s November 2025 arrives. Software’s two years started in March 2024 with a better autocomplete. Accounting’s started this year with ledgers finally built for a robot to read, and since the models are already trained, I don’t expect it to take two years. The next piece, is about the ledger itself, and the 26 characters a group of California bankers decided were enough to describe where your client’s money went.

Sponsored Content: This article is generously brought to you by one of our valued sponsors. Their support enables us to continue delivering expert insights and the latest industry trends to our dedicated community of accounting professionals.