We document as we go, and hand over the keys

There’s a version of automation that looks like a clear win for about eight months. A workflow gets built. The reports come out on time. One person on the team knows how it actually works. Then that person takes another job, or goes on leave, or moves to a different role — and the thing quietly stops. Nobody touches it, because nobody knows what it will break.

The cost shows up long after the win

That isn’t a rare event. In June 2026, 3.2 million people quit a job in a single month, a quits rate of 2.0 percent, according to the Bureau of Labor Statistics. Stretch that across a year and, in a business with a dozen people, somebody leaving isn’t a disaster scenario. It’s the ordinary condition.

Most small operations already plan for this on the people side. They cross-train the front counter. They keep a second person who can run payroll. What they generally don’t do is apply the same thinking to the automations they’ve bought or built, because those feel like infrastructure — like the electrical service, something that just runs.

It isn’t. An automated workflow is a set of decisions somebody made about your business: which invoices get flagged, what counts as a duplicate, which messages route to whom, what threshold triggers a human review. Those decisions live somewhere. If the only place they live is in one person’s head, you don’t really own the automation. You’re renting it from an employee.

Most AI use is already off the books

Federal data suggests this is more common than most owners think. In the Census Bureau’s 2026 AI supplement to its Business Trends and Outlook Survey, covering November 2025 through January 2026, researchers found that worker task use sometimes occurs without formal firm-level adoption — people are using AI to get their jobs done at businesses that wouldn’t describe themselves as AI adopters at all.

Where adoption is formal, it’s usually narrow. The same paper found 57% of adopting firms use AI in three or fewer business functions, and 65% limit worker use to three or fewer tasks.

Read those together and you get a picture that matches what we see in the field. AI in a small business is usually not a program. It’s two or three people who figured out something useful and kept doing it. That’s genuinely good — it’s how most real improvement starts. It’s also completely undocumented, and it belongs to those two or three people rather than to the business.

The question isn’t whether your team is using this stuff. It’s whether what they’ve figured out survives their two weeks’ notice.

What actually gets handed over

So we write things down as we build them, not afterward in a documentation phase that gets cut when the budget gets tight. Four things end up in your hands:

Why we think this is also why it works

There’s a self-interested version of this argument and an honest one, so here’s the honest one. A widely-cited MIT Project NANDA report found that most generative-AI initiatives stall not on model quality but on what its authors called a “learning gap” — tools that never learn from or adapt to how the work actually moves. The same report found AI brought in through a vendor or partnership succeeded roughly 67% of the time, while internal builds succeeded about a third as often.

We’ll name our stake in that second number: we’re the outside partner in that comparison, so treat it as one report rather than a settled fact. But the mechanism underneath it is the part we’d bet on either way. What makes an automation stick is whether it was built around how the work genuinely moves — and whether anybody wrote that down in a form the next person can read.

When documenting isn’t the answer

Documentation isn’t free, and there’s a point where writing something down costs more than the thing is worth. If a task takes eleven minutes a month, leave it alone. We’ll say so.

If a process is about to change — you’re moving systems next quarter, or the seasonal swing through orchard country means October runs nothing like June — then documenting the current version is waste. Wait until it settles.

And documentation can’t rescue a process that’s wrong. If the month-end close takes nine days because three people are re-keying the same numbers into three systems, careful documentation produces an excellent description of a bad process. Fix the sequence first, then write down the fix. We’d rather say that in the first conversation than bill for the wrong order of operations.

The reason we work this way

It isn’t generosity. Businesses across the river valleys and the I-94 corridor have been sold plenty of systems that only the seller can operate, and that arrangement has a way of outliving its usefulness. We’d rather be the firm you call when something new comes up than the firm you can’t afford to stop calling.

Which is also why we ship one working thing early instead of a six-month plan. A documented workflow that saves four hours a week, running by the end of the month, teaches your team more about what’s possible than any roadmap will. And if we disappeared afterward, it would keep running.

A note on the numbers: adoption surveys disagree with each other because they ask different questions — “used AI in the past two weeks” and “used AI in a business function” measure different things. Every figure above links to its primary source so you can check the definition yourself. We keep a running Sources page grading each citation by its distance from original research.