Capture what your retiring experts know
The person who has run billing, estimating, or scheduling for thirty years is going to retire. Most of what they know is not written down. Here is the way we capture it that actually works, from an engagement running in Wisconsin right now.
Put the AI on the job next to them, and write down every correction
Somewhere in your company there is a person who has done one job for twenty or thirty years. Which paper charge applies. How cartons bill when an order drop-ships. Which freight lines stay off the invoice. Which customer always gets the old rate. Most of it is not written down anywhere, and when they retire it leaves with them.
The usual answer is to ask them to write a manual, or to shadow them for a month. Both produce a document nobody trusts, because the knowledge was never a list. It is a few hundred small judgments learned one job at a time.
The approach that works is to put an AI system on the job beside them. It does the work, they correct what it gets wrong, and every correction becomes a written rule the system follows next time. Weeks in, the company has two things it never had: the work done in minutes instead of hours, and the rules on paper.
Five steps, in this order
The order matters. Picking the wrong first job is the most common way this stalls.
Pick the job only one person can do
Not the biggest job, the most concentrated one. The test is simple: if this person were out for a month, what would stop? Invoicing, estimating, scheduling, and the month-end close are the usual answers. Our three-week assessment scores each candidate on how much time it takes, how much of it depends on one head, and how ready the data is.
Put the AI on the job, with them
The system starts doing the work on real jobs, and the expert reviews every result before it goes out. Nothing bypasses them. In the first weeks it gets a lot wrong, and that is the point.
Every correction becomes a rule
When the expert changes something, we ask why, and the answer is written down as a rule the system follows from then on. "Freight on a drop-ship order stays off the invoice unless the customer asked for it" is one line. There are usually a few hundred of them, and nobody knew the number until now.
The rules become the manual the company never had
Because the rules are written in plain language and tested on real work, they double as documentation anyone can read. A new hire can learn the job from them. A manager can see, for the first time, how the work is really done.
The system keeps running after the retirement party
By the time the expert leaves, the corrections have slowed to a trickle and someone junior has taken over the review. The job that used to take a person hours takes minutes, and the judgment behind it is still in the building.
A Wisconsin printer, one billing desk, thirty years of rules
At a Wisconsin commercial printer we work with, the person who runs billing has been doing it for almost thirty years. Dozens of rules about paper charges, cartons, and freight lived in her head and nowhere else. Since early this year we have been building AI that does the invoicing with her: it drafts the invoice from the job data, she corrects what is wrong, and each correction becomes a rule the system knows next time.
Invoice work that took hours now takes minutes, and the company has a written record of thirty years of billing judgment it never had before. The same engagement grew into four projects: invoice generation, vendor invoice processing, an accounting workflow that replaces the spreadsheets long held by retiring staff, and an operations dashboard that replaces the daily manual check-ins. Some tracks are deployed, others are still in discovery.
We hear versions of this everywhere now. A food manufacturer told us about someone who has been there 28 years and whose absence for a month would leave nobody ready to step in. Michael writes about this work as it happens in his field notes, and the manufacturing version of the story is on the manufacturing page.
Four things to insist on
The expert stays in the loop and is paid to correct. The corrections are the product. A build that tries to work around the expert produces a system that is confidently wrong in exactly the cases that matter.
The rules are written where your company can read them. Plain language, in your systems, not buried in a model nobody can inspect. If the builder cannot show you the rule list, the knowledge is not captured, it is hidden.
It runs in your accounts and you own it. Your cloud, your keys, your code. The whole point is that the knowledge stays in the building, so it should not live on a vendor's servers either. How we handle ownership and fees →
Start with one job, not a program. One desk, one workflow, one expert. The second and third come faster once the first is running and the company has seen what a rule looks like.
The ones that come up next
Is this different from writing a manual?
Yes, in one important way. A manual is written from memory, once, and goes stale. Here the rules get written while the work is being done, one correction at a time, and each one is tested on the next job. The document that results is the same one the system runs on, so it cannot drift away from how the work is really done.
Our expert does not want to be replaced. How does this land with them?
Better than you would expect, because the design depends on them. Their job during the build is to catch what the system gets wrong and say why. That is respect, not replacement, and most long-tenured people like seeing their judgment written down for the first time. Nobody is asked to train their replacement. They are asked to teach a tool that will keep doing the job their way.
How long before it is useful?
The first workflow is usually running, with the expert correcting it, within weeks. We start with a three-week assessment that maps which jobs depend on one person and which of those AI can take on first, then a fixed-fee build of the top one. You are not waiting for a finished system before anything changes.
What does it cost?
The three-week assessment is a fixed fee, typically $10K to $25K depending on the size of the company and how many workflows we look at. A build is quoted from the assessment, fixed fee, most often $25K to $150K+ depending on the scope. There is no commitment past the assessment itself, and the map is yours either way.
Can a Wisconsin training grant help pay for it?
For the training part, yes. WisTRAIN reimburses Wisconsin employers for up to 80 percent of eligible AI training costs, and teaching your own staff to work alongside the new system counts. We help write the application. The build itself is not a training cost, so it is not covered.
What kinds of businesses is this for?
Any business where the way the work gets done lives in a few long-tenured heads: manufacturers and printers with a billing or estimating desk run by one person, distributors with a scheduler everyone calls, agencies with a renewal process nobody wrote down, credit unions with one person who knows the exception cases. The size we see most is 50 to 500 employees.
- Featured by Apple in Best New Apps & Updates (LilSense)
- Serving collegiate and professional programs, including an NCAA national championship team
- Founded Wisconsin's longest-running AI practitioner community (meeting weekly since 2024)
- Co-hosts the AI Leadership Breakfast Forum with Steve Cretney (EVP, Colony Brands)
Who is the one person in your company who knows how it is done?
A 30-minute conversation about which job to start with. No pitch. If it is not a fit for AI, we will say so.
Book a Call