The Job Got Done. The Notes Didn't.

The crew knew exactly what happened on that job. They'd been standing in it for six hours, in the weather, up a pole, with their hands on the problem. They could have told you every detail over a cup of coffee.

The work order said: "Fixed."

That gap, between what the crew knew and what the record showed, is the whole story of a project we worked on recently with an electric utility. They keep the power flowing across a big stretch of rural country, and their field crews spend all day doing the kind of work that keeps a region lit. Smart people. Deep knowledge. Almost none of it was making it into the system.

If you run any business where the real expertise lives out in the field, in a truck, on a job site, on the floor, this one is for you. Because the problem wasn't the one everyone assumed.

Everyone thought it was a documentation problem

When a company notices its records are thin, the usual reaction is to push harder on documentation. Add required fields. Send a reminder. Remind people in the meeting that notes matter. Maybe write them up when the notes are empty.

None of that works for long, and here's why. The crews weren't skipping notes because they didn't care. They were skipping them because typing a detailed account of a complicated job, on a tablet, at the end of a long shift, is genuinely miserable. So they did what any reasonable person does with a miserable task. The bare minimum.

That is not a discipline problem. It's a friction problem. The cost of capturing what they knew was higher than the perceived payoff, so the knowledge stayed in their heads and walked off the job site with them.

And the office paid for it later. When a question came up weeks after a job, nobody could reconstruct what actually happened. Was that the second visit or the third? Did we replace the part or just inspect it? What did the meter read when we left? The answers existed. They just existed in a person's memory, which is not a place you can run a report on.

The information wasn't missing. It was trapped in the one place you can't search: somebody's head.

The real question we asked

We've learned to be suspicious of the obvious framing. "Help our crews document better" is a solution dressed up as a problem. The better question was quieter:

Where is the knowledge, and why isn't it getting written down?

Once you ask it that way, the answer is almost funny. The knowledge is in the crew's heads, fully formed, the moment they finish a job. They could talk about it for five minutes without pausing. The only thing standing between that and a useful record was the act of typing.

So we didn't try to make them type more. We tried to make typing unnecessary.

What we built

The tool we built is boring to describe, which is usually a good sign. A crew member finishes a job, opens the tool, talks for about thirty seconds about what they did, and snaps a few photos. That's the whole ask on their end. Talk, point the camera, done.

Behind that, AI does the part nobody wants to do:

• It transcribes the speech. Thirty seconds of a tired person talking turns into clean text.
• It summarizes the job. Not a word-for-word dump, but a structured account: what the work was, what got done, what still needs doing.
• It pulls out the numbers that matter. A meter ID, a voltage reading, a cost. The specific values a record actually needs.
• It keeps the raw material. The photos and the full transcript stay attached, so nothing the person said gets lost in the summary.

A job that used to produce the word "Fixed" now produces a real account, with next steps and pictures, from half a minute of talking. The office can finally see what the crew always knew.

A detail we were careful about: the tool only opens when it's tied to a specific job. You can't fire it up and accidentally file a note against the wrong work order or no work order at all. Small thing, but it's the kind of small thing that decides whether a tool gets trusted or quietly abandoned.

What we were careful about

This is the part that matters more than the clever features, so I want to be honest about it.

AI is worst at exactly the thing that has to be perfect. A summary can be a little loose and still be useful. A meter number cannot. If the tool hears "forty-two" and writes "forty-seven," that's not a typo, that's a wrong record that looks right. So we never let the AI's extracted numbers go straight through. The tool pulls them out and puts them in front of the crew member to confirm or fix before anything is submitted. The machine does the capturing. The human owns the few facts that have to be exact.

That one design choice is the difference between a tool people trust and a tool that quietly poisons your data. It's the same reason we keep a human on the numbers whenever AI is confidently wrong: a loose summary is forgivable, a wrong reading that looks right is not.

It can mistake one number for another. When you teach software to read numbers off a photo, it will cheerfully read numbers that aren't the ones you want. A license plate in the background is also a number. So part of the real work was teaching it which numbers to care about and confirming how many different formats it might see. This is unglamorous and necessary, and it's ongoing.

We didn't nail it on the first try. This was actually a rebuild. An earlier version existed, and it wasn't good enough. Newer transcription models, better summary rules, and a cleaner flow made the new one a real step up, and the client's early reaction was that it was a big leap and the crews would be a lot happier. But we got there by shipping, watching it struggle, and fixing it. Anyone who tells you their AI project worked perfectly on day one is editing the story.

There are still live bugs. The recording can drop when someone switches to another app mid-note. We know. It's on the list. I mention it because the honest version of a build-in-progress includes the parts that aren't done, and pretending otherwise just sets a client up to feel lied to.

We're adding one job type at a time. There are many kinds of work orders. We started with the simplest category, got it right, and only then moved to the next. The temptation is always to cover everything at once. The faster path to something people actually use is to do one thing well and expand from there.

The lesson you can take to your own business

Here's the part that travels, no matter what you do.

Your most valuable operational knowledge almost certainly lives in your people, not your systems. The tech who's been there eleven years. The driver who knows which accounts are touchy. The crew that just finished the job. They know things your records don't, and the reason your records don't is rarely laziness. It's friction. Writing it down costs more than it seems to be worth in the moment.

So stop asking "how do I get my team to document more." That question has been losing for decades. Ask instead:

• Where is the knowledge actually sitting right now, and in what form? Usually in someone's head, as speech they could give you in a minute.
• What's the cheapest possible way for them to get it out? Often talking, not typing.
• Which parts have to be exact, and who verifies those? Keep a human on the numbers.

AI is good at the middle of that. It reads the mess, a rambling thirty-second voice note and a couple of blurry photos, and turns it into something structured and searchable. It does not replace the person who did the work or the judgment of the person reviewing it. It replaces the awful, friction-filled step in between, the part everyone was avoiding anyway. If you want the version of this aimed at the knowledge stuck in your own head rather than a field crew's, we wrote about getting it out of your head, and about why you can't automate what you can't explain in the first place.

That's the pattern worth stealing. Not "add AI." Lower the cost of capturing what your people already know, and let the software do the tedious translation. It only works because the knowledge is yours: the specifics a generic model will never have, which is the whole reason AI doesn't know your business until you feed it what your people do.

If your business is sitting on a pile of hard-won knowledge that never makes it out of your team's heads and into something you can use, that gap is exactly the kind of thing we help close at Humanity AI. Usually it starts with one unglamorous question about where the information is really hiding.