If the person running all your AI tools quit tomorrow, how much of that progress would your company actually keep? For most teams I look at, the honest answer is very little. The time savings were real, but they lived in one person's chat history, and they leave when that person does.
Here's the part that doesn't fit in sixty seconds. Saving hours is a personal win. Keeping what those hours produced is a company asset, and it is a different thing to build. I sit on the finance and investor side, so I read AI the way I read any operating improvement: not by what it did once, but by whether the business still has it a year later. Three questions get me there.
The three harder questions
Time saved is the headline. These are the questions I ask underneath it, in the order I ask them.
Where does the knowledge live?
When someone uses AI to fix a client problem, does the next person get a record they can actually use, or do they open a fresh chat and start over? A private thread that only its author can find is not company knowledge. Capture means the solution lands somewhere shared and searchable, written so a colleague who was not there can follow it: what the problem was, what fixed it, and what to do if it happens again.
Is the judgment improving?
AI produces a lot of output. That is not the same as better decisions. Pick real examples where the stakes were real, a proposal, a scope call, a hiring note, and check whether the decision got better, not just faster. The test I like most: could the team catch a promise it can't deliver before the proposal goes out? If the AI helps you say no to the wrong commitment, the judgment is improving. If it just makes the yes arrive quicker, it isn't.
Can someone else run it?
Hand a second person the same files and the same instructions, and watch where they get stuck without the person who built it. Every place they stall is a piece of the process that lives in one head instead of on the page. A process that only runs when its author is in the room is a dependency, not a capability. Teachable means a competent colleague can pick it up cold and get to the same result.
What this looks like on a real team
Abstract questions are easy to agree with and hard to act on, so here is a small studio running all three tests on the same week of work.
The repeated client issue. A client's exports keep breaking because their file naming doesn't match the template. The first time, someone uses AI to work out the fix in twenty minutes. Good. The capture test asks what happens the third time it breaks, when a different teammate catches it. If the fix went into a shared troubleshooting doc, they reuse it in two minutes. If it stayed in the first person's chat, they rediscover it from scratch, and the twenty minutes gets paid again by everyone in turn.
The proposal with a scope conflict. A draft proposal promises a two-week turnaround, but the same client's contract already commits the team to a launch that overlaps that window. AI drafted the proposal fast. The judgment test asks whether anything in the process caught the conflict before it went out. A team that has AI check each new promise against existing commitments catches it and adjusts the date. A team using AI only to write faster sends a deadline it can't hold, and pays for the speed later.
The handoff to a second operator. The person who built the weekly client-report process is out for two weeks. A colleague gets the folder and the instructions and tries to run it. The teachability test is simply where they get stuck. If they stall because the instructions say "clean up the data" without saying what clean means, that gap was invisible while the author was around to fill it, and it just became the thing to fix. Every stall is a to-do, not a failure.
The audit I'd run on your own team
You don't need a consultant for this. Take one real piece of work that AI touched this month and run it through the three checks below. Write down what you actually found, not what you hope is true, and the last column tells you what to fix first.
The work I'm auditing: [one real task AI helped with this month]
01 CAPTURE: where does the knowledge live?
- Where is the solution recorded right now: [ ]
- Could a colleague find it without asking the author: [yes / no]
- What I found: [ ]
- Fix: [ ]
02 JUDGMENT: is the decision better, not just faster?
- The real decision it shaped: [ ]
- Would this process catch a promise we can't keep: [yes / no]
- What I found: [ ]
- Fix: [ ]
03 TEACHABILITY: can someone else run it?
- Second person, same files and instructions, no author in the room
- Where did they get stuck: [ ]
- What was in the builder's head that wasn't on the page: [ ]
- Fix: [ ]
Top thing to fix this week: [ ]What to watch for
The failures here are quiet. Nothing breaks loudly, which is exactly why they get missed until the person leaves.
Knowledge trapped in one person's chat history is the most common one. The work got done, the client was happy, and the record of how it was done sits in a private thread nobody else will ever open. It reads like progress right up until that person is out sick and the same problem comes back.
Then there is output that nobody checks. The AI produces summaries, drafts, and analyses on schedule, and everyone assumes the decisions behind them got sharper. If no one ever pulls a few real calls and asks whether they came out better, you have volume standing in for quality, and no way to tell the difference.
Watch for the process that only works when its builder is in the room. It runs beautifully, on time, every week, and then they take a vacation and it quietly stops. The smoothness was the person, not the system, and you find out which only when they step away.
The one I'd watch hardest: measuring hours saved while quality drifts. Hours saved is easy to count and looks great in a summary. Meanwhile the drafts get a little more generic, the judgment calls get a little lazier, and because the dashboard is green, no one goes looking. Count the time if you like, but read the actual output too.
Ten hours a week is where this starts, not where it ends. Capture the knowledge so it outlives the chat, check whether the judgment is genuinely improving, and make the process teachable to someone who wasn't there. Do those three and the progress belongs to the company. Skip them and it belongs to whoever happens to be running the tools this quarter.
Where to point this next
The audit tells you what to fix. Two guides pick up where it lands. If the teachability check kept stalling, too much still runs through one person, and which parts of your business should AI handle helps you decide what to hand off and what to keep close. If capture is where you fell down, the invoice-assistant walkthrough shows one process built to leave a record end to end. This is deliberately not the same question as what makes a business worth buying: that one is about the value a buyer receives, while this one is about whether your knowledge and capability survive a staff change at all.