Your team uses ChatGPT. Your firm hasn't learned a thing.

If your 12-person firm pays for ChatGPT and the work isn’t moving any faster at the firm level, this is why. Individual employees get faster on their own keyboards. The firm itself learns nothing. The only place the AI-era knowledge lives is in people’s heads, and it walks out the door the day one of them quits.

Robert Glaser, Head of Applied AI at Exxeta, made the point cleanly: “Individual productivity gains from AI do not automatically become organizational gains.” His phrase for the gap is “token-to-learning, not token-to-output.” The bill went up. The capability didn’t.

Glaser writes for enterprises with €2M annual AI spend, dedicated Centers of Excellence, and brown-bag sessions. None of that machinery exists in a 12-person accounting firm. The problem is the same. The fix has to look different.

What this looks like in a small firm

Six bookkeepers at the same accounting firm. All six pay for ChatGPT through the firm’s plan. Six different ways of using it.

The first drafts client emails by pasting the previous email and a one-line summary of what to say. Reply rate is up. The second pastes ambiguous expense lines next to the chart of accounts and lets the model categorize them. The third built a 14-step prompt over six months that handles VAT reconciliation in the format their largest client demands. A fourth doesn’t use it at all and is quietly falling behind. The remaining two are running slightly worse versions of what the first one figured out, because nobody told them.

That firm’s AI capability is the sum of six private notebooks. The notebooks don’t get backed up, don’t transfer when someone goes on parental leave, and don’t turn into a process the firm can train a new hire on. The firm pays the subscription and gets the productivity of one bookkeeper times six, with the variance you’d expect from six untrained people each Googling differently.

Glaser’s line for what the firm gets: “visible compliance and invisible learning.”

The Hacker News thread nailed why

The Hacker News discussion on Glaser’s piece went to 348+ points, and the most useful comments are the ones explaining why nobody on the inside fixes this voluntarily.

olsondv put the disincentive plainly: “I’m not going to selflessly share my productivity gains with the broader company for free. I might share a tool if it’s useful. All the learning of how to wrangle AI or set up agents is better kept to myself if there is no recognition for sharing. My company set up a ‘prompt of the week’ award and brown-bag sessions to help spread adoption… Without a real (read ‘monetary’) incentive or job security, the risk and cost of spreading the knowledge falls squarely on the developer.”

That’s the small-firm problem in one comment. If the bookkeeper shares the 14-step VAT prompt, the firm immediately needs fewer bookkeeper-hours for VAT. She is rationally hoarding. Management is rationally frustrated. Brown-bag sessions are theater because the underlying incentive points the other direction.

daheza, a manager: “VPs are starting to ask, how many story points are we getting with AI now… I can’t answer that question but plenty of other managers are fully ready to just give bogus numbers.” nradov added the obvious follow-on: “When management starts tracking improvements in story point productivity then the agile teams inflate their story point estimates.”

The SMB version is the “minutes saved” report. The owner asks how much time the AI is saving. Everyone gives a confident number. The number drifts up, because saying it’s drifting down looks bad.

busterarm on the dependency risk: “Developers themselves are drunk. They’ll be cut off from tools right when they no longer understand the underlying code they’re responsible for.” For an accountant or a paralegal the version is identical. Three years of letting the model draft the response, and then OpenAI raises the price 4x or your prompt stops working after a model update, and the staffer has lost the muscle memory to do it cold.

danaris on the institutional knowledge angle: “they’ll get the people who hold that knowledge laid off, and at least 50% of the institutional knowledge won’t be documented anywhere that even could be fed to the LLM.” Even the AI optimist’s escape hatch (just dump everything in the model and let it answer) fails when the docs were never written.

kj4211cash described a “two timelines” approach inside his employer: official slow-process work, and side-channel “vibe coding” outside the system. The SMB version of two timelines is the bookkeeper with her 14 prompts, the marketing person paying for a personal Claude account because the firm’s plan blocks something, the partner running the entire proposal pipeline through ChatGPT and never telling the rest of the firm. Every shadow workflow is a single point of failure the owner doesn’t know exists.

agloe_dreams on the worst-case ending: private equity firing the gates entirely (QA, sign-offs, code review) to chase the AI productivity story. “We are probably going to bankrupt ourselves from an idiotic mistake somewhere here. But nobody will ever know until it happens. Don’t take those gates for granted.” For a service business it looks like firing the senior staffer who was the last quality check on what the AI sent to clients, and learning six months later that 8% of statements went out wrong.

The Microsoft numbers say the gap is structural

The Microsoft Work Trend Index, May 2024, surveyed 31,000 knowledge workers across 31 countries. Four numbers worth keeping in your head:

75% of global knowledge workers use generative AI at work
78% of those AI users bring their own tools (the report calls it “BYOAI”)
52% of AI users are reluctant to admit using it for important tasks
Only 39% of people using AI received any company training

Those numbers are two years old now and the shape has only sharpened. Three out of four of your staff use AI. Four out of five of the people who use it use a tool the firm doesn’t manage. Half won’t tell you when they used it for important work. A firm running on those defaults can’t capture what’s working, because most of the work happens off-system on accounts the owner can’t see.

The same survey found 60% of leaders worry their organization “lacks a plan and vision to implement AI” and 59% worry about quantifying productivity gains. Everyone knows. Almost nobody fixes it.

Why Mollick’s framework half-applies to you

Ethan Mollick, in May 2025, named the three pieces of organizational AI adoption: Leadership, Lab, and Crowd. Leadership sets the vision and the incentives. The Lab is a small dedicated team that builds and ships AI workflows. The Crowd is the rest of the staff, finding domain-specific applications through trial and error.

For a 12-person firm, Leadership is the owner. The Crowd is everyone else. The Lab doesn’t exist, and you can’t afford to staff one full-time.

That’s the hole the SMB has to fill, and the answer is making the Crowd’s work visible without making it punitive.

What actually captures learning at SMB scale

Four pieces, in rising order of effort.

1. A shared prompt repo, owned by one person who’s paid to own it

A Notion page, a Google Doc, a Markdown file in a Dropbox folder. The format doesn’t matter much. One person owns it, and “owning it” means they are paid for it. A small monthly stipend, or an explicit slice of their job description that nobody else is allowed to take from them. Every prompt that gets used more than once goes in the repo with a one-line description of what it’s for and a copy of the input it expects.

This sounds trivial. Most firms still don’t do it. The reason is that nobody owns it. With no owner the repo dies in week three. With an owner who is paid €100 a month to keep it current, it doesn’t.

The Karpathy idea-files pattern is the same shape, scaled down: a wiki the LLM reads from and writes to. The format is less important than the rule that everything goes in.

2. A 15-minute Friday review

Mandatory, with a calendar invite that has been on every staffer’s calendar since week one. Each person says one of two things: “this prompt worked better this week” or “this prompt stopped working or produced something weird.” The owner takes notes, the notes go in the repo, and the meeting is done by 4:15.

If you run a 12-person service firm and you can’t give 15 minutes a week to this, the firm-level capability stays at zero indefinitely. That is the trade.

This is the olsondv fix. If the firm wants the bookkeeper to hand over the 14-step VAT prompt, the firm pays for it. A €500 to €2,000 one-time bonus per shared workflow that the firm actually puts into general use, paid out three months after adoption so there is time to confirm it works. The cost is real. The cost of running six private notebooks for two more years is much higher.

The skeptic’s response: people should share for the team. Maybe in some firms. In every firm I’ve seen, until there is an explicit incentive, the rate of sharing is roughly zero. olsondv is right.

4. An off-boarding question nobody asks

When a staffer gives notice, the standard exit-interview script does not include “what prompts and AI workflows are you taking with you that the firm needs to recreate.” It should. A 30-minute session in the last week, owner asks the staffer to walk through every recurring task they use AI for, every prompt they have saved, every model they prefer for which job. Recorded, transcribed, summarized into the repo.

This costs almost nothing. It catches the part of the employee-knowledge-loss problem that the standard exit interview misses by a mile, because the standard exit interview was written before AI mattered.

Why service businesses get hit hardest

For a dental practice, an accounting firm, or a law firm, the argument is sharper than for a software company. A software firm can lose a brilliant engineer and the code stays in the repo. A 12-person accounting firm that loses the bookkeeper who built the VAT process loses the process itself, because there is no repo and the model only knows what was prompted into it last week.

I wrote separately about what happens when a small business loses a key employee. The AI version of that problem is worse, because it is invisible. The owner doesn’t know what the bookkeeper knew because the bookkeeper didn’t write it down. She wrote a prompt, the prompt is on her laptop, she quits Friday, and the firm’s “AI capability” just dropped 17%. The owner finds out three months later when client emails start sounding off.

The deeper version of this is that the prompt is the product. The capability of the firm in 2026 is the capability of its prompts. A firm with no shared repo, no review, and no incentive to share has no AI capability at all, regardless of what the subscription bill says.

What I do with clients

When I work with a 5–30 person service business on AI, the first thing I do is not pick a tool. It is an audit. Who’s using what. Which prompts live in private notebooks. Which workflows touch clients. What walks out the door if any one person quits Friday.

Then we set up the four pieces above. A repo. A Friday review owned by the firm. An incentive that points at sharing instead of hoarding. An off-boarding capture. The tool selection comes after, because the tool is downstream of whether the capture machinery exists. Switching from ChatGPT to Claude or Gemini changes nothing about firm-level learning if the prompts still live on six laptops nobody else can see.

The firms that come to me 12 months in usually have the same story. The subscription is paid, three people are power users, eight are occasional users, and one person already quit and took a workflow nobody can reproduce. They are paying for AI and getting Glaser’s “worst possible version of adoption.”

Quick FAQ

Is this only a problem for service firms? No, but it bites service firms harder. The work product is judgment-shaped, the prompts are doing the judgment, and there is no compiled artifact left behind when the staffer leaves. Manufacturing or e-commerce can sometimes coast on tooling alone. A 12-person law firm cannot.

My team is six people. Do I really need a repo and a review? At six people, you need it more, not less. The disruption from one person leaving is bigger at that size, not smaller.

What if my staff refuses to share? They will, until the incentive changes. olsondv is right on Hacker News and he is right in your firm too. Pay for the share.

Should I just ban personal AI accounts at work? No. The Microsoft data says 78% of users bring their own. A ban means 78% of your staff lie about it instead of stop doing it. The better move is making the firm’s account good enough that the personal accounts become unnecessary, and paying for what your staff figured out on their own.

How much does this cost? The repo and the review are basically free. The sharing incentive depends on size: figure on €1,000–€5,000 a year in bonuses for a 10-person firm. That is a small fraction of the AI subscription bill.

What’s the single first move? Pick the owner of the repo today. One name. Tell them they own it. Pay them €100 a month for it. Everything else gets easier once that role exists.

A second opinion

I do free 30-minute discovery calls. If you run a 5–30 person service business and you suspect your team is using AI without the firm capturing any of the upside, send me what you are seeing. We will look at where the capability is sitting (in private notebooks, on personal accounts, in one staffer’s head) and what it would take to move it into the firm. Sometimes the answer is “do these four things yourself, you don’t need me.” I’ll tell you that.

The number of European SMBs paying for AI and capturing none of it is, in my experience, the largest single source of wasted AI spend in this market. The gap between staff usage and firm capability doesn’t close on its own.