# The Curriculum You Throw Away

> Your people spend hours a week correcting AI output, and every fix teaches the organisation something it then forgets by Monday.

- Source: https://www.aicapability.academy/insights/the-curriculum-you-throw-away
- Author: Marco van Hout, AI Capability Academy (Demadera)
- Published: 2026-10-09
- Type: article

Somewhere in your organisation right now, someone is fixing a number an AI tool got wrong. They will fix it well, because they know what the number should be. Then they will hit send, and the lesson goes with it.

We have spent two years teaching people to prompt. We have spent almost nothing on what happens after the output arrives. That is where the work is, and it turns out to be where the knowledge is too.

## Four hours a week, every week

Templafy's [Business-Ready AI Gap 2026](https://cfotech.co.uk/story/uk-workers-lose-time-editing-ai-business-documents) report, based on a Sapio Research survey of 2,000 knowledge workers in the UK and US, puts a figure on the after-work. 95% still edit AI-generated business documents before use. On average they spend almost four hours a week reviewing, correcting and validating that output. 48% say the review required can leave little actual time saved. In the UK, 18% spend six or more hours a week on it, 42% say their AI tools lack access to approved company content, and 40% name factual accuracy, numbers and claims as a recurring problem. Co-founder Christian Lund put it plainly: ["AI may be fast to ask, but it is still slow to finish."](https://bernama.com/en/press/news.php?id=2614432)

Templafy sells document software, so it benefits from the gap looking wide. Read the framing with that in mind. The editing numbers are the point, and they match what most of us see on our own screens.

Four hours is a tenth of a working week. Across a department, you have quietly hired a small team whose only job is to tell the machine what good looks like here. Now look at what you keep of that work. Nothing. The fix lives in one person's draft. The reasoning lives in their head. On Monday the tool makes the same mistake for the colleague at the next desk, who fixes it again, from scratch.

## The most valuable data you produce

Christian Catalini, a research scientist at MIT, argues in [Harvard Business Review](https://hbr.org/2026/10/ai-is-making-verification-the-bottleneck-for-companies) that this is where the value of an organisation now sits. Hierarchy, he writes, does more than move information around. Its job is to verify it: to decide what information means, what deserves attention and which outputs can be trusted. He opens with two cases. A controller who caught a revenue figure that had counted deferred bookings. A security engineer who blocked a release that passed every test but would have given an AI agent write access to the production database. In both cases, a person caught it.

Catalini calls these overrides, escalations and reversals **"the most valuable data your firm produces"**. He adds, carefully, that firms which record such corrections could, in theory, train AI on them. Note the "in theory". He is describing an asset, and most organisations do not know they are throwing it away. HBR's Tom Stackpole relayed the same point in [The Insider](https://hbr.org/2026/10/the-ai-risk-your-company-is-overlooking): once AI makes execution cheap, verification becomes the hard part.

## Where your standards actually live

Put the two side by side. The survey says your people spend hours every week verifying. Catalini says those verifications are the most valuable thing they produce. And the way we have set up AI at work makes sure almost none of it survives the afternoon.

Every correction is a small statement about this organisation. "In our accounts, deferred bookings are not revenue." "We never put a delivery date in writing for this client." "That clause goes in every contract, whatever the template says." None of that is prompting skill. It is institutional judgement, and until now it has been passed on slowly, by sitting next to someone who has it.

This is what I mean by capability building at the organisational altitude. A training programme raises the floor for individuals. It cannot capture what the organisation itself knows, because that knowledge only shows itself at the moment someone says "no, that is wrong here". If you want the tool to improve, the new colleague to improve, and your written standards to be more than a PDF nobody opens, the record of corrections is the raw material. Treat it as waste and you are paying for a curriculum and keeping none of it.

## One thing to do this week

You do not need a budget line. You need a shared document and twenty minutes on a Friday.

- Start a correction log for one team. One line each time someone fixes AI output in a way that matters: what it got wrong, what right looks like here, how they spotted it. Thirty seconds, no formatting.

- Run it for two weeks. Skip the cosmetic edits. Log the fixes where sending the unedited version would have cost something.

- On the second Friday, read the log together. Twenty minutes. Look for what recurs.

- Pick the three corrections that come up most. Write each as one plain sentence and put it in two places: the team's written standard, and the instructions the tool gets before it starts.

Then run it again. The log gets shorter, because the tool stops making the mistakes you have told it about, and the new colleague reads in a page what used to take a year to absorb. That is what the four hours were always worth. You were just not keeping the receipt.

## Sources

- [Christian Catalini, "AI Is Making Verification the Bottleneck for Companies", Harvard Business Review](https://hbr.org/2026/10/ai-is-making-verification-the-bottleneck-for-companies) · 2026-10-02
- [Templafy / Sapio Research, "The Business-Ready AI Gap 2026" (2,000 UK & US knowledge workers), via CFOtech](https://cfotech.co.uk/story/uk-workers-lose-time-editing-ai-business-documents) · 2026-10-02
- [Templafy press release, "Templafy Report Finds 95% of Knowledge Workers Edit AI-Generated Documents Before They're Business-Ready"](https://bernama.com/en/press/news.php?id=2614432) · 2026-10
- [Tom Stackpole, "The AI Risk Your Company Is Overlooking", HBR The Insider](https://hbr.org/2026/10/the-ai-risk-your-company-is-overlooking) · 2026-10-06