← Back to the exerciseThe full worked example: Write a board ESG page you can defend line by line
campo.
Field guide / L&D / UC-001 · SC-11
The result

"312 trained, every job guaranteed"? 156 finished, and jobs are Unknown.

A Head of Apprenticeships at a 20,000-person rail equipment maker checks a draft ESG headline against her own data and sends the board committee a page it can rely on, with any AI chat her company allows.

The resultWhat Hannah sends the ESG committee. Fictional company.
TAMSIN RAIL SYSTEMS · BOARD ESG COMMITTEE · DEC 3
Apprenticeships in 2026
For: the ESG committee, via Claire Dunmore
From: Hannah Okafor

THE HEADLINE

In 2026, 156 apprentices completed their three-year programme: 65% of the 240 who started in 2023. 84 left without completing, and we're finding out why. · Input 2

THE PIPELINE

2023 cohort156 of 240 completed
65%; 84 left
2024 cohort271 of 290 in training
Finishes 2027
2025 cohort305 of 312 in training
Finishes 2028
Kept onUnknown
Not tracked after month one

PUBLIC FUNDING, 2026

UK £2.1M (levy; includes college fees). Germany €0.6M. Canada CA$0.4M. Shown separately: there's no single total until finance converts at a stated rate. · Input 3

NOT CLAIMED THIS YEAR

A job guarantee. Germany offered roles to about 90% of finishers; UK and Canada figures to follow. · Input 5
A retention rate. We'll track finishers for 12 months from 2027. · Input 5
Caught on review
1
"Trained 312" is the 2025 cohort's starts; 156 completed in 2026
2
Finance's 4.0M counts £0.9M twice and adds three currencies
3
"Every apprentice is guaranteed a job" isn't supported by any input
UC-001 · SC-11 · Rev A
FOR
Head of Apprenticeships & Early Careers
SECTOR
Rail & transit equipment
COMPANY
20,000 people · 9 sites · UK, Germany, Canada
TASK
Write the apprenticeship page for the board's ESG committee
WRITTEN FOR
ChatGPT · Claude · Gemini · Copilot
AI TEST
Claude, Nov 13, 2026 · 7 of 7 found
Step 0 · The mess

What landed on Hannah's desk

A proud draft headline, a cohort export, a funding total, a college's pass rate and a short HR note. Each looks fine alone. Read together, the headline doesn't hold.

EMAIL · SUSTAINABILITY[1]
ESG committee page
“In 2026 we trained 312 apprentices, and every apprentice is guaranteed a job.”
1
Two claims to test.
SYSTEM EXPORT[2]
Cohorts 2023 to 2025
2023: 240 started, 156 completed. 2025: 312 started.
2
Where the 312 came from.
FINANCE SUMMARY[3]
Public funding
£2.1M + £0.9M + €0.6M + CA$0.4M = “4.0M”. College fees paid from the levy.
3
Counted twice; three currencies.
COLLEGE REPORT[4]
Certification exams
“98% pass rate!” 41 of 42 who sat. Cohort: 60.
4
Right number
TEAMS · HR GERMANY[5]
Retention
“We don't track apprentices after their first month.” Germany offered roles to about 90%.
5
An offer isn't a guarantee.
You're not alone
65.4%
was England's national apprenticeship achievement rate in 2024/25, so about a third of apprentices didn't achieve.
UK Secretary of State for Work and Pensions, letter on the apprenticeship achievement rate, 26 March 2026. Source
Steps 1–3 · How it works

Three steps, one chat

Here the AI checks a page that will go into public reporting. The prompt asks for numbers you can defend.

1
Open a new chat

Use the AI tool your company approves: ChatGPT, Claude, Gemini or Copilot.

2
Paste the prompt and the reports

Copy the prompt and the five fictional inputs from the appendix into one message.

3
Cut it to one page

Check each number against its input yourself. Then keep what the committee needs and decide what you'll commit to.

The prompt, part by part
Copy-ready version in the appendix

I'm Head of Apprenticeships at a rail equipment manufacturer with about 20,000 people in the UK, Germany and Canada. I'm writing the one-page apprenticeship update for our board's ESG committee on December 3. Below are our sustainability lead's email with a draft headline, our cohort data, finance's funding summary, a college partner's report and a note from HR.

1
Who you are, who will rely on the page, and what you were sent.

Please give me:

  1. Is the draft headline accurate? Answer yes or no, then give a corrected headline.
  2. The apprentice pipeline, from started to completed to kept on, with the input each number comes from.
  3. Public funding, shown correctly.
  4. What we can't claim yet, and what I'd need to claim it.
2
A yes or no on the headline, the pipeline, the funding and what you can't claim. The yes or no stops the AI hedging.

A board committee will rely on this, so every number must be one I can defend. Check what each number counts, whether totals add things twice or mix currencies, and whether a rate uses the right group. If something is unknown, say "Unknown". Don't add facts that aren't in the inputs.

3
The heavy lifting: every number defensible, no double counting or mixed currencies, the right group for each rate. It found all seven problems.
Where to paste
Every tool has one message box. Paste the prompt, add a blank line, paste the reports, send. Exports can be attached as files if your tool allows it.
Safe to try
Practise with the fictional data first. Real apprentice records are personal data: use only your company's approved tool.
The AI's first answer

What Claude actually wrote

Exactly as it came back, not edited. Read it the way Hannah did: check each point against the source it cites.

UNEDITED
Claude · fresh session, no other context · Nov 13, 2026 · one run

1. Is the draft headline accurate?

No. Both halves of it are wrong.

  • "We trained 312 apprentices in 2026." 312 is how many apprentices started in the 2025 cohort [2]. It isn't how many we trained in 2026, and it isn't the number in training today either: 305 of those 312 are still in training [2]. The export shows three cohorts active in 2026. The 2023 cohort finished during the year with 156 completing, and 271 (2024 cohort) plus 305 (2025 cohort) are still in training [2]. Whether a 2026 cohort started is Unknown, because the export has no 2026 row. found
  • "Every apprentice is guaranteed a job." None of the inputs supports this. Germany offered permanent roles to "about 90%" of this year's finishers. That figure is approximate, and an offer is not a guarantee. The UK and Canada sent no figures, and HR doesn't track apprentices after their first month as employees [5]. found

Corrected headline (every number traceable to [2]): "In 2026, 156 apprentices completed their three-year programme, and 576 more are currently in training." A: completion rate

(576 = 271 + 305. Please confirm the date of the export so "currently" can be given as "as of [date]".)

2. Apprentice pipeline: started, completed, kept on

Cohort Started Completed Still in training Left without completing Kept on Source
2023 (finished 2026) 240 156 0 84 (240 − 156 − 0) Unknown [2]; kept on: [5]
2024 (in year 2 of 3) 290 0 (not due yet) 271 19 (290 − 271) n/a [2]
2025 (in year 1 of 3) 312 0 (not due yet) 305 7 (312 − 305) n/a [2]

Rates you can defend:

  • 2023 cohort completion rate: 156 / 240 = 65%. This is the only cohort that has finished, so it is the only real completion rate. Don't work out completion rates for 2024 or 2025, because they haven't finished yet. found
  • Still in training: 2024 cohort 271 / 290 = 93%, 2025 cohort 305 / 312 = 98%. Label these "still in training", not "completion".
  • Kept on: Unknown. Germany offered permanent roles to "about 90%" of its finishers [5]. That measures offers, not acceptances or retention. It covers Germany only, and we don't know how many of the 156 finishers were in Germany, so it can't be turned into a number of people or a company-wide rate. UK and Canada: no data [5]. Retention beyond the first month isn't tracked [5]. found

About Northgate's "98% pass rate" [4]: this uses the wrong group. 98% is 41 of the 42 who sat the exam. Northgate's cohort with us is 60 apprentices, so 41 / 60 = 68% of that cohort passed. For the other 18 the exam outcome is Unknown: they didn't sit it, or the reason isn't given. This is also one college's certification exam in Canada. It isn't a programme completion rate. The inputs also don't show which of our cohorts the 60 belong to, so don't combine it with the figures in [2] or use it as a company-wide figure. found

3. Public funding, 2026, shown correctly

Country Public funding 2026 Source
UK £2.1M apprenticeship levy funds drawn (this includes the £0.9M of college partner fees, which are paid from the levy account) [3] and its footnote
Germany €0.6M state training grant [3]
Canada CA$0.4M apprenticeship grant [3]

Two errors in finance's "Total public funding: 4.0M":

  1. It counts money twice. The £0.9M of college fees is paid out of the £2.1M levy draw, so adding it again overstates the UK figure by £0.9M. found
  2. It adds three currencies together. It adds pounds, euros and Canadian dollars as if they were one unit, and it doesn't say which currency the 4.0M is in. found

A single total is Unknown from these inputs. To give one, finance would need to convert the euros and Canadian dollars at a stated exchange rate and date, using £2.1M (not £3.0M) for the UK. Until then, show the three figures separately as above.

One point to check with finance before the board sees this: whether the board should read UK levy draws as "public funding" in the same sense as the German and Canadian grants. The inputs only label them that way and give no further detail.

4. What we can't claim yet, and what you'd need to claim it B: two lines

Claim Status What you'd need
"Every apprentice is guaranteed a job" Not supported A written company policy guaranteeing roles in all three countries. The inputs only show about 90% of German finishers offered roles.
How many apprentices were kept on, or a retention rate Unknown Offers made and accepted in the UK, Germany and Canada for the 156 finishers, plus tracking beyond the first month as employees (not done today [5]).
A company-wide job-offer rate Unknown Finisher counts by country and offer counts from the UK and Canada. Germany's "about 90%" needs the exact count behind it.
Number of apprentices trained in 2026 Unknown as a single figure The export date, whether a 2026 cohort started, and when in 2026 the 2023 cohort finished. Use completed (156) and in training (576) instead.
A "98% pass rate" Misleading The fair figure is 41 of 60 (68%) for Northgate's cohort. You'd also need to know why 18 didn't sit the exam and which of our cohorts they belong to.
Completion rates for the 2024 and 2025 cohorts Too early Wait until those cohorts finish (2027 and 2028, given the three-year programme).
Why 84 of the 2023 cohort didn't complete Unknown Leaver reasons from the apprenticeship system or HR. Expect the committee to ask about this, since it is about a third of the cohort. C: say it first
A single public funding total Unknown Exchange rates and a date from finance, and removing the double-counted £0.9M.
Country split of the cohorts Unknown The export broken down by country. Without it, neither the German offer rate nor Northgate's figures can be connected to the pipeline.

Before Nov 26: ask Claire to drop the job-guarantee line, ask finance for a corrected funding table, and ask HR in the UK and Canada for finisher and offer counts.

foundcaught a planted problemA–CHannah changes it; see the next section
The human part

What Hannah changed, and why

Claude found every planted problem. Hannah's edits are about what a board committee should see and what she's willing to commit to.

What Claude found: 7 of 7 planted problems
One run
✓“Trained 312” is starts
✓65% of the 2023 cohort completed
✓College fees counted twice
✓Three currencies added together
✓98% of those who sat the exam
✓Retention not tracked
✓Job guarantee unsupported
Three edits, all judgment calls
A
CLAUDE WROTE
“In 2026, 156 apprentices completed their three-year programme, and 576 more are currently in training.”
HANNAH'S VERSION
“156 completed: 65% of the 240 who started in 2023. 84 left without completing.”
WHY · AUDIENCE
An ESG committee should see the completion rate, not only the bigger number. 576 also depends on an export date nobody has confirmed.
B
CLAUDE WROTE
A nine-row table of claims that can't be made yet.
HANNAH'S VERSION
Two lines on the page: no job guarantee, no retention rate. The rest goes to Claire as a to-do list.
WHY · LENGTH
The committee needs to know what isn't claimed. The data requests are Hannah's work, not theirs.
C
CLAUDE WROTE
84 leavers appear near the end: “Expect the committee to ask about this.”
HANNAH'S VERSION
In the headline, with "we're finding out why".
WHY · PRIORITY
If a board member spots a third of a cohort missing, it should already be on the page. Saying it first builds trust.
What only you can do
Decide what the board should see first, say the uncomfortable number yourself, and commit to tracking retention. The AI checked the numbers. You own the page.
Your turn

Make it yours

Use it for any page a board, auditor or regulator will rely on: ESG reports, annual reviews, public commitments.

Your prompt: fill in the brackets

I'm the [your role] at a [industry] company with about [number] people in [countries]. I'm writing [the page] for [the board committee or audience] on [date]. Below are [the draft text and the data behind it].

Please give me: 1. Is the draft headline accurate? Answer yes or no, then give a corrected headline. 2. [The main numbers], with the input each number comes from. 3. [Any money figures], shown correctly. 4. What we can't claim yet, and what I'd need to claim it.

A board committee will rely on this, so every number must be one I can defend. Check what each number counts, whether totals add things twice or mix currencies, and whether a rate uses the right group. If something is unknown, say "Unknown". Don't add facts that aren't in the inputs.

Five checks before you send
1
Each number in the headline matches the input it came from.
2
Totals don't count the same money twice or mix currencies.
3
Every rate says which group it's out of.
4
Claims like "guaranteed" or "every" have proof, or are gone.
5
The number a board member would ask about is already on the page.
Make it repeatable (optional)
Save the prompt once and reuse it for every board or ESG page: a custom GPT in ChatGPT, a project in Claude, a Gem in Gemini, or an agent in Copilot. Some companies switch these off; copy and paste always works.
With real data
Apprentice records are personal data, and draft board papers are confidential. Use only the AI tool your company approves.
What this example shows
One run of this prompt in Claude found all seven problems planted in fictional reports and invented nothing.
What it doesn't show
That every tool, setting or run will do the same, or how much time it saves you. Tamsin Rail Systems and its people are fictional.
Appendix

Copy-ready: the prompt and the five inputs

Copy the prompt, then all five inputs, into one message in a new chat. Everything here is fictional and safe to paste.

The prompt

I'm Head of Apprenticeships at a rail equipment manufacturer with about 20,000 people in the UK, Germany and Canada. I'm writing the one-page apprenticeship update for our board's ESG committee on December 3. Below are our sustainability lead's email with a draft headline, our cohort data, finance's funding summary, a college partner's report and a note from HR.

Please give me:

  1. Is the draft headline accurate? Answer yes or no, then give a corrected headline.
  2. The apprentice pipeline, from started to completed to kept on, with the input each number comes from.
  3. Public funding, shown correctly.
  4. What we can't claim yet, and what I'd need to claim it.

A board committee will rely on this, so every number must be one I can defend. Check what each number counts, whether totals add things twice or mix currencies, and whether a rate uses the right group. If something is unknown, say "Unknown". Don't add facts that aren't in the inputs.

[1] Email from Claire Dunmore, Head of Sustainability

Subject: Apprentice page for the ESG committee, Dec 3

Hannah, I need your one-page apprenticeship update by Nov 26. Comms has drafted the headline: "In 2026 we trained 312 apprentices, and every apprentice is guaranteed a job." Can you confirm the numbers and add funding?

[2] Cohort export from the apprenticeship system
Cohort (start year) Programme length Started Completed Still in training
2023 3 years 240 156 0
2024 3 years 290 0 271
2025 3 years 312 0 305

Note: the 2023 cohort finished in 2026.

[3] Finance: public funding, 2026
  • UK apprenticeship levy funds drawn: £2.1M
  • College partner fees (UK): £0.9M
  • Germany state training grant: €0.6M
  • Canada apprenticeship grant: CA$0.4M
  • Total public funding: 4.0M

Footnote: UK college partner fees are paid from the levy account.

[4] Report from Northgate College (Canada partner)

2026 certification exams: 98% pass rate! 41 of the 42 apprentices who sat the exam passed. Our cohort with you: 60 apprentices.

[5] Teams message from Jonas Becker, HR Germany

On retention: we don't track apprentices after their first month as employees. In Germany we offered permanent roles to about 90% of this year's finishers. UK and Canada didn't send figures.