exit-reasons

---
name: exit-reasons
description: Turns two years of leavers into a counted table that shows which reason, which manager, which shift and which month your people actually leave from, with the small numbers suppressed so nobody is identified. Use when you keep rehiring the same role and nobody can say why the last four went.
---

# Why people really leave you, counted, not guessed

You give me a list of everyone who has left in the last two years, however rough it is: a payroll export, a page of a diary, a rota with names crossed out, and whatever anybody remembers being said on the last shift. You get back a leavers file with one row per person, a fixed list of reason codes applied the same way to all of them, the count for each reason including an honest count of the ones nobody knows, three cuts showing whether leavers cluster by manager, by section or by shift pattern, a length of service table showing how long people last before they go, and a one page summary naming the two things to change and when you will check whether it worked.

## What it does

1. **Build the leavers list from payroll, not from memory.** Take everybody whose employment ended in the last 24 months, from the payroll export or the P45 list, not from who springs to mind. Memory drops the ones who lasted three weeks, which is the group that costs most and tells you most. One row per person, with these columns and no others: reference, start date, end date, days employed, job title, section, main shift pattern, line manager at the time they left, who ended it, primary reason code, secondary reason code, who coded it, date coded. Names go in a separate lookup that only the owner holds, and the analysis file carries the reference instead. If a field is not known, write "unknown", never a guess. This file is the whole skill: everything after it is counting.

2. **Fix the reason codes before you read a single leaving conversation.** Write a closed list of twelve or fewer codes and do not add to it halfway through, because a code invented on leaver nine cannot be applied to leavers one to eight and the count becomes meaningless. A hospitality list that works: hours or rota, pay, travel or transport, another job in the trade, left the trade, study, moved away, treatment by a manager, no progression, health, caring or childcare, ended by us, end of season, unknown. Each leaver gets one primary code and at most one secondary. Anything they actually said goes in a separate free text column that is never counted, only read. Two codes that must never be merged are "another job in the trade" and "pay", because one means a competitor is better and the other means you are cheap, and they need different answers.

3. **Code the silent ones as unknown and report that count first.** Most small businesses do not hold exit interviews, so a real leavers file starts out mostly unknown. Do not fill the gap with what you assume. Acas says "Arranging an exit interview can be useful to understand why the employee is leaving" and that "Their reasons can help inform how you recruit or retain staff", which is the case for starting them, but it is not a licence to backfill two years of them from memory. Report unknown as a number on the front page. If unknown is above half, that is the first finding and the first fix, and every other pattern in the file is a pattern in the minority you happen to know about. Say that out loud rather than presenting the known half as the whole picture.

4. **Count days employed exactly, then band it.** Calculate end date minus start date in days for every person, then band: under 14 days, 14 to 30 days, 1 to 3 months, 3 to 6 months, 6 to 12 months, 1 to 2 years, over 2 years. Count how many fall in each band and what share of all leavers that is. This single table usually answers the question before any reason code does. A pile in the first two bands is a recruitment and induction problem, people arriving to a job that is not the one advertised. A pile at 6 to 12 months is usually rota, progression or a manager. Acas gives the recruitment version of this directly: "if you find that employees regularly leave because the job is not what they expected, you could update future job adverts to be clearer about the role and responsibilities."

5. **Cut by manager, section and shift pattern, and suppress the small cells.** Three tables, each showing leavers, average days employed and the top two reason codes. Then apply the rule that keeps this document safe to put on a desk: any cell with fewer than three leavers is shown as a count only and is never written up as a finding about a named person. The Information Commissioner's Office warns, about monitoring information, that it "might potentially identify particular workers, even if the names have been removed", giving the example of a characteristic shared by relatively few workers. A table saying two of three kitchen leavers cited treatment by a manager, in a kitchen with one head chef, has named him and published a serious allegation on a count of two. Where a cell is under three, write "too few to draw a conclusion" and carry it forward to the next run instead.

6. **Cut by month left and month joined.** Count leavers by calendar month of leaving across the whole 24 months, and separately by calendar month of joining. Seasonal trade produces a genuine spike in September and January that is not a management failure, and mistaking it for one sends the owner chasing a problem that does not exist. What matters is the residue: the months where leaving is high and it is not the end of a season, and the joining cohorts that melt faster than the others. If every September intake is gone by November and every February intake is still there, the difference is induction during a busy period, not the people.

7. **Split the list into what you can change and what you cannot, and put a price on the first group.** Moved away, study, left the trade and health are real and mostly outside your control. Hours or rota, pay, treatment by a manager, no progression and travel are inside it. Count both groups and lead with the controllable one, because that is the only number that can move. The Chartered Institute of Personnel and Development says "High turnover rates can be costly in terms of recruitment, training and loss of knowledge" and that "Calculating the cost of turnover can be useful to understand the scale of the issue and supports the business case for improving retention", so put your own real figures against the controllable leavers: what the advert cost, the hours spent sifting and interviewing, and the weeks of trial shifts before the new person carried a section alone. Use your own numbers only. Do not import a cost per hire from anywhere.

8. **Write one page, choose two changes, and set the date you re-run this.** The page carries: how many left, over what period, out of how many staff; the unknown count; the length of service bands; the top three controllable reasons with their counts; the two cuts that showed something and the ones that were too small to read. Then exactly two changes, each with a named owner and a date, because an owner who picks six changes makes none of them. Finally, the date you run this again, which should be six months out, using the same codes and the same file, so the second run is a comparison rather than a fresh opinion. Keep the codes frozen between runs even if you now dislike one of them.

## Then it checks

1. Every person who left in the period appears exactly once, the count matches the payroll export it came from, and nobody has been added from memory without a payroll record.
2. Every row has a primary reason code from the fixed list, including "unknown", and no code appears that is not on the list written before coding started.
3. The unknown count appears on the front page of the summary, as a number and as a share of all leavers.
4. Days employed is calculated from the two dates in the row, not typed, and every row falls into exactly one length of service band.
5. No cell containing fewer than three leavers is written up as a finding, and no manager, section or shift is named in a conclusion that rests on fewer than three people.
6. The summary names two changes, each with an owner and a date, and states the date this analysis will be re-run using the same codes.

Any check fails: name it, redo that step once. Failed twice: say what is wrong and stop.

## Rules
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.
- Never assign a reason code from an impression. A code is a judgement, and the Information Commissioner's Office says that for records to be accurate they "should make clear that it is an opinion, and, where appropriate, whose opinion it is", which is why every row records who coded it and when. If nobody heard a reason, the code is unknown and it stays unknown.
- Never write up a finding about a named manager, section or shift on fewer than three leavers. The whole value of counting is that it replaces a rumour with a number, and a number of two is a rumour with a decimal point.
- Never put what somebody said on their last shift into a file that other staff can open. Keep the free text column with the owner, keep the analysis file on references rather than names, and delete the free text when the analysis is finished.
- Never use this file to decide anything about a person who still works for you. It is a count of people who have gone, built to change a rota or an induction, and it is not a performance record.
- This output is a working document prepared for the owner's HR adviser, solicitor or accountant to check before it is used. It does not establish why anybody left, it is not evidence in any dispute, and it is not employment law advice.

## Built from
- Acas, "Responding to an employee's resignation", https://www.acas.org.uk/resignation/responding-to-an-employees-resignation, last updated 30 June 2026, read 14 September 2026: the purpose of an exit interview, and the example of leavers regularly saying the job was not what they expected leading to a change in the advert, which is the shape of the whole skill and shaped steps 3 and 4.
- Acas, "If employees feel excluded", https://www.acas.org.uk/improving-equality-diversity-and-inclusion/if-employees-do-not-feel-included, no publication date shown on the page, read 14 September 2026: that exit interviews can reveal somebody left because they "felt excluded or held back in some way", which is why "treatment by a manager" and "no progression" are separate codes in step 2 rather than folded into one.
- Information Commissioner's Office, "Employment practices and data protection: keeping employment records - Using employment records", https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/employment-practices-and-data-protection-keeping-employment-records/using-employment-records/, no publication date shown on the page, read 14 September 2026: the warning that monitoring information can identify individuals even with names removed, which produced the fewer-than-three suppression rule in step 5.
- Information Commissioner's Office, "Employment practices and data protection: keeping employment records - Collecting and keeping employment records", https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/employment-practices-and-data-protection-keeping-employment-records/collecting-and-keeping-employment-records/, no publication date shown on the page, read 14 September 2026: that records should show where something is an opinion and whose it is, which is why every row carries who coded it, and the minimisation and erasure lines behind the rule on free text.
- Chartered Institute of Personnel and Development, "Employee retention guide", https://www.cipd.org/en/knowledge/guides/employee-retention/, dated 18 March 2025 on the page, read 14 September 2026: the definition of turnover as a proportion over a set period and the case for costing turnover, which shaped step 7. Most of this guide sits behind a membership login and only the visible summary was read.

# exit-reasons

## You are given
A folder holding whatever the owner has about people who left in the last 24 months: a payroll export (CSV or XLSX), a rota, or a typed list. Dates, job titles and names may be missing or spelled three ways. If no reason code list is supplied, use exactly these and add none: hours-or-rota, pay, travel, another-job-in-trade, left-the-trade, study, moved-away, manager-treatment, no-progression, health, caring, ended-by-us, end-of-season, unknown.

## Produce
Write these files into an `output` folder beside the input:

1. `leavers.csv` - one row per leaver, columns in this order: `ref, start_date, end_date, days_employed, job_title, section, shift_pattern, line_manager, ended_by, reason_primary, reason_secondary, coded_by, coded_date`. Dates YYYY-MM-DD. `ref` is L001 upward. No names here.
2. `names-lookup.csv` - two columns, `ref, full_name`. Nothing else.
3. `reason-counts.csv` - columns `reason_code, leavers, share_of_all, avg_days_employed`. One row per code including unknown, sorted by leavers descending.
4. `tenure-bands.csv` - columns `band, leavers, share_of_all`. Bands exactly: `under 14 days, 14-30 days, 1-3 months, 3-6 months, 6-12 months, 1-2 years, over 2 years`.
5. `cut-by-manager.csv`, `cut-by-section.csv`, `cut-by-shift.csv` - each with columns `group, leavers, avg_days_employed, top_reason, second_reason, suppressed`. Set `suppressed` to `yes` and blank both reason columns wherever `leavers` is under 3.
6. `by-month.csv` - columns `month, joiners, leavers`, one row per calendar month across the whole period, YYYY-MM.
7. `summary.md` - one printable page: total leavers, period covered, unknown count and share stated first, the tenure band table, the top three controllable reasons, which cuts were readable and which were suppressed, and a blank two-line block headed "Two changes, owner and date".

## Rules
- Never invent a leaver, a date or a reason. Anything absent is the literal string `unknown`.
- Calculate `days_employed` from the two dates. Never accept a typed figure.
- Do not create a reason code that is not on the list.
- Any group with fewer than 3 leavers is suppressed: count shown, reasons blanked, `suppressed` set to yes. Never name a manager in `summary.md` on fewer than 3.
- Put no free text about what anybody said into any output file.
- Write UTF-8, comma separated, quoting any field that contains a comma. Use British date order in `summary.md` prose (14 September 2026).
- If the input has no usable end dates, stop and say so. Never estimate them.

## Return
List every file written with its full path and row count. State the number of leavers, the number coded `unknown`, and the number of groups suppressed across the three cut files. Name any input row you could not use and why. End with the single line: `This is a working document for the owner's HR adviser to check before it is used.`