The Receipts Room
You work the early shift at a leisure centre. To get paid, you look into a camera at the door, or press your finger to a scanner. Nobody asked whether you’d rather use a card. And according to the regulator, a written procedure said that refusing to enrol could lead to disciplinary action.
That was the working life of more than 2,000 people at 38 leisure facilities in the UK. In December 2019, an employee of the UK Information Commissioner’s Office noticed facial recognition technology in use at a leisure centre, and the regulator opened an investigation. Four years later, in February 2024, it issued enforcement notices to Serco Leisure, Serco Jersey and seven associated community leisure trusts. The ICO found that they had unlawfully processed the biometric data of those employees, using facial recognition and, at two sites, fingerprint scanning to track attendance and pay. They had not shown why less intrusive options such as ID cards or fobs would not do, and staff had not been proactively offered an alternative (Information Commissioner’s Office [ICO], 2024a, 2024b).
Here is our proposition: most organizations can find their own version of this problem in a month, on their own schedule, and almost none do until a regulator sets the clock. The organization saves effort by waiting. The staff and customers pay for it in the meantime.
The receipt: a remedy with a deadline
What makes the case so useful is the remedy. It came with a clock. Within three months of the notice, Serco had to stop using biometric data for attendance checks at every site, roll it out nowhere else, and destroy the data it was not legally required to keep, including telling its software supplier to delete data held on its behalf (ICO, 2024b, Annex 1). Serco Leisure said it would “fully comply with the enforcement notice,” according to the Press Association (2024).
Now read the notice the way an auditor would. Purpose: attendance and pay. Data: faces and fingerprints. Supplier: an outside software company acting on Serco’s behalf. And the uncomfortable truth: a 2020 internal assessment said “an opt out would be unsuitable,” and although it was updated in 2021 to say an alternative would be considered, the operating procedure still said refusing to enrol could lead to disciplinary action (ICO, 2024b, paras. 22–26).
The written position and the working practice did not match. That gap is the whole story. It sat there for years, and it took an outsider to write it down.
The fix you can run before anyone makes you
A regulator’s deadline is one way to force a bounded fix. Chapter 10 of Operational AI Readiness, the free field manual by Viktoria Bakos, argues that you should not wait for one. (This is the next stop in our series From the Field Manual, where we introduce the book one chapter at a time.)
Chapter 10 starts by naming a trap: believing that governing AI seriously means rewriting every policy and aligning every stakeholder first. So nothing happens. The book’s answer is a bounded sprint on one consequential workflow. The point is not to “become compliant” in a month, a claim the book calls foolish. It is to move one workflow “from vague confidence to operational visibility” (Bakos, 2026, p. 68).
The four weeks are simple to state:
- Week 1, inventory and risk mapping. Write down what the workflow is for, what it must not be used for, who owns it, what data and sources feed it, what tools it connects to, and which outcomes it can influence. The book warns that this week tends to produce “the first uncomfortable truths”: undocumented connectors, unclear boundaries, and disagreement about how much the system really influences decisions (pp. 68–69).
- Week 2, oversight and approval design. Decide where human review belongs, which outputs need named approval, who can override or escalate, and what happens when the system hits ambiguity. The deliverable is a design that says exactly where review occurs and what authority the reviewer holds, not a promise that humans remain in the loop (p. 69).
- Week 3, the evidence spine. Define what must be traceable for a consequential case: which system and version ran, where its context came from, any flags raised, what the reviewer decided, any override, what happened downstream, and when. The book is clear this does not require a giant observability overhaul, only a deliberately scoped record (pp. 69–70).
- Week 4, validation and rehearsal. Probe for misuse, run regression cases, drill an override, and hold at least one tabletop exercise where the workflow misbehaves and the team shows how the controls respond. The aim is not “to stage a dramatic apocalypse” but to learn whether the workflow is governable when something goes wrong (p. 70).
Put the Serco notice next to Week 1 and it reads like an inventory someone else had to do for them. Finding that kind of gap early, on your own schedule, is exactly what Week 1 is for.
Key idea: Acceptance tests for governance
An acceptance test is a check, agreed in advance, that tells you whether a piece of work is actually finished. Chapter 10 applies the idea to oversight: by the end of Week 3, for example, one real trace should show you can reconstruct a consequential case. For an ordinary person, this is the difference between an organization saying “we reviewed our AI” and an organization that can show you what “reviewed” meant, case by case.
The complication: this wasn’t really an AI story
The fairest objection to using this case is a good one. Serco’s system was biometric clock-in technology, not the kind of AI assistant the book mostly discusses. Serco said the technology was introduced after external legal advice that it was permitted, according to the Press Association (2024). And the case is about whether biometric processing was lawful and fair, not about whether the system misidentified anyone.
All true. We are not claiming a sprint would have changed the legal outcome. Our point is narrower: the failure the ICO described, a gap between the written policy and what staff actually experienced, is the most common failure in any automated system that sits between people and their pay. It doesn’t need a large language model to happen.
The verdict: five moves that earn their place
- Pick the workflow you would least like to explain. The sprint works on one live, consequential workflow, not the whole estate. Start where a system touches money, pay, housing, jobs, health or legal rights.
- Write down the alternative route on Day 1. Week 1 asks for intended purpose and excluded uses. Add one more line: what happens for the person who cannot or will not use the system? If nobody writes that route down, in practice there is not one.
- Replace “done” with acceptance tests. The book sets one per week: an undisputed inventory by Week 1, explicit review and approval rules by Week 2, one real reconstructable trace by Week 3, and a walked-through failure scenario by Week 4 (pp. 70–71).
- Rehearse the bad day before it arrives. A tabletop exercise is cheap. Discovering under a regulator’s deadline that nobody knows how to switch the system off, or what data your supplier holds, is not.
- Treat a messy sprint as a result. The book says plainly that a sprint “is not a shortcut around judgment.” If the team cannot agree on what the workflow does, that is diagnosis, not failure. The real failure is declaring success because a sprint happened. “The point of the sprint is not activity. It is operational legibility” (p. 71).
Most of us will never sit in a sprint planning meeting. But we live with the results, often at work. A month of disciplined work will not make a system perfect. It can make it answerable.
What we still don’t know
The ICO’s notice sets out the regulator’s findings. We have not found a public ICO record confirming that every step was completed by the deadline, and we do not claim it was or was not. Which leaves the question worth carrying into your own workplace: if the regulator walked in tomorrow, would your written policy and your working practice tell the same story?
What you can do this week
- If your employer asks you to clock in with your face or fingerprint, ask what alternative exists, who supplies the system, and how long your data is kept.
- If an automated system at work affects your pay or hours, ask who reviews its records and how you can correct a mistake.
- If you run AI or automated systems at work, put four dates in the calendar this week, one per sprint week, each with its acceptance test written next to it.
- If you lead an organization, ask for one reconstructed case at the end of the month instead of a status update.
Go further. Chapter 10 of Operational AI Readiness, “The 30-Day Readiness Sprint,” walks through each week with worked examples, and the book’s closing pages connect it to Chapter 9, Appendix B, Appendix D and Appendix E. Read it free. Then make Day 1 count: take the free readiness assessment on the workflow you picked. It takes about ten minutes, names your weakest domain, and gives you three next moves for Week 1. It is a self-assessment, not an audit or certification, and nothing you answer is stored.
References
Bakos, V. (2026). Operational AI readiness: A practitioner guide to governing AI that is already live (Free Standard ed.). FairByDesign. https://fairbydesign.org/operational-ai-readiness/
FairByDesign. (2026). Operational AI readiness assessment. https://fairbydesign.org/readiness-assessment/
Information Commissioner’s Office. (2024a, February 23). ICO orders Serco Leisure to stop using facial recognition technology to monitor attendance of leisure centre employees. https://ico.org.uk/about-the-ico/media-centre/news-and-blogs/2024/02/ico-orders-serco-leisure-to-stop-using-facial-recognition-technology/
Information Commissioner’s Office. (2024b, February 19). Enforcement notice: Serco Leisure Operating Limited. https://ico.org.uk/media2/woalogoc/20240219-serco-leisure-operating-limited-en.pdf
Press Association. (2024, February 23). Serco Leisure ordered to stop using facial recognition tech on workers. LBC. https://www.lbc.co.uk/article/226204fffa5b44af9d692369d3445b9a-5Hjcf9c_2/
FairByDesign is reader-funded. If this helped, get the free Mercury Brief or see ways to support. Support never buys editorial control.


Leave a Reply
You must be logged in to post a comment.