# Shadow-AI visibility checklist

> Check it against your practice's own facts and the cited sources before you use it, with a qualified adviser where the work needs one.

From Cast & Rule: https://www.castandrule.co.uk/resources/shadow-ai-checklist

"Shadow AI" is AI use a practice hasn't approved, doesn't know about, or hasn't thought through. Not because staff are being careless, but because nobody has asked. This checklist is the asking.

## Discover

- [ ] Ask every team, not just IT, which AI tools they already use for any part of client work.
- [ ] Check browser extensions and mobile apps, not only desktop software. A lot of shadow AI arrives this way.
- [ ] Look at what free-tier AI tools staff can reach without a company card or sign-off.
- [ ] Ask new starters which AI tools they used at their last job, and assume some habits came with them.

## Assess

- [ ] For each tool you find, work out whether confidential client information has gone into it, even once.
- [ ] Check each tool's terms for whether it trains a model on the data it is given.
- [ ] Identify where each tool processes and stores data, and for how long.
- [ ] Note which tools have no written answer to the questions above. That's a finding in itself.

## Control

- [ ] Decide, tool by tool, whether it's approved, restricted to non-confidential use, or blocked.
- [ ] Write the decision into the practice's AI-use policy, not into an email that will be forgotten.
- [ ] Set out clearly what must never be typed into an unapproved tool, in plain words.
- [ ] Give staff an approved alternative for the task they were using the shadow tool for. A ban with no alternative just pushes the behaviour further underground.

## Evidence

- [ ] Keep a dated record of the exercise itself: what was asked, of whom, and when.
- [ ] Keep the approved-tool list up to date, and date it at each review.
- [ ] Record any incident where confidential information went somewhere it shouldn't have, and what was done about it.
- [ ] Repeat the whole exercise at a set interval, not only after something goes wrong.

## Train

- [ ] Walk staff through the practice's AI-use policy in plain terms, not as a document to sign unread.
- [ ] Explain why the rule exists (confidentiality, not distrust of staff), so it's followed rather than worked around.
- [ ] Make it easy and blame-free to report a mistake, so problems surface while they're still small.
- [ ] Repeat the training when the policy changes, not only once at induction.

## What the evidence shows, and its limits

No checked survey of UK accountancy measures how often staff paste client data into public AI tools. The evidence below is real, but wider than accountancy. Read the limits column before quoting a figure.

- PagerDuty / Wakefield, 2026 Shadow AI Survey: 34% of respondents (1,250 office professionals at organisations with $500m+ revenue, UK n=250, fieldwork April 2026) reported entering customer data or information into public AI tools; 66% used tools they believed were not permitted. Limit: An international survey, paid for by a supplier, of office professionals in general, not UK accountancy. "Customer data" isn't the same as accountancy client data. (https://www.pagerduty.com/newsroom/shadow-ai-workplace-survey-2026/)
- Salesforce / YouGov, Generative AI workplace survey: 89% of UK respondents (of 14,000+ across 14 countries, October 2023) reported no clear generative-AI guidelines at work; 84% reported no formal AI training. Limit: Shows a gap in rules and training, not a measured rate of client-data copying. (https://www.salesforce.com/uk/news/stories/ai-at-work-research/?bc=OTH)
- ICAEW Practice Assurance Monitoring Report 2025: Of the larger firms ICAEW held AI discussions with during 2024 monitoring, AI policies commonly covered confidential data and access controls, but only 20% monitored staff access to AI tools and 22% restricted open-source AI tools. Limit: Covers the larger firms discussed during monitoring visits (ICAEW doesn't give the number), and describes firms' controls rather than staff behaviour. (https://www.icaew.com/-/media/corporate/files/regulations/practice-assurance/practice-assurance-monitoring-2025.ashx)
- ICAS study at a mid-tier UK firm: 52% of respondents in a named UK firm case study (published 17 March 2026) raised client-data privacy and confidentiality concerns about AI. Limit: A study of one firm, not a measurement of the whole sector. (https://www.icas.com/news-insights-events/news/press-release/ai-can-t-replace-human-judgement-in-accounting-icas-study-shows)

A working checklist you can run today, built from the evidence and sources on this page.