The Prompt Ledger: A Half-Day Review of AI Chat Histories in UK Data Estates
Photo: Internet Archive Book Images, No restrictions, via Wikimedia Commons
Most UK organisations have formed a view on which AI assistants staff may use. Far fewer have examined what those assistants keep. Every prompt, pasted extract, attached file and generated answer is a record. It may sit in a corporate SaaS tenant, a browser extension, a personal account or a vendor's model-training pipeline. The half-day review treats prompt history as part of the data estate, not as an IT curiosity.
Why prompt logs are not email
Email has folders, retention labels, legal holds and a long-established eDiscovery market. Chat histories are chronological, searchable and exportable, yet often invisible to records management. They mix customer data with instructions and model output. Deletion is usually a button in a user interface, not proof of erasure. Vendor access is contractual rather than architectural. The review therefore asks three questions of every tool: what is in the log, how long does it live, and who can read it.
The four-hour shape
Hour one: inventory and access. List approved assistants, embedded copilots, meeting transcribers, browser extensions and any AI features inside existing SaaS. For each, record the account type (corporate, personal or shared), whether an admin console exists, export capability, retention settings and subprocessors. Ask IT and procurement for the SaaS register. Look for shadow tools through expense claims, DNS logs and a short team survey.
Hour two: sampling. Draw a defensible sample: 50 prompts per tool per business unit, plus 20 deleted items where the console exposes them. Include attachments and generated outputs. Do not read for individual performance; read for data types. Use a simple extraction sheet: prompt ID, tool, date, department, data category, personal data yes or no, special category yes or no, customer identifier yes or no, commercial terms yes or no, retention setting, and vendor access noted.
Hour three: classification and retention test. Apply four tiers: trivial, internal, personal data, and special category or high-risk. Map each tier to an expected retention period. Then test the claim. If the interface says deleted, ask whether the vendor retains the content in backups, operational logs or model-training data. If the contract says no training, ask for the setting and the subprocessor list. If the answer is 'industry standard', mark it unverified.
Hour four: vendor questions and decisions. Produce a one-page note for the next renewal. Include an allow-list of tools, a paste rule, a retention instruction and a named owner. Decide whether prompt history is in scope for data subject access requests and retention schedules. Update the record of processing activities. Brief the AI user group.
What the sample usually shows
Common findings are mundane and material: customer names and postcodes pasted for summarisation, complaint text copied for tone rewrites, pricing schedules pasted for comparison, staff details in HR prompts, and meeting transcripts containing client names. Special category data appears more often than expected, usually in HR and health-related queries. Personal accounts appear where the corporate tool lacks a feature or a licence. The point is not to catch individuals. It is to find repeatable patterns that policy, tooling or training can address.
Deletion claims and UK GDPR
Personal data in prompts falls within UK GDPR. The controller needs to know where it is, why it is kept and who processes it. Deleting a chat from a user interface is not erasure from backups or vendor logs. Retention schedules should specify prompt history separately from email, and DSAR handling needs a tested export path. Records of processing should reflect the assistant, its purpose and its subprocessors. 'It is just a chat' is not a lawful basis. This is a governance test, not legal advice.
Shadow AI and the procurement gap
Shadow use often enters through personal accounts, free tiers and browser extensions. A half-day review should not become a hunt for individuals. It should identify repeatable patterns. If staff use personal accounts for customer data, the issue is capability and policy as much as discipline. Add a lightweight exception route and make the approved tool adequate for common tasks. Otherwise the organisation will keep discovering the same workaround under a different product name.
Ten questions before renewal
- Where is prompt history stored, and in which regions?
- What retention controls exist at tenant, user and workspace level?
- Can administrators export and delete history, and are deletions propagated to backups?
- Is prompt data used for model training by default, and how is any opt-out evidenced?
- Which subprocessors touch prompt data?
- What access do vendor support staff have, and under what controls?
- Are attachments and generated files retained separately from the prompt?
- What is the DSAR and eDiscovery process for chat history?
- What customer-facing logging is available?
- What changes on renewal, and what notice period applies?
The output
The deliverable is a prompt ledger: a register of tools, data classes, retention settings, access, owner and review date. Add a one-page paste rule, a retention line in the schedule and a named accountable person. The aim is not to stop AI use. It is to make the residue visible. What gets pasted stays pasted, but it can be governed.