Edcen Data All articles
Data Governance

The Evidence File: What UK Boards Should See Before an AI Pilot Touches Customer Records

Edcen Data
The Evidence File: What UK Boards Should See Before an AI Pilot Touches Customer Records

AI pilots rarely begin with a legal review. They begin with a promising use case, a sample dataset and a delivery date. In 2026, that pattern remains common in UK organisations: a team tests a model in a sandbox, then connects it to live customer records because the results look useful. The governance question arrives later, often during procurement or after a security review.

That is why permission should be treated as an operational asset. It is not a sentence in a privacy notice, nor a box ticked at project kick-off. It is a set of records that shows how customer data was obtained, why it may be used, who may process it and when it must be returned or deleted. Before an AI pilot touches customer data, senior teams should be able to produce that evidence quickly and in plain English.

Permission is a chain, not a checkbox

A useful way to frame the issue is as a chain. Customer data is collected for a purpose, under a lawful basis, with a notice shown at the time. It may then be shared with suppliers, combined with other datasets, retained for a period and eventually deleted. An AI pilot adds a new link: the use of records to train, test or operate a model.

Each link must hold. A privacy notice that mentions analytics does not automatically cover model training. A consent record for service updates does not necessarily cover product development. A supplier contract that covers hosting may not cover the vendor's use of data to improve its own models. The board does not need to resolve every legal question itself, but it should ask whether the chain can be demonstrated.

Four gaps that tend to surface late

1. The lawful basis is named but not mapped

Many UK organisations can point to a lawful basis under the UK GDPR and the Data Protection Act 2018. Fewer can show which basis applies to each data type in a pilot. Customer account data, website behaviour, call recordings and special category data may each require a different analysis. If the organisation relies on legitimate interests, it should hold a legitimate interests assessment that considers the pilot's purpose and the customer's reasonable expectations. If it relies on consent, it needs evidence rather than assumption.

2. Consent records cannot answer the pilot's question

Consent must be specific, informed and unambiguous. A record that says a customer agreed to marketing in 2024 may not answer whether they agreed to their data being used to train a model in 2026. Useful consent evidence includes the wording shown, the channel used, the date and time, the scope of the permission and the withdrawal route. Without that detail, a product team may believe consent exists when the organisation can only prove a general acceptance of terms.

3. Provenance is thinner than the dataset suggests

AI pilots often combine data from a CRM system, web forms, contact centres, partners and third-party sources. Provenance asks a simple but demanding question: where did each record come from, and what restrictions travelled with it? A provenance register should capture the source, collection date, notice version, contractual limits and any special category content. It should also flag data that was inferred rather than supplied, because inferred data can carry its own accuracy and fairness risks.

4. Vendor terms stop at processing, not learning

AI suppliers may act as processors, controllers or a mixture of both depending on the service. UK firms should check the processing terms, sub-processor arrangements, international transfer mechanisms, security commitments and incident notification duties. They should also look for clauses that address whether the vendor may use customer data to train or improve its own models, how long data is retained, how deletion is evidenced and whether audit rights are practical. A standard software contract may not be enough.

A pre-pilot evidence file

Senior teams can use the following checklist as a short governance gate. It is not a substitute for legal advice, but it helps expose gaps before procurement or product teams move ahead.

Questions the board should ask

Can we show the notice the customer actually saw? Which lawful basis covers the pilot, and where is that documented? Which contract clause prevents the vendor from using our data for its own model improvement? Where will the data be processed geographically, and what transfer mechanism applies? Who signs off deletion, and how would we prove it happened?

If the answer to these questions is that the privacy policy covers it, the evidence file is probably incomplete. Boards do not need to become data protection specialists, but they should insist that permission is demonstrated rather than asserted.

Make permission part of delivery

The strongest approach is to embed the evidence file in the pilot's delivery gate. Product teams provide the data map and purpose statement. Legal and data protection colleagues review the lawful basis and vendor terms. Security confirms transfer and retention controls. Data governance keeps the file current as the pilot scope changes.

For UK organisations, this is not only a compliance exercise. A pilot that can evidence its permissions moves faster when questions arise, because the answers are already assembled. Permission becomes part of the data strategy rather than an afterthought, and the board can judge an AI proposal on evidence instead of optimism.

All Articles

Related Articles

Walking-Away Rights: A Board-Level Test for Analytics Vendor Portability

Walking-Away Rights: A Board-Level Test for Analytics Vendor Portability

The Interest-Only Trap: A Ninety-Day Plan to Clear Data Debt Without Halting the Business

The Interest-Only Trap: A Ninety-Day Plan to Clear Data Debt Without Halting the Business

The Unseen Ledger: Bringing Shadow Spreadsheets Into the Light

The Unseen Ledger: Bringing Shadow Spreadsheets Into the Light