Responsible AI Practice3 min read

Redact Before You Prompt: Run a Reidentification Check

Removing names is not enough. Use a four-pass reidentification check before approved AI tools receive workplace text.

Bokili EditorialPublished September 26, 2026· Verified September 26, 2026

A redacted workplace document whose indirect clues still converge on one identifiable person, beside a safer task-only version

Redaction is not the same as anonymity

Removing a name or email can still leave a combination of role, place, date and rare event that points to one person.

Before you paste a workplace document into an approved AI tool, ask a harder question than “Did I remove the names?” Ask whether a colleague with ordinary background knowledge could still identify someone from the remaining clues. A redacted note can preserve enough context to reveal a customer, employee or case.

Current ICO guidance distinguishes direct identifiers from indirect combinations such as location, age, dates and distinguishing traits. It also treats pseudonymised information as personal data when another key can reconnect it to a person. NIST makes the same practical point: de-identification reduces risk, but some data can be reidentified, including free text and images. This guide is a workplace safety method, not legal advice.

Run four passes before AI sees the text

The reidentification check

1. Strip direct identifiers

Remove names, email addresses, phone numbers, account numbers, precise addresses, employee IDs and other obvious labels.

2. Combine indirect clues

Read role, location, dates, age, shift, rare events, relationships and free-text details together. A harmless clue can become identifying in a bundle.

3. Test the motivated colleague

Imagine a colleague who knows the team, calendar and recent events. Could they narrow the text to one person without special access?

4. Keep only task-essential facts

Delete or generalise anything the AI does not need. Preserve the decision-relevant meaning, not the story’s full detail.

The third pass adapts the ICO’s “motivated intruder” idea to everyday work. The tester does not need secret databases or specialist hacking skills. They use information that is public, widely known inside the organisation or reasonably obtainable. The purpose is not to prove zero risk; it is to expose combinations that a simple find-and-replace misses.

Worked example: the name is gone, the person is still visible

An HR partner wants an AI tool to group themes from survey comments. One note originally names a supervisor. The name is removed, but the text still says: “the only night-shift supervisor in Lyon, after Tuesday’s loading-bay incident, who returned from parental leave last month.” Every clue may look ordinary. Together they describe one person.

Reading is a start. Practice makes it stick.

Start learning

	Name-only redaction	Task-essential version
Role	Only night-shift supervisor	A supervisor
Place	Lyon site	One operational site
Time	Tuesday after the loading-bay incident	After a recent operational incident
Personal detail	Returned from parental leave last month	Removed
Useful meaning	A specific person is exposed	The comment concerns support after an incident

The safer version still supports theme analysis: it preserves the workplace issue and removes details that do not affect the classification. If the exact shift, site or chronology is necessary, do not quietly keep it. Escalate to the data owner and use the approved process for that higher-risk task.

Redaction choices that fail under context

Check every release, not just the first draft

Search document properties, comments, tracked changes, hidden sheets and filenames for identifiers.
Check free text for rare events, unusual job titles, precise dates and personal circumstances.
Read all rows or passages together; separate fragments can combine into one identity.
Confirm the AI task needs each retained detail and generalise anything it does not.
Ask a second reviewer who knows the context to attempt identification.
Stop and escalate when the required context cannot be made proportionate.

The UK Government AI Playbook recommends minimising personal data and applying appropriate techniques such as redaction or pseudonymisation. It also warns that hidden information can remain in office documents. That makes redaction a process, not a black marker: inspect the file, the visible text, the context and the approved destination.

Test one paragraph in ten minutes

Choose a fictional or already approved paragraph; do not use live sensitive material for practice.
Underline direct identifiers and remove them.
Circle indirect clues: role, place, date, rare event, relationship and personal circumstance.
Read the circled clues as one bundle and ask who a knowledgeable colleague would infer.
Generalise or delete every clue that the AI task does not require.
Record the task, the retained facts and the reviewer’s decision.

Use less context—and keep responsibility visible

A useful AI brief does not need every detail of the original case. It needs the smallest set of facts that supports the requested work, plus a named person who owns the decision. If minimisation destroys the task, the answer may be a controlled workflow, a different tool or no AI use at all.

Build the practice with a safe fictional pack, map what a learning platform retains with the learner-data guide, and keep evaluation independent with the acceptance-test guide. Bokili’s learning features support short, workplace-based practice with safety and review built in.