Remove personal data from text before AI sees it

To decide what personal data to remove from a document before AI reads it, go through five groups in order: the numbers that name one person (National Insurance, NHS, passport, driving licence, UTR), contact details, bank and card details, case and file references, and finally the details that identify someone only in combination, such as a job title with an employer, a town and a date. The first four you delete or replace. The fifth you have to read for.

This guide is the list, with the British identifiers you actually find in UK paperwork and where each one hides. It does not cover the mechanics of taking them out of a particular file: for that, see our guides on redacting a PDF so the text is really gone and redacting a Word document before an upload.

It is written for people who hold client files in UK firms (solicitors, accountants, HR advisers, letting agents, brokers) and would like an AI tool to summarise, draft or check them. If the file is a Word document, settle its tracked changes and delete its comments before working through the list, because a name crossed out under track changes is still in the file.

What personal data to remove from a document before you use AI

The UK GDPR defines personal data in Article 4(1) as information about a living person who can be identified, directly or indirectly, and names the kinds of identifier it means: a name, an identification number, location data, an online identifier, or factors specific to that person. Everything in the table below is one of those, written the way it appears on a British letter, form or payslip.

The Data Protection Act 2018 uses the same test and limits it to a living individual. That does not make a deceased client’s file free to paste: your professional duty of confidence still applies, and so do the rights of every living person named in it.

Group	What to look for	Where it usually sits
Direct identifiers	Name, NI number, NHS number, passport, driver number, UTR	Letterheads, payslips, forms, ID copies
Contact details	Home address, postcode, mobile, email, social handles	Signature blocks, forwarded emails
Financial details	Sort code with account number, IBAN, card number	Refund instructions, invoices, statements
Case and file references	Claim, policy, tenancy, hospital and court numbers	Subject lines, headers, footers
Quasi identifiers	Job title, employer, town, dates tied to a person	The narrative paragraphs

Groups follow the ICO’s guidance on identifiers and its anonymisation glossary; the examples are UK paperwork.

The reason for taking them out is minimisation, set out in Article 5(1)(c): what you process should be adequate, relevant and limited to what the purpose needs. An AI tool asked to tighten the wording of a complaint reply does not need the complainant’s date of birth, and the ICO’s guidance on AI and data protection applies the same principle to AI systems.

What happens to a document once it is uploaded depends on the tool and the account. We cover that separately for ChatGPT and how long it holds your chats and for Copilot with confidential information. This list assumes the safer habit: send the tool the least it needs.

Direct identifiers: the UK numbers that point at one person

The ICO’s anonymisation glossary calls a direct identifier any single item that could uniquely identify someone, and its examples are a name, an address and unique reference numbers, the NHS number among them. In a British office, the numbers come from a handful of state bodies, and each has a recognisable shape. Knowing the shape is what lets you spot one in a scanned attachment or a line of a spreadsheet pasted into an email.

The name comes first, and it is rarely written only once. A client appears as Imogen Thackeray in the header, Ms Thackeray in the salutation, Imogen in the second paragraph and IT in the file reference. Each of those is the same person, and each one has to go.

National Insurance number

HMRC’s National Insurance Manual sets out the format: a prefix of two letters, six digits, and a suffix that can only be A, B, C or D. It also lists what never appears. D, F, I, Q, U and V are not used as either prefix letter, O is not used as the second, and the prefixes BG, GB, KN, NK, NT, TN and ZZ are not issued.

That last detail matters when you test a tool or a search. The specimen HMRC prints, QQ 12 34 56 A, uses a letter no real prefix contains, so a checker that ignores it is behaving correctly. Never test with a made up number in a valid shape, because it may belong to someone. The same manual lists prefixes that are no longer issued, such as PZ, unused since August 2002, so PZ 12 34 56 A has the right shape and is not issued today.

In paperwork, the NI number turns up on payslips, P45 and P60 forms, pension letters, benefit correspondence and employment contracts. It is often printed without spaces, as PZ123456A, which a search for the spaced version will miss.

NHS number, and the numbers used in Scotland and Northern Ireland

The NHS website describes the NHS number as ten digits, printed in the example as 485 777 3456, and says it appears on any letter or document the NHS sends: prescriptions, test results, appointment letters.

It also appears where you would not expect a health document at all: an occupational health referral inside an HR file, a letter a client forwarded with a sick note, a solicitor’s bundle in a personal injury claim.

Scotland uses the CHI number and Northern Ireland the Health and Care number, which our guide for NHS staff using ChatGPT covers in detail. Next to any of them, the diagnosis in the same paragraph is health data, a special category that carries its own rules under Article 9.

The driving licence number spells out a name and a birthday

Of all the British identifiers, this is the one people underrate. DVLA’s leaflet INS57P, which comes with every photocard, explains how the driver number is built, using its own sample MORGA 657054 SM 9IJ.

Part of the sample	What DVLA says it holds
MORGA	First five letters of the surname, padded with 9 if shorter
657054	Date of birth: decade, month (with 5 added for women), day, last digit of the year
SM	The first two initials, with 9 if there is only one
9IJ	Check characters generated by the computer

Source: DVLA, INS57P, “How to check your driver number”.

Read that back and the sample says: surname beginning Morga, a woman born on 5 July 1964, initials S and M. A driver number in a document therefore carries most of a surname, a full date of birth and a sex, even after you have removed all three from the text around it. Copies of licences turn up in anti money laundering files, vehicle hire agreements and motor insurance claims, so cover and watermark the copy you send.

Passport, UTR and the other HMRC references

A passport number is short and unlabelled on its own, but a copy of the photo page is rarely just the number: it carries the full name, date and place of birth and a photograph. The ICO’s page on identifying someone indirectly lists passport numbers, NI numbers and vehicle registrations among the details that can point to someone once combined with other information.

The Unique Taxpayer Reference is ten digits, printed on self assessment returns and on HMRC’s notices to file and payment reminders. An individual’s or a sole trader’s UTR is about a person. A limited company has its own UTR, which is about the company, though in a one person company the line between the two is thin.

Vehicle registrations

A number plate is about a vehicle, and the vehicle has a registered keeper. The ICO names vehicle registrations among the details that can identify someone indirectly, once linked to other information, and in a motor claim, a parking dispute or a fleet file that link is usually one line away. A plate next to “parked outside number 12 every evening” is a person.

Current British plates also carry a date and a region. DVLA’s leaflet INF104 explains that the first two letters are a memory tag for the region of first registration and the two numbers an age identifier, so an invented plate such as LQ71 KTD dates the car to between September 2021 and February 2022. That one cannot exist, since DVLA does not use Q in memory tags.

On its own that is harmless. Next to a named street and a job title it narrows the field further, which is why plates belong on the list even though they identify a car first.

Everything else with a number on it

The rest of the list is local to your sector: payroll numbers, pension scheme member numbers, student numbers, tenancy deposit references, Home Office references on immigration letters. The rule for all of them is the ICO’s: if it is a unique reference to one person, treat it like a name. Our international guide on masking a document for AI across countries sets the same question out market by market, if you work with files from more than one country.

Contact details, and the email address that is really a name

Contact details identify a person and also give a route to them, which is why they are worth removing even when the name has already gone. The list is short: home address, postcode, landline and mobile numbers, personal and work email addresses, social media handles, and the online identifiers that the UK GDPR names in its definition, such as an IP address.

The trap is the email address. Most work addresses are built from the name, so imogen.thackeray@ or ithackeray@ is the name again in a different form. Forwarded chains are the worst place for this, because every header in the chain repeats the sender, the recipients and anyone copied in.

245people whose email addresses went out in the To field

£350,000fine on the Ministry of Defence

The ICO’s penalty notice to the Ministry of Defence is the plainest UK example that addresses alone are personal data. The ministry’s relocation team put the email addresses of 245 Afghan nationals eligible for relocation in the To field instead of BCC, and the ICO fined the ministry £350,000. The notice records the government’s own acknowledgement that not every address was a full name, and that this did not change the impact.

Before an upload, the practical steps are to delete the headers of forwarded messages, strip signature blocks, and swap each address and number for a placeholder like [EMAIL_1]. If the AI needs to know that two messages came from the same person, a consistent marker keeps that without keeping the person. Which tool receives the markers matters too: our guide to whether Gemini uses your data compares the free app with the Workspace version.

The other people in the file

A client’s document is rarely about the client alone. A grievance names the manager and the witnesses, a tenancy file names the guarantor, a family matter names children, and a complaint reply quotes a neighbour. Each of them is an identifiable person under the UK GDPR, with the same contact details, numbers and dates, and none of them agreed to be part of your prompt.

Children need the most care, because a first name, a school and a year group are often enough in a small town. Replace every third party with a role (Witness 1, the guarantor, Child A) before the document goes anywhere.

Bank and card details: sort code, account number, IBAN

Refund instructions, invoices, direct debit mandates and bank statements all carry financial identifiers, and they are among the easiest to spot because of their fixed shapes. GOV.UK’s own pattern for bank details asks for a sort code of 6 digits and an account number of 6 to 8.

Item	Shape in UK paperwork	Identifies a person?
Sort code	6 digits, often in pairs: 60-16-13	Only a bank branch, on its own
Account number	6 to 8 digits	Yes, with the sort code
IBAN	GB, 2 check digits, bank code, sort code, account	Yes: both are inside it
Card number	Long number, usually in groups of four	Yes, and it is also a payment risk
Building society roll number	On the card, statement or passbook	Yes, with the society’s name

Shapes from the GOV.UK Design System bank details pattern; the IBAN layout from our invented example below.

The IBAN is the one to watch in documents from clients with European accounts or international payments. A UK IBAN such as GB83 NWBK 6083 7141 9260 35 contains the sort code and account number in plain sight, so taking the account number out of the text and leaving the IBAN underneath undoes the work.

Amounts are a separate question. A figure of £2,340 is rarely identifying on its own, and the AI usually needs it to do the job. An unusual amount tied to a date can identify someone to a person who knows the case, which is a quasi identifier problem and belongs to the section below. For firms whose whole file is financial, our page for accountants shows a covered example.

Case, file and policy references

Reference numbers are the identifiers most often left in, because they look like admin rather than personal data. The ICO’s glossary disagrees: it lists unique reference numbers as direct identifiers. A reference that opens a file about one person is, to anyone with access to that system, as good as a name.

In a letter or a bundle they live in the subject line, the header, the footer and the first line of a letter, and they tend to be repeated on every page.

Numbers that open someone’s file

• Court and tribunal claim numbers, and the case name that usually carries a surname.
• Insurance claim and policy numbers, which link to the insured’s full record.
• Hospital numbers and medical record numbers, next to the NHS number.
• Tenancy, deposit protection and council tax references.
• HMRC and Home Office case references on correspondence.
• Your own matter or client numbers, especially when they contain initials.

For solicitors, the case name is the one that slips through: “Thackeray v Ledwyche Homes Ltd” in a header after every “Ms Thackeray” in the body has been replaced. Our guide on AI in court documents covers what else a filing needs, and the solicitors page shows a witness statement before and after covering.

Register numbers anyone can look up

Some numbers are public by design, and that makes them more identifying, not less. The SRA’s Solicitors Register can be searched by anyone, and other professional regulators publish registers in the same way.

Companies House works the same way for directors. Its blog on personal details on the register explains that for directors appointed since 10 October 2015 the month and year of birth are public, along with a correspondence address. A company name and “the director, born March 1979” is therefore one search away from a person.

Quasi identifiers: job title, employer, town and a date

This is the group no search pattern will find, and the one the ICO insists on. Its guidance on making anonymisation effective is blunt about it: taking out the direct identifiers alone does not get you there. The glossary calls what is left indirect identifiers, also known as quasi identifiers: any piece of information, or combination of pieces, that can be used to identify a person.

The ICO’s own example, on its page about identifiers and related factors, uses a very common name. John Smith alone may not be personal data, because there are many of him. John Smith who works at the Post Office in Wilmslow is usually one man.

An invented British example, taken apart

Here is a sentence from our invented client file note, with the name already removed, and what each part contributes.

Detail left in the text	What it narrows down
“is the practice manager”	One role, held by one person per practice
“at a dental practice”	A type of employer, a handful per town
“in Wendlecombe”	A market town, not a city
“and left in March 2025”	One month, known to every colleague

Our invented client file note; the person does not exist, and neither does Wendlecombe.

No single row identifies her. All four together describe one woman to every patient of that practice and everyone she worked with. That is exactly the reader an AI summary may end up in front of, since summaries get pasted into emails, reports and complaint replies. A spreadsheet lines such details up on every row, which is how a staff list with no names in it can still point at one person.

The fix is to generalise, not to delete. “A practice manager at a healthcare employer in Shropshire, who left last year” keeps what the AI needs to draft an employment reference or summarise a grievance, and it no longer points at one person. Keep the fact the task depends on and blur the ones it does not.

The ICO’s anonymisation guidance does the same with free text. Its worked example on interview transcripts removes the names and replaces specific place names with a general term, because in a narrative the place is often what gives the speaker away. A grievance letter, an attendance note or a complaint is that kind of narrative, and deserves the same treatment. So is the notes column of a client export, which is usually safer deleted than cleaned.

The ICO’s test for whether you have done enough is the motivated intruder: a reasonably competent person, with no special skills, trying to put a name to the text. Our guide on pseudonymisation under the UK GDPR runs that test on a full paragraph.

A postcode is about fifteen front doors

15

The Office for National Statistics describes postal geography this way: a single unit postcode may hold up to 100 addresses, but 15 is more typical. Next to a date of birth or a job title, a full postcode is often enough to find a household.

If the AI needs a location at all, the outward code (ZZ9 in our example) or the town is usually sufficient. For a letting agent’s rent arrears summary, “a two bedroom flat in south Shropshire” says what the draft needs. Dates tied to a person work the same way: the month or the year, or day 1 and day 14 counted from an event, instead of a date of birth or a dismissal date.

What you can usually leave in

Minimisation cuts both ways. Strip too much and the AI has nothing to work with, and staff go back to pasting the original because the covered version is useless. The aim is to keep the facts the task depends on and lose the facts that point at a person.

In practice the split looks like this, and the right hand column is where most of the judgment goes.

Usually fine to leave in	Look at twice
Amounts that are not memorable	A sum odd enough that someone remembers it
Hearing dates, filing deadlines	Birth, admission or dismissal dates
Roles: the tenant, the employer, Adviser A	A rare condition or an unusual hobby
The legal or financial question itself	A nationality in a small community
Your firm’s public details, on an approved tool	A relative’s name, or a photograph

Our summary of Article 5(1)(c), applied to documents.

None of the items in the right hand column is on a numbered list, and each can finish off an identification that the rest of the document started.

Removing names and numbers is still pseudonymisation rather than anonymisation, a distinction our guide to what pseudonymised data is under UK law explains.

Where a document is about health, beliefs or union membership, taking the name out does not change the category of what is left, as our special category guide sets out. Deciding, in writing, what staff may leave in is much of what an AI acceptable use policy needs to say, and the tools question sits in choosing AI tools under the UK GDPR.

Twelve checks, in the order you read the page

Work on a copy, and go through the document in the order a stranger would read it: top to bottom, including the parts the screen does not show. The first nine are about values; the last three are about meaning. Getting it right before sending matters because taking text back afterwards means several separate buttons, as our guide to clearing a ChatGPT chat, its memories and its files shows.

1. Header and file reference: matter numbers, case names, initials in the reference.
2. Names, every form of them: full name, title and surname, first name alone, initials, nicknames.
3. NI, NHS, passport and driver numbers: spaced and unspaced, and on attached ID copies such as a driving licence.
4. UTR and HMRC references: on tax letters and self assessment pages.
5. Addresses and postcodes: home address first, then any postcode finer than the town.
6. Phones, emails and handles: including forwarded headers and signature blocks.
7. Bank details: account numbers, IBANs, card numbers, building society roll numbers.
8. Claim, policy and tenancy references: especially in subject lines and footers.
9. Other people: colleagues, relatives, neighbours, witnesses and the other side.
10. Dates tied to a person: birth, admission, dismissal or accident, reduced to month or year.
11. The combination: job title, employer, town and date together, with one or two generalised.
12. The rare detail: anything unusual enough that a colleague would say “that must be her”.

Check the pictures as well as the words. A scanned passport, a photographed licence or a screenshot of a bank app sits in the file as an image, and a text search will never find it. Delete the image rather than drawing over it, and remember that a tool which accepts images may read the text in them, and keeps the image itself, in ChatGPT’s case even after the chat is deleted. On a slide, a black box over a screenshot leaves the whole picture underneath.

Last, read the remainder the way a colleague of the client would. If the answer to “who is this?” comes easily, go back to checks 11 and 12.

The places inside a file where these details hide are different for each format. Tracked changes, comments and text boxes in a .docx are covered in our Word redaction guide, and the text left under black boxes in a PDF in our PDF redaction guide.

Our invented client file note, covered by Nonimo

Drop a document into the Nonimo app and it gives you two things: the text with the data covered, ready for the AI, and a covered copy of the file to download. The app pseudonymises. Each name or number turns into a placeholder like [PERSON_1], and the app puts the originals back when the AI replies. The box above this guide covers pasted text and discards it as soon as you have the result. Here is the client file note behind the Wendlecombe example, all of it invented:

In the note                              What came back
Imogen Thackeray                         [PERSON_1]
14/09/1987 (after "date of birth")       [BIRTH_DATE_1]
PZ 12 34 56 A                            [REFERENCE_1]
1392018465 (UTR)                         [REFERENCE_2]
THACK852317I99LW                         [REFERENCE_3]
12 Weaver's Row, Wendlecombe ZZ9 9ZZ     [RECORD_FIELD_1]
imogen.thackeray@example.com             [EMAIL_1]
31926819 (account number)                [REFERENCE_4]
GB29NWBK60161331926819                   [IBAN_1]
999 000 0018 (NHS number)                [REFERENCE_5]
Imogen (second mention)                  [PERSON_2]

Our security page sets out what the Nonimo app keeps on your computer and what it sends.

Sources

• UK GDPR, Article 4, legislation.gov.uk: personal data is information relating to an identified or identifiable living person; identifiers include a name, an identification number, location data and an online identifier.
• Data Protection Act 2018, section 3, legislation.gov.uk: personal data means information relating to an identified or identifiable living individual.
• UK GDPR, Article 5, legislation.gov.uk: data minimisation in Article 5(1)(c), adequate, relevant and limited to what is necessary.
• What are identifiers and related factors?, ICO: the John Smith example; a common name combined with a place of work usually identifies one person.
• Can we identify an individual indirectly from the information we have (together with other available information)?, ICO: vehicle registrations, NI and passport numbers, and combinations such as age, occupation and place of residence.
• Glossary, ICO anonymisation guidance: direct identifiers include a name, an address and unique reference numbers such as an NHS number; indirect identifiers, also known as quasi identifiers.
• How do we ensure anonymisation is effective?, ICO: removing direct identifiers alone is insufficient; the motivated intruder; singling out.
• Guidance on AI and data protection, ICO: data protection principles, including minimisation, applied to AI systems.
• National Insurance Manual, NIM39110, HMRC, updated 22 July 2026: two letters, six digits and a suffix of A, B, C or D; letters and prefixes not used; administrative prefixes no longer issued, among them PZ, unused since August 2002; the specimen QQ 12 34 56 A.
• Find your NHS number, NHS website: ten digits, example 485 777 3456; printed on prescriptions, test results and appointment letters.
• INS57P, information about your driving licence, DVLA: how the driver number encodes surname, date of birth, sex and initials, with the sample MORGA 657054 SM 9IJ.
• Find your UTR number, GOV.UK: a 10 digit number, on tax returns, notices to file and payment reminders.
• Bank or building society account details, GOV.UK Design System: sort code of 6 digits, account number of 6 to 8 digits, roll number.
• INF104, vehicle registration numbers and number plates, DVLA: memory tag for the region of first registration; the age identifier 71 covers September 2021 to February 2022.
• Monetary penalty notice: Ministry of Defence, ICO, December 2023: email addresses of 245 people disclosed in the To field on 20 September 2021; penalty of £350,000.
• Postal geographies, Office for National Statistics: a unit postcode may contain up to 100 addresses, but 15 is more typical.
• Your personal details on the Companies House register, Companies House: for directors appointed after 10 October 2015, only month and year of birth are public; the correspondence address appears on the register.
• Solicitors Register, Solicitors Regulation Authority: the public register of solicitors and firms.