Can AI Draft Pension Auto-Enrolment Letters? Check Regulator Wording

Yes, AI can draft auto-enrolment letters, but only around the regulator's wording, never instead of it. Use The Pensions Regulator's published templates as locked text, let AI handle personalisation, the covering note and error-checking, then compare each final letter against the mandatory content list before it goes out.

The risk is not a badly written letter. It is a fluent letter that quietly drops a required statement, such as the opt-out period or the right to opt back in. AI rewrites tend to smooth exactly those clunky sentences. So the workflow below treats the template as fixed and asks AI to do the jobs around it.

Follow me on Instagram@sagnikteaches

What the regulator says an enrolment letter must contain

The regulator's detailed guidance on information to workers lists the mandatory content, and it is longer than most people remember. For a jobholder being automatically enrolled, the employer must give a statement that they have been or will be enrolled, the enrolment date, the contributions payable by employer and jobholder (as a fixed amount or a percentage of qualifying earnings or pensionable pay), and, for a defined contribution scheme, a statement about the phased increase in contributions. It must also say that the jobholder's contributions are deducted from pay and whether tax relief is given.

Connect on LinkedInSagnik Bhattacharya

The second block is about the right to opt out. The letter must state that the right exists during the opt-out period, give the start and end dates if the employer knows them, say where the opt-out notice can be obtained, and explain that opting out means being treated as never having become an active member. It must say that contributions already paid are refunded, that an opted-out worker may opt in again once in any 12-month period, that a written notice must be signed (or, if electronic, must include a statement that the worker personally submitted it), that after the opt-out period the jobholder may stop contributing under the scheme rules, and that opting out normally leads to automatic re-enrolment later.

Subscribe on YouTube@sagnikteaches

The guidance also sets timing: enrolment information must be given no later than six weeks after the automatic enrolment date (or re-enrolment date, or the enrolment date after an opt-in). A postponement notice has its own rule: no later than six weeks and a day from the first day of employment, or from the date the eligible-jobholder criteria are met, depending on how postponement is used. Always check the current guidance for the case in front of you; this tutorial describes how to check, not what a particular client's facts mean.

Why the templates become locked text

The regulator publishes Word and ODT templates for four situations: staff being put in a scheme, staff being put in a scheme but not receiving tax relief (net pay arrangements), staff not being put in a scheme, and a postponement letter for all employees, plus an insert to accompany the first letter. They also come in a range of languages. The guidance says templates are one way to communicate the information and that employers can add their own wording, provided the mandatory requirements are met.

For a bureau, "may use your own wording" is a trap. Every paraphrase is a chance to omit a statement, and the regulator's text has already been written against the rules. The sensible default is that the statutory paragraphs are copied in unaltered and only the clearly marked placeholders change. The regulator's website terms add that any reproduction of its materials should acknowledge the regulator as the source, which is worth remembering before you paste template text into a client pack or internal procedure.

Locked text still carries figures

Open the current English templates and you will find numbers written into the prose, not left as placeholders. The enrolment letter states the earnings trigger as a weekly and monthly figure; the insert says the employee will pay in 5% and the employer 3%, a total of 8%, and gives the lower earnings level at which the employer must also contribute to someone who opts in. The postponement letter repeats both thresholds. Those sentences are statutory content, but they are also the first place a letter goes stale: when thresholds are revised, or when a client pays more than the minimum, the figures must change and nothing else should.

So treat every number inside locked text as a controlled edit. One named person updates it from the provider's schedule and the current thresholds, a second person approves, and the change is logged against the template's download date. The AI comparison in Prompt 2 then has a clear job: flag every figure that differs from the approved master, so a 3% that should read 5% cannot slip through on letter 40.

Three layers in every letter

• Locked: statutory statements copied from the template. AI must not edit them, only check they are present.
• Fields: name, dates, scheme name, contribution rates, opt-out deadline, contact details. These come from payroll or the pension provider, never from AI's memory.
• Free: the covering email or cover note, a plain-English summary box, the reminder subject line. AI can draft these because they carry no legal content.

Seven steps from template download to send log

Take a bureau with 40 small-employer clients and a busy month of new starters and re-enrolment letters. The aim is one repeatable process, not 40 hand-built letters. A reasonable sequence looks like this.

1. Download the current templates. Take them from the regulator's letter-templates page at the start of each cycle rather than reusing a file from the shared drive. Save them with the download date in the filename.
2. Mark the placeholders. In each template, highlight every item that changes per employee or per employer. Count them: the regulator's own enrolment letter and insert have only five or so (date, employer name, employee name, enrolment date, provider contact point), but a bureau's adapted version often adds the scheme name, both contribution rates and the opt-out end date.
3. Build the field list. A payroll export with one column per placeholder. This is where AI helps most, by mapping column names to placeholders and flagging gaps.
4. Merge in Word or your payroll software. Do the actual filling with a mail merge, not by pasting data into a chatbot.
5. Run the wording check. Compare each merged letter type (not every letter) against the locked text and the mandatory content list.
6. Spot-check the data. A named human reads a sample of merged letters for dates, rates and opt-out deadlines.
7. Send, record and diarise. Keep the sent version, the date, the method and who sent it.

Prompt 1: map payroll columns to template fields

This is the safest and most useful job. The prompt contains column headings and placeholder names only, with no employee data.

I am preparing a mail merge for pension auto-enrolment letters.
Below are (A) the placeholders in the template and (B) the column
headings in our payroll export. No employee data is included.

A. [first name], [enrolment date], [scheme name], [employee rate],
[employer rate], [opt-out end date], [provider contact]
B. FirstName, AEDate, Scheme, EeRate, ErRate, OptOutBy, ProviderTel

1. Match each placeholder to a column, or say "no match".
2. Flag any column whose format (text, date, percentage) could
   break the merge.
3. Do NOT rewrite or suggest wording for the template itself.

A realistic reply maps six of seven, says "[provider contact] has no match; ProviderTel is a phone number only, and the template also needs an address or website", and warns that AEDate may import as a serial number in Excel unless it is formatted as a long date. Both warnings are useful. Neither touches the statutory words.

Prompt 2: a wording check that only reports

The second job is a comparison. Give AI the regulator's template text (public text, nothing confidential) and the merged letter with the employee's name replaced by [first name]. Ask for a report, not a rewrite.

Compare LETTER against TEMPLATE.
TEMPLATE is regulator wording and must not change.

Report in a table:
- Sentences in TEMPLATE that are missing from LETTER
- Sentences in LETTER that differ from TEMPLATE, quoting both
- Numbers or dates in LETTER that look inconsistent with each other
Do not suggest improvements. If everything matches except the
placeholders, say "no differences found".

Illustrative output from a letter where someone shortened the opt-out paragraph:

Notice what the second row catches. "At any time" sounds friendlier than "once in any 12-month period", and it is wrong. A human reviewer skimming for typos could easily pass it. A text comparison does not get tired at letter 30.

Do not treat the AI's "no differences found" as proof. Run a deterministic comparison too: Word's Compare feature, or a diff tool, shows every changed character. Use AI as a second reader, not the only one.

Prompt 3: the covering note, where AI can write freely

The covering email carries no statutory content, so this is the one place a polished draft is welcome. It should point to the attached letter, not replace it, because merely signposting does not count as giving the information.

Write a short covering email (90 words max) to an employee at a
small business. It tells them a letter about the workplace pension
is attached, that the letter explains their right to opt out and
the dates that apply, and that they can reply to [payroll contact]
with questions. Friendly, plain English, no pension advice, no
statements about what they should do. Use [first name].

Illustrative output: "Hi [first name], your workplace pension letter is attached. It explains what happens next, including your right to opt out and the dates that apply. Please read it this week. If anything is unclear, reply to this email and [payroll contact] will help. Thanks, [your name]." Edit "this week" if your deadlines do not support it, and never let the email summarise rates or dates that could contradict the letter.

A filled-in locked-text checklist

Keep one checklist per letter type, completed by the person who prepares it and countersigned by the person who approves it. This version is for the staff-being-enrolled letter and is illustrative; the list you use must come from the current regulator guidance.

The tax-relief row is a good example of a template choice that trips people up. The regulator publishes a separate English template for staff who do not receive tax relief on contributions because the scheme uses a net pay arrangement, and says to use it only in that situation. An AI that has never seen your scheme rules cannot know which applies; the provider paperwork does.

Which letter type applies: a quick decision table

Deciding who sits in which row is the assessment work in the regulator's detailed guidance, driven by age, earnings against the thresholds in each pay reference period, and the employer's choices. That is a payroll decision from your software and your own checks. AI can help you draft a query list for a borderline case, but it should not make the call.

Where the drift usually shows up

Four examples of errors a bureau will recognise, all illustrative.

Rate drift. A letter shows 5% employee contribution where the scheme is set at 3%, or the template text predates a rate change. The current statutory minimum total is 8%, with at least 3% from the employer, on qualifying earnings between GBP 6,240 and GBP 50,270; those bands are reviewed, so take the letter's figures from the provider and use the guidance only as a cross-check.

Date arithmetic. A bureau counts six weeks from the day the letter is prepared instead of from the enrolment date. A date-check prompt on a column of enrolment dates and send dates finds any gap over 42 days in seconds. That is arithmetic you can also do in Excel; see how to clean messy data before the merge if the export is untidy.

Translation drift. A client asks for letters in the first language of most of its staff. Check the regulator's templates page first (it lists more than a dozen languages, each with the same set of letters) and use its version, not a machine translation of the English one.

Merge residue. A letter goes out reading "Dear [first name]," because one row had an empty cell. A final-pass search for square brackets across the merged output catches it. Ask AI to write the Word macro or the search expression, and test it on a dummy file.

Keeping pay data out of the chat

Nothing above needs an employee's name, tax identifier or salary in an AI tool. Placeholders and headings carry the whole task. If you later want AI inside the payroll system itself, the question becomes what that vendor does with the data; which payroll AI features are real and what to ask covers the vendor conversation, and safe prompts that never include pay data gives wording habits for the team.

Write the rule into your procedure in one line: "Statutory letter text is copied from the regulator's current template and is changed only by a named person with sign-off; AI is used for field mapping, comparison reports and non-statutory covering notes." Keep a log of each cycle: template download date, who prepared, who approved, how many letters, how many checks flagged anything.

What the bureau sends back to the employer

The guidance is clear that an adviser or bookkeeper can send information but the employer remains responsible for it being given on time and being correct and complete. So your service terms should say who checks, and your cycle summary should give the employer evidence. A short summary email can list letter type, number sent, date sent, method, bounced or returned items, and chased items. The guidance specifically notes that returned post or a bounced email means the information cannot be considered given, so a bounce list is not an admin extra; it is part of the duty. Keeping email addresses current reduces the risk.

For a bureau that also runs the monthly cycle, the pension file ties into the pay run. What to check before buying payroll automation software sets letters in that wider picture, and AI for payroll: real-time filings, auto-enrolment and holiday pay checks covers neighbouring tasks.

A mixed month: three clients, three different letters

One more walk-through, because a bureau's real month is never one letter type. Say three clients each have new starters on the same Monday. Client A is a small architecture practice with a relief-at-source scheme: its starter is aged 27 and earns above the trigger, so the enrolment letter and the insert go out. Client B is a four-fee-earner law firm that uses postponement for every new starter, so its starter gets the postponement letter with the deferral date filled in and the lower qualifying earnings figure for the pay period stated. Client C is a marketing agency whose part-time starter is under 22, so the "staff not being put in a scheme" letter applies, explaining the rights to opt in or to join.

The bureau's cycle sheet shows three rows, one per letter type, each with a template filename and download date, a preparer, an approver and a date sent. The AI jobs run per type: column mapping on three export layouts, one comparison report each, and one covering note that covers all three because it carries no statutory content. If the comparison for Client B flags that the deferral date is before the start date, the fault is in the payroll record, not the letter, and the fix goes back to the payroll system, not into a hand-edited document.

The same sheet is useful when an employer rings in six months and asks what was sent to whom. A dated template, a named checker and a stored copy answer that in a minute, which is a better position than "we think the template was the current one".

A small illustration of the saving

Say a bureau sends 120 enrolment, non-enrolment and postponement letters in a busy month. Hand-checking each letter against the template takes about six minutes, so 12 hours. Under the layered process, the comparison runs once per letter type (four types, about ten minutes each with a diff tool and an AI report), then a human reads a sample of ten merged letters at four minutes each. That is roughly 80 minutes instead of 12 hours, and the check is better because it was done against the template text instead of memory. These figures are illustrative; time your own process before promising anyone a saving.

If you would like help setting up a locked-text workflow for your own letters, that is the kind of thing my AI implementation consultation covers; get in touch. Otherwise, the steps above are enough to start this week: download the current templates, highlight the placeholders and run Prompt 1 on your export headings.

Where AI should not be used on these letters

• Deciding whether someone is an eligible jobholder, non-eligible jobholder or entitled worker. That sits in your payroll software and the regulator's assessment guidance.
• Rewording statutory statements for readability, even slightly.
• Choosing the no-tax-relief template without the provider's confirmation that the employee gets no relief.
• Answering an employee's individual questions about whether to opt out. That is advice territory, and the letter's own wording is the safe answer.
• Generating dates, rates or contribution amounts from memory.

Everything on this list shares a trait: a wrong answer would look correct and reach an employee before anyone noticed. The jobs AI is good at here are the opposite: they produce a report that a person then acts on.

Questions bureaus ask before letting AI near enrolment letters

Can I paste employee names and pay into a chatbot to fill the letters?

Avoid it. A list of placeholders such as [first name] is enough for AI to plan the mail merge, and your payroll software or Word then fills the real names, dates and amounts. Keep personal and pay data inside systems your data-protection terms already cover, and check any AI tool's data settings before use.

Does the regulator forbid using my own wording instead of its templates?

Its guidance says the templates are one way to communicate the information and that an employer may add its own wording or use its own templates, provided the mandatory requirements are met. That is why a locked-text checklist matters: every required statement must be present, whatever the style.

Is emailing a link to a pensions page enough?

No. The regulator's guidance says giving information does not include merely signposting to a website, attaching a URL or displaying a poster. The information itself must be given in writing, by post, by hand, in the body of an email or as an email attachment.

Who is responsible if a bureau sends a wrong letter?

The guidance says someone acting for the employer, such as an accountant or bookkeeper, can send the information, but it remains the employer's responsibility that it is given on time, correct and complete. Your engagement terms should say who checks what.

Should AI translate the letters?

The regulator publishes its templates in a range of languages, so start from those. Machine translation of statutory wording risks changing the meaning. If you need a language it does not cover, have a competent speaker check the result against the English template line by line.

Further reads

• AI Tools for Payroll: What a Small Bureau Can Automate and Must Check — Which tasks a small bureau can automate and must still check.
• Payroll Year-End Checklist: Where AI Helps and Where It Misleads — Year-end tasks where AI helps and where it misleads.
• How to Use AI for Payroll: Query Answers, Checks and Reports — Using AI for payroll queries, checks and reports.
• Can AI Do Payroll? What a Small Employer Can Hand Over Safely — What a small employer can hand over to AI safely.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: The Pensions Regulator automatic enrolment detailed guidance (no. 10, information to workers) and its automatic enrolment letter templates page, both read in November 2026; the government's workplace pensions guidance on contributions; The Pensions Regulator website terms and conditions.