Most internal writing is fine to automate. Some of it is not.
A handbook section on how to book annual leave is information. Get it clear, get it accurate, move on. AI will halve the time it takes and the result will be better, because plain language is one of the things these tools do well and one of the things HR documents do badly.
A message telling forty people their roles are at risk is not information. It is the sentence someone will remember for years, and they will read it looking for evidence that whoever sent it understood what they were doing.
The whole lesson sits on that difference.
In plain English
- Policy explainer:
- A short plain-language version of a formal policy, written so people can act on it without reading twelve pages.
- Consultation:
- A formal process of discussing proposed changes with affected staff or their representatives. Requirements vary by country.
- Identifiable information:
- Anything that could point to a specific person: names, job titles in small teams, dates, case details.
- Consumer AI tool:
- A free or personal-account chatbot with no agreement covering how your organisation's data is handled.
Where it helps most
Handbook sections. Take an existing policy and produce the version a person can act on. Same rules, fewer words, no shall.
Policy explainers for managers. Managers do not need the policy, they need to know what to do, what to refuse and what to escalate.
Routine announcements. New system, changed process, updated form, revised deadline. Nobody's life changes and clarity is the only thing that matters.
Consistency checks. Paste in two policies and ask where they contradict each other. Most handbooks contain at least one pair of clauses that cannot both be true.
Rewrite this policy as a plain-language explainer for employees.
Rules:
- Reading age 14, short sentences, active voice
- Replace shall, must be adhered to, and in accordance with
with ordinary words
- Keep every rule and every deadline exactly as written
- Do not soften anything, and do not add reassurance
- Structure: what this covers, what you need to do,
what happens if you do not, who to ask
- Under 400 words
If any part of the original is genuinely ambiguous, list it at the end
under Needs a decision rather than resolving it yourself.
Policy:
[paste it]
The plain-language rewrite is also the best proofreading tool you have. When the tool cannot restate a clause clearly, that is usually because the clause does not say anything clear. That is a finding, not a failure.
The messages a human must write
Redundancy. Bereavement. Disciplinary outcomes. Grievance responses. Serious illness. Anything that changes someone's life rather than their calendar.
Write these yourself, or rewrite an AI draft so heavily that nothing of it survives. There are two reasons and both are practical.
People can tell. Generated sympathy has a texture. The three balanced clauses, the acknowledgement that this news may be difficult, the pivot to next steps in the same paragraph. Readers may not name what is off, but they notice, and once they suspect it they reread everything you have ever sent them with the same suspicion.
Being handled by a machine at the worst moment of your working life is its own injury. Separate from the redundancy, separate from the outcome. Someone finds out later that the message about their job was generated, and a second, entirely avoidable harm has landed on top of the first.
If you use AI at all for these, use it before you write, not to write. Ask what questions people will have. Ask what you have not thought about. Then close it and write the message.
โ Weak prompt
Prompt
Write a compassionate email announcing 40 redundancies in the operations team.
Output
We understand this news may be difficult to receive. This decision was not taken lightly, and we remain committed to supporting affected colleagues through this transition as we move forward together.
Four stock phrases in three lines. It performs sympathy without containing any information, and every person reading it will recognise the register from every other company that has done this.
โ Good prompt
Prompt
I have to tell 40 people in operations their roles are at risk. Do not write the message. List the questions they will ask in the first hour, in the order they will ask them, and tell me which ones I must answer in the announcement itself rather than later.
Output
Is my job going? When will I know? What is the process and how long? What is the redundancy payment? Can I apply for other roles internally? Who do I speak to today?
It prepared you rather than replacing you. You now know what the message must contain, and you write it yourself in your own words.
Checkpoint
Redundancy, bereavement, disciplinary and grievance messages are written by a human. Use AI beforehand to work out what the message must answer, then close it and write.
Never paste identifiable employee information
This one has no exceptions worth learning.
No names. No employee numbers. No salary figures tied to a person. No case details from a grievance or disciplinary. No health information. No performance notes about an individual. Nothing that could identify someone, including a job title in a team of three, which identifies them perfectly well.
The reason is simple. On a consumer account you generally have no agreement about where that text goes, how long it is kept, or who might see it. Employee data typically carries confidentiality obligations that do not pause because a tool is convenient. And an employee finding out their disciplinary case was pasted into a chatbot is a trust problem you will not talk your way out of.
Work in placeholders instead. The tool does not need to know who Sarah is to help you structure a difficult conversation.
I need to have a difficult conversation with a team member.
Situation, described generically: [an employee in a support role has missed
three deadlines in two months after previously performing well]
No names, no identifying details.
Help me prepare:
- How to open, in two sentences, without softening the point
- Three open questions that might surface a cause I have not considered
- What I should not say
- What I must record afterwards
- What would make this a matter for a formal process rather than a chat
Do not write a script. I am having the conversation, not reading it out.
If your organisation provides an enterprise AI account with an agreement covering employee data, the rules are whatever that agreement says, not whatever this lesson says. Check with your legal or compliance function and your data protection lead. Nothing here is legal advice, and confidentiality obligations around employee records vary by jurisdiction and sector.
Here are two sections of our employee handbook:
[paste section A]
[paste section B]
Identify:
- Any place where the two contradict each other
- Any rule stated in one that is silently different in the other
- Any deadline, notice period or threshold that does not match
- Anything a reasonable employee could interpret two ways
Quote the exact conflicting lines. Do not rewrite anything and do not
suggest which version is correct. I will decide that.
๐ Quiz
Question 1 of 4Which type of message must be written or heavily rewritten by a human?