Drafting emails with human review.

AI writes a passable email in seconds, which is exactly the risk. Learn a routine where you decide the content, AI drafts the words, and a five-point review catches wrong facts and unapproved promises before you press send.

LessonBeginner20 minutes6 min readUpdated 19 October 2026

Short answer

AI email writing is safe and useful when you decide the content and the assistant only drafts the words. Brief it with the facts and decisions already made, ask it to mark anything uncertain, then review every draft on five points: facts, promises, tone, privacy and whether you would sign it as written. A person sends every email; nothing goes out automatically.

  • You decide what the email says; AI decides how to say it.
  • Check facts, promises, tone, privacy and sign-off before sending.
  • Unapproved promises, such as dates, refunds or admissions of fault, are the costliest errors.
  • Send from your own mail program; never let a draft send itself.
  • Repeated edits belong in your prompt, not in every email.

Emails that carry weight with clients, suppliers or partners are where AI drafting saves the most time, and where an unchecked draft does the most damage. For owners, project managers, account managers and office staff, this lesson sets up a routine you can run in a couple of minutes on every AI-drafted email: brief it properly, review it on five points, and send it yourself. You need 20 minutes, an AI assistant and one real email you have to write today.

Who decides what, in an AI-drafted email

The routine rests on one split. You decide the content: the facts, the decisions, the commitments and what will not be said. AI drafts the words: structure, tone, clarity and length. When that split slips, and the assistant starts deciding content, emails go out with dates nobody agreed, apologies that admit fault, or offers nobody approved.

New Zealand guidance supports keeping a person in charge. MBIE’s quickstart for businesses says a human in the loop is needed where important decisions and actions are made or informed by AI (MBIE), and the Privacy Commissioner expects human review before acting on generative AI output (OPC).

The five-point review

Run these five checks on every draft before it leaves your outbox:

Check Question Typical AI error
Facts Does every name, number, date and price match my source? A date “corrected” to a different weekday, or a rounded figure
Promises Does it commit us to anything I have not approved? A firm finish date, a goodwill discount, “at no extra cost”
Tone Does it fit this relationship and situation? Over-apologising, admitting fault, or corporate stiffness
Privacy Is the recipient right, and is everything in it theirs to see? Details from another client pasted in as context
Sign-off Would I put my name to every sentence as written? Filler and claims about the business I cannot back up

The promises check matters most. A draft that says “we guarantee” or “the best in Tauranga” is your claim once you send it. The Commerce Commission’s guidance is that a business must be able to back a claim up when it makes it, not find the evidence afterwards (Commerce Commission).

Exercise: brief, draft, review, send

  1. Pick a real email. Choose one you need to send today that matters, such as a delay, a price change or a request.
  2. Write the brief. Who it is to, the purpose, the facts, decisions already made, and what must not be said. Use roles or first names only.
  3. Generate the draft. Ask the assistant to mark anything it is unsure of with [CHECK] instead of filling it in.
  4. Run the five-point review. Mark each problem you find and which check caught it.
  5. Edit and send it yourself. Copy the final version into your own mail program, re-read the recipient line, and send.
  6. Record your edits. If you made the same kind of edit twice, add a rule to your saved prompt in your prompt library.

Worked example: a Tauranga landscaper’s delay and variation email

Illustrative example: a landscaping company in Tauranga is building a retaining wall for a residential client. Rain has delayed the job, and the excavation found that extra drainage is needed behind the wall. The owner writes the brief:

Draft an email to a residential client (first name: Priya) from the owner of a
landscaping company in Tauranga.

Purpose: explain a weather delay and ask for approval of a variation.

Facts:
- Retaining wall job. Rain stopped work on 3 days last week.
- New expected completion: week of 2 November, weather permitting.
- During excavation we found groundwater seeping from the bank. The wall needs an
  extra drainage layer and pipe behind it.
- Variation cost: $1,850 plus GST. We need written approval before we do it.
- Without it, we would not be comfortable warranting the wall's drainage.

Do not: promise a firm finish date, offer a discount, or apologise for the rain as
if it were our fault. Do not add claims about our company.
Tone: friendly, straightforward, not alarming. Under 170 words.
Mark anything you are unsure of with [CHECK].

Example output (illustrative and abridged), with the owner’s review notes:

Hi Priya,

A quick update on your retaining wall. Last week's rain stopped work on three days,
so we now expect to finish in the week of 2 November, weather permitting.

While excavating, we found groundwater seeping from the bank. To keep the wall
properly drained, we recommend adding a drainage layer and pipe behind it. This is
a variation of $1,850 plus GST. We'd need your written approval before going ahead.
Reply to this email to approve, or give me a call to talk it through on [CHECK: phone].

Cheers,
[Name]

Review notes: the facts match the brief; no firm date or discount was added; the tone is calm. The owner fills in the phone number and adds one sentence explaining why drainage matters for the wall’s life. One earlier draft had said “we’ll have it finished by Friday 30 October”; the facts check caught it, and “do not promise a firm finish date” is now a permanent line in the saved prompt.

Where AI email drafting goes wrong

  • The draft decides the content. Without a brief, the assistant fills gaps with plausible commitments. Brief first.
  • Unapproved promises. Firm dates, discounts and “no extra cost” are the costly errors. They get their own check for that reason.
  • Admitting fault by accident. Polite apologies can read as accepting responsibility. Decide the position yourself.
  • Reply-all and the wrong recipient. AI does not see your recipient line. You do.
  • Automatic sending too early. Even production systems keep people in charge of what goes out: the property management agent we built for APL Property handles email triage and follow-up while staff keep the decisions.

Customer service teams answering high volumes should also read using AI for customer support drafts, and our AI quality review checklist extends the five checks for team-wide use. For drafting inside Outlook, see AI and Microsoft 365. Teams adopting the routine together can practise it on their own correspondence in our AI training for business teams.

Check your routine

  • I wrote a brief with facts, decisions and things not to say before generating.
  • The draft marks uncertain items with [CHECK] rather than guessing.
  • I ran all five checks: facts, promises, tone, privacy, sign-off.
  • Nothing in the email commits us to something I have not approved.
  • I sent it myself from my own mail program after checking the recipients.
  • Repeated edits went into my saved prompt.

Next lesson: Researching a customer account applies the same discipline to preparing for a sales conversation.

Sources

  1. Generative Artificial Intelligence, Office of the Privacy Commissioner, 15 June 2023
  2. Quickstart: Businesses using AI tools, Ministry of Business, Innovation and Employment, accessed 19 October 2026
  3. Making accurate claims, Commerce Commission, accessed 19 October 2026

Written by the Levy Street team in Wellington. Last updated 19 October 2026.

Questions people ask

Is it okay to use AI to write business emails?

Yes, for most routine and semi-routine emails, provided a person decides the content and reviews the draft before sending. The risks are specific: wrong facts, promises nobody approved, a tone that does not fit the relationship, and personal information in the wrong place. A short, consistent review catches nearly all of them. Use a business AI account if the email involves client details.

Should we tell recipients that an email was drafted with AI?

For ordinary correspondence that a person has reviewed and sent, most businesses do not, in the same way they do not disclose spellcheck or templates. The sender remains responsible for every word. Disclosure matters more when AI materially shapes advice, when a customer asks, or when your own AI policy or client contracts require it. Be honest whenever someone asks directly.

Can AI send emails automatically?

Technically yes, through automation tools and email integrations, but it should not be your starting point. Automatic sending removes the review step that catches wrong facts and unapproved promises. Start with drafts a person sends. Consider automatic sending only for narrow, low-risk messages, such as booking acknowledgements, after you have measured the error rate on a large sample of reviewed drafts.

How do I stop AI emails sounding generic?

Give it your voice: three words describing your tone, phrases you use and phrases you never use, and one or two past emails you were happy with as examples. Ban the usual filler, such as “I hope this email finds you well”. Ask for short sentences and a specific next step. If you keep making the same edit, add it to your saved prompt.

Want this working across your team?

Our team training turns these lessons into habits, with exercises built on your own documents and tools.