AI candidate communication drafting.

Why candidate messages are worth automating carefully

Candidate communication is high volume and easy to neglect. Most candidates hear nothing, or get a generic regret weeks later, and those who interviewed often get no useful feedback. AI drafting helps because the facts already exist: the applicant tracking system knows the role, stage and dates, and the interviewer’s approved notes contain the evidence for specific feedback. The model’s job is to turn those facts into a clear, kind message in your tone. It is not to decide anything.

The useful split is between messages that can go on approved templates and messages that need a person. Acknowledgements, interview logistics and screening regrets can be templated with careful merge fields. Post-interview feedback, offers and replies to upset candidates need a recruiter to read and send.

How the workflow runs

Inputs

Candidate stage and history from the applicant tracking system, the role details, approved interview summaries, the organisation’s tone guide, approved templates and a vetted FAQ.

Systems

The applicant tracking system or recruitment CRM, email and SMS, the calendar for interview booking, and a language model for personalised drafts.

Steps

Detect a stage change; choose the message type; fill the template or draft a personalised message from notes; check recipients and merge fields; check for banned content; queue for approval or send if templated.

Human review

Recruiters approve every personalised draft. A senior recruiter approves template changes and reviews a weekly sample of sent messages. Anything about pay, visas, complaints or distress goes to a person directly.

Outputs

Sent messages logged against the candidate record, drafts awaiting approval, and escalations.

Measures

Days from decision to candidate told, share of interviewed candidates given specific feedback, recruiter edits per draft, candidate replies asking for clarification, and complaints.

Risks

Invented feedback, wrong recipient or role in a merge, content touching prohibited grounds, and marketing sent without consent.

Illustrative example: an Auckland in-house talent team

Illustrative example, synthetic numbers: an Auckland employer with 600 staff receives about 2,500 applications a month across 30 roles. Each month it sends 300 interview invitations, 300 post-interview regrets with feedback and 30 offer-stage messages, plus templated acknowledgements and screening regrets.

Only the 630 personalised messages use a model; templated ones do not need it. Assuming 4,000 input tokens per draft (notes, role, template and tone guide) and 350 output tokens on OpenAI’s gpt-6.1-sol at US$2 per million input and US$10 per million output tokens, as listed on 19 October 2026, each draft costs about 1.2 US cents, or about US$7 a month. The measure that matters to candidates is days from decision to message, so track it.

NZ rules that apply to candidate messages

• Confidentiality. The Privacy Commissioner’s recruitment guidance asks employers not to do anything that might reveal a person has applied, and not to use recruitment information for other purposes without express consent. A bulk email with visible recipients, or a draft that mentions another candidate, does both.
• Fair reasons. Under sections 22 and 23 of the Human Rights Act 1993, employment decisions and inquiries must not turn on prohibited grounds such as age, sex, family status, disability or ethnic origin. Feedback should be specific and job-related, and drafts should be checked for any reference to those grounds.
• Marketing messages. The Unsolicited Electronic Messages Act 2007 covers messages that market or promote goods or services. If a job alert or talent pool email does that, it needs consent, accurate sender information and a functional unsubscribe; an unsubscribe takes effect within 5 working days, and the sender must prove consent if challenged.
• Accuracy. The Privacy Act 2020 requires reasonable steps to make sure information is accurate before it is used. A regret sent to the successful candidate because the stage field was wrong is an accuracy failure, not just an embarrassment.

What recruiters keep

Recruiters own the decision, the feedback content and the relationship. They decide what to say to a disappointed internal candidate, how to handle a candidate who discloses a personal difficulty, and when a phone call is better than an email. Our lesson on drafting emails with human review covers a quick review routine that catches most problems.

Where candidate messaging fails

• Invented feedback. Without notes, a model will produce plausible advice (“consider building stakeholder skills”). Block personalised feedback when no approved summary exists.
• Merge errors. Wrong first name, wrong role, or the hiring manager’s internal comment pasted into the message. Validate every merge field against the record before sending.
• Visible recipients. One CC instead of BCC reveals a whole shortlist. Send one message per candidate.
• Tone mismatch. A cheerful template sent to a candidate who withdrew for family reasons reads badly. Route withdrawals and sensitive replies to a person.
• Stage lag. Messages fire from stale stages because the tracking system was updated late. Make sending depend on a confirmed stage change; our lesson on connecting AI to your CRM covers the triggers.

Feedback quality depends on interview note summarisation, and accurate stages and consents depend on recruitment CRM data cleanup. See the recruitment sector guide and our AI workflow automation service for the full build.

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