AI for Nanny Agencies: Family Briefs, Candidate Shortlists and Check-Ins

A nanny agency can use AI to turn a family call into a structured brief, to lay out each candidate's evidence against the family's written criteria, and to draft placement check-in messages. To avoid unfair screening, the criteria must be job-related and written first, AI must not score or rank people, and a recruiter decides every step.

The temptation is to ask an AI to "find the best three nannies for this family". Resist it. A model asked for a verdict will lean on whatever correlates with past successes, and some of those correlations are proxies for age, nationality, religion, family status or health. The safer design uses AI for the tedious parts, such as tidying notes, matching evidence to criteria and drafting messages, and keeps the judgement with people who can explain it.

Follow me on Instagram@sagnikteaches

Why a nanny agency has a higher bar than most

Three things set this work apart from ordinary recruitment. The people placed work alone in family homes with children, so safeguarding checks and references matter more than speed. The clients are private households, who often arrive with personal preferences, some of which cannot lawfully be acted on. And the data is unusually sensitive: children's details, home addresses, identity documents and background-check results, held about both families and candidates.

Connect on LinkedInSagnik Bhattacharya

The safe pattern therefore starts with a short list of what AI will never touch. Write it down before you try any tool:

Subscribe on YouTube@sagnikteaches

• Identity documents, right-to-work evidence and background-check outcomes.
• Any information about a child's health, behaviour or needs beyond what a brief needs, and never a child's name.
• Candidate health, pregnancy, family plans, religion, ethnicity, age, nationality or marital status, which no process should use to choose or rank people.
• Safeguarding concerns and references' sensitive comments, which go to a named person on your team.
• Final decisions on who is introduced, who is placed and whether a placement continues.

Step one: the family brief template

A good brief is a one-page record of what the family needs in job terms. It is the foundation for everything else, because the shortlist is judged against it. Capture it on the call, then let AI tidy your notes, never invent missing parts.

A brief template with the reasons built in

Notice the split between must-haves and nice-to-haves, and the rule that only job-related items can exclude. This is the single most useful fairness step in the whole process. Many families will volunteer preferences that are really about who a person is. A recruiter's job is to turn each one into a genuine requirement or to decline it. "Needs to speak Spanish because the grandparents do" is a language requirement. A preference about someone's age, background or appearance is not something to feed into any process, human or automated.

Prompt: tidy call notes into the brief

Turn these recruiter call notes into a family brief using EXACTLY the
sections below. Use only what is in the notes. If a section has no
information, write NOT STATED. Do not add requirements.

If the notes contain any preference that relates to a person's age,
gender, ethnicity, nationality, religion, health, family status or
appearance, do not put it in the brief. Instead list it under
FOR RECRUITER REVIEW with the exact wording used.

Sections: Role / Hours and days / Children (age bands only) / Duties /
Must-have requirements / Nice-to-haves / Pay and benefits / Household notes

NOTES
[paste anonymised notes: no surname, no address, no child's name]
END

The "FOR RECRUITER REVIEW" step is the safeguard. Instead of quietly passing a discriminatory preference through, the model puts it in front of a human, who handles it with the family: explaining which requirements the agency can act on, and which it cannot. The recruiter should also agree the final brief with the family in writing.

Handling the family preference conversation: a before and after

Wording matters more than most agencies expect. A recruiter who simply says "we can't do that" leaves the family feeling judged and likely to take the request elsewhere. A script that moves to the underlying need works better. Here is an illustrative exchange.

Family: We'd prefer someone younger, she'll need to keep up with a two-year-old.

Weaker reply: We can try to find someone younger.

Better reply: We can't choose by age, but we can look for what you actually need. What does a typical afternoon with your two-year-old involve? If it is lots of active outdoor play and carrying, I can add "comfortable with active outdoor play, including lifting a toddler" as a requirement and check it with every candidate.

The better reply turns an attribute into a job description, which the shortlist can test fairly, and it is the pattern your recruiters should rehearse. Add the rehearsed lines to your rule sheet so new staff use the same words.

Step two: shortlist notes built on evidence, not scores

Once the brief exists, the work is comparing candidates' evidence to it. AI can help read a CV or profile and lay out where each criterion is supported. It should present evidence, point to where it came from and say "no evidence found" rather than guess. It should not produce a number, a star rating or a ranking.

The evidence table

For each candidate, produce a table like the one below. This is an illustrative example for a part-time role with two children under five, needing a driving licence and infant experience.

The fourth column is crucial. It makes the AI's job "find and quote", and the recruiter's job "verify". Nothing in the table decides anything. A recruiter reads it and chooses whom to call, using the same table for every candidate so no one gets a different standard.

Prompt: build the evidence table

You are preparing an evidence table for a recruiter. Compare the
candidate document to the criteria list.

For each criterion give: the exact words in the document that support
it (quote them), where they appear, or NO EVIDENCE FOUND.

Rules:
- Do not give scores, ratings, rankings or recommendations.
- Do not comment on the candidate's name, photo, age, nationality,
  religion, health, family situation or career gaps.
- Do not infer anything that is not written down.

CRITERIA
[paste the must-haves and nice-to-haves from the brief]

CANDIDATE DOCUMENT (personal details removed)
[paste]
END

Strip names, photos, dates of birth and addresses before pasting; your applicant system or a quick manual redaction can do it. Removing identifying details also reduces what the model could lean on, although it does not guarantee fairness, because other parts of a CV can correlate with protected characteristics.

Check for bias the way you would check anything else

Fairness is something you test, not something you declare. Try these checks every month, on a sample.

1. Swap test. Take a CV, change the name and any age-revealing details, and run it again. The evidence table should be identical. If it differs, stop using that prompt.
2. Gap test. Run a CV with a career gap and one without. The table should not comment on the gap.
3. Outcome review. Look at who your recruiters actually introduced over the last quarter, and whether any group of candidates is consistently dropped at an early step. If the pattern appears, review the criteria and the humans, not just the software.
4. Reason log. For every candidate not taken forward, record the job-related reason in one line. If you cannot write one, the decision needs rethinking.

The longer treatments in bias checks every recruiter should run and can AI screen CVs fairly apply almost directly to nanny placements.

How a "strongest three" request dropped two candidates

In this illustrative case, a small agency asked a chatbot to "shortlist the strongest three of these ten nannies" and sent the result to the family. The three all came from the same background, with long continuous careers and recent experience in similar households. The agency later realised that two candidates with equal qualifications, who had taken career breaks to raise children of their own, had been quietly dropped. Nothing in the output said why. The fix was to stop asking for a shortlist, build the evidence table for everyone, and have a recruiter choose with written job-related reasons.

Step three: placement check-in messages

Placements fail quietly. A nanny feels undervalued, a family has an unspoken complaint, and by month three someone leaves. Regular check-ins catch problems, and drafting them is a safe use of AI because the output is a human-sent message, with no decision attached.

A check-in schedule

Here is an illustrative draft for the month-one message to a family, written from a prompt that supplies the brief's duties and hours and asks for a warm, plain note with two questions and an offer of a call:

Hello [first name], it has been about a month since [nanny first name] started, and I wanted to check in. Is the role working in the way you expected, including the hours and the school-run routine? Is there anything you would like to change or talk through? A reply here is fine, or I can ring you on a day that suits. Thank you, [your name], [agency name].

What would you check before sending? That it names no child, that it does not ask a leading question, and that it does not promise a fix. Messages that ask open questions get fuller answers than "Is everything OK?". Keep the personal conversation as the real check-in, with the message as a prompt.

What happens when a reply shows a worry

Set a rule: any reply that mentions a child's safety or wellbeing, an accident, a dispute about pay, or a nanny's welfare is read the same day by a named person and handled by phone, not by AI-drafted text. A carer retention check-in plan gives a useful pattern for frequency and tone, and you can borrow its exit-conversation approach for placements that end.

Data protection: the checks before any tool is connected

Both families and candidates trust you with personal data, and parts of it are sensitive. Before you use an AI tool with any of it, work through these checks, and ask your own adviser which data-protection rules apply to you.

1. Minimise. Share only what the task needs, with names, addresses and children's details removed where possible.
2. Know the terms. Where does the vendor store the data, for how long, and does it train models on it? The business plans of ChatGPT, Claude and Gemini in Google Workspace say they do not train on business content by default; confirm it for your own plan and for any recruitment software. The privacy-terms checklist lists what to ask.
3. Tell people. Your privacy notice should say that you use software, including AI tools, to help process applications and family briefs, and who sees the data.
4. Keep humans in the decision. In many places, people have rights about decisions made solely by automated means. A process where a recruiter makes every decision is both safer and easier to defend.
5. Retention. Delete candidate documents and call notes after the period your policy sets, including the copies in AI tools.
6. Access. Limit who can see candidate data, and use separate logins, not a shared one.

If you place candidates with families in the EU, AI used to select or screen candidates is classed as high risk under the EU AI Act. The rules for those systems apply from 2 December 2027, a date set by later amendments to the Act. Check the official timeline and take advice, because the dates have already moved once. Whatever the timetable says, written criteria, human decisions and a reason log are the sort of controls the law is moving towards. The children's data tutorial covers the child-data side in more depth.

One part-time placement, from first call to month one

Say a small agency with two recruiters takes a call from a family needing a part-time nanny for two children under five, three days a week, with a driving licence and first-aid training required. Times are illustrative.

1. Call and brief (30 minutes plus 10 minutes of AI tidying). The recruiter takes notes. The model tidies them into the template and lists one item under FOR RECRUITER REVIEW: the family mentioned wanting someone "young and energetic". The recruiter explains to the family that the agency can only act on job-related requirements and asks what the role physically involves. The answer, carrying and chasing toddlers and running around a park, becomes "comfortable with active outdoor play", a nice-to-have.
2. Candidate pool (20 minutes). The recruiter selects 12 candidates from the agency's register who are available on those days and hold a driving licence, using the database's filters, not AI.
3. Evidence tables (30 minutes). The assistant builds 12 tables from redacted CVs. Three show NO EVIDENCE FOUND for first aid, and one for infant experience.
4. Recruiter review (45 minutes). The recruiter reads all 12 tables, rings six candidates to verify gaps and availability, and writes one line of job-related reasoning for each of the six not taken forward.
5. Introductions. Three candidates go to the family, with the recruiter's own summary of each, checked against the evidence table.
6. Check-ins. The week-one and month-one messages are drafted by AI, approved and sent by the recruiter, and two replies lead to phone calls.

Total recruiter time on the shortlist is about two hours, against perhaps four or five if every CV is read cold and every note typed by hand. These figures are an illustration, so measure your own before and after. The final decision about whom the family meets stays with the recruiter every time.

Tools for a small agency, and the order to add them

You do not need specialist software to try this. A general assistant on a business plan, such as ChatGPT, Claude or Gemini, can run the three prompts above, with a person pasting redacted text in. Check each vendor's current pricing page for plans, since they change. If your agency already uses a recruitment or CRM system with AI features, ask the vendor how it handles candidate data and whether its screening features rank or score people, which is what you want to avoid. Setup and safeguards for CV screening covers that decision for agencies.

Begin with the family brief, since it carries the least risk and fixes the criteria for everything after. Add evidence tables once the swap test passes on your own CVs. Add check-in messages last. If you would like to design this with someone, that is what my AI implementation consultation covers.

Nine rules every recruiter follows

• The brief is written first and the criteria are job-related.
• AI organises evidence; it never scores, ranks or recommends.
• Personal details are removed before any text is pasted.
• Preferences about who a person is are never acted on, and are raised with the family.
• Every candidate not taken forward has a one-line, job-related reason on file.
• The swap test runs monthly and outcomes are reviewed quarterly.
• Safeguarding concerns and sensitive references go to a named person, never into a prompt.
• Check-in messages are sent by a person, and worrying replies get a phone call that day.
• Candidate and family data is deleted on schedule, including from AI tools.

Questions nanny agency owners ask about AI

Can AI rank nanny candidates for us?

It can organise evidence against criteria you have written down, but you should not let it produce an overall score or rank. Ranking hides the reasoning and invites the model to lean on proxies such as names, gaps or schools. Use it to build a side-by-side evidence table, and let a recruiter decide who goes forward and why.

What should never go into an AI tool?

Identity documents, background-check results, health information, children's names and details, home addresses and anything a family told you in confidence. Keep brief and shortlist drafting to anonymised, minimal information, and check the data terms of any tool before you use it for candidate or family data.

What if a family asks us to avoid certain kinds of candidates?

Do not pass discriminatory preferences into any process, human or AI. Separate genuine job requirements, such as a driving licence, a language, first-aid training or experience with a newborn, from preferences about who a person is. If you are unsure where the line sits, take advice from an employment lawyer familiar with the rules that apply to you.

Does the EU AI Act affect a small nanny agency?

If you place candidates with clients in the EU, AI used to select or screen candidates is a category the Act treats as high risk, with obligations that apply from 2 December 2027 under the amended timetable. Check the official timeline and take advice. Whatever the law says, human decision-making and documented criteria are good practice.

How often should an AI check-in message go out?

Use fixed points, for example after the first week, the first month and the third month, and keep each short. An AI-drafted message should invite a reply and offer a call, never replace the personal conversation. Any concern about a child's safety goes to a named person at once, not into a message thread.

Further reads

• How to Write Job Adverts With ChatGPT Without Biased Language — Write job adverts without biased language.
• Caregiver Recruiting Strategies: AI for Ads, Screening and Follow-Up — Caregiver recruiting with AI, from a home care angle.
• How Staffing Agencies Automate Candidate Follow-Up With AI — Keep candidates informed without chasing by hand.
• Caregiver Onboarding Checklist: Pre-Start Checks and Induction With AI — A pre-start checklist for carers joining a placement.
• AI Governance for a Small Business: Who Decides, Approves, Checks — Who decides, approves and checks AI output in a small firm.
• Nursery Staff Training: A One-Hour AI Briefing on Children's Data — A one-hour briefing on AI and children's data.
• AI Bias in Small Business Decisions: Hiring, Pricing and Credit — The wider picture of AI bias in small business decisions.
• AI Staffing Tools for Employee Engagement: Keep Temps on Assignment — A check-in cadence for temp assignments: opt-in texts, AI reply sorting into fine, call and urgent, and a same-day route to a consultant.
• AI for Senior Placement Agencies: Family Intake, Matching and Follow-Ups — An intake and matching workflow for senior placement owners: structured family intake, shortlists from checked facility data and follow-up messages.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: EU AI Act Service Desk timeline and Annex III (recruitment systems high risk; rules apply from 2 December 2027), and the business-data pages of OpenAI, Anthropic and Google Workspace, all checked 4 November 2026; employment-law and safeguarding rules vary and must come from your own adviser and regulator.