How to Use AI in Hiring for a Small Shop: Ads, Questions, Scorecards

Use AI to draft, not to decide. Give it the real duties and must-haves, have it write a plain-language ad and five screening questions, then build a scorecard with fixed 1-5 anchors before any CV arrives. Every candidate gets the same questions and the same scale, and a person reads the answers and makes the call.

Most small shops don't need software that ranks applicants. A single counter vacancy brings in a pile you can read in an evening. What costs time, and what goes wrong, is the preparation: a vague ad, questions invented on the day and judgements made on gut feel by the fifth interview. AI is quick at that preparation, which means you can finish it, and lock it, before the first application lands.

Follow me on Instagram@sagnikteaches

What AI does in this kit and what stays with you

The kit has three documents: a job ad, five screening questions and a scorecard. AI helps draft each one. The judgement calls stay with a named person. The split is deliberate. Language models are good at turning notes into tidy wording and poor at judging people, and in some places the law treats software that filters or ranks candidates as high-risk. This is who does what.

Connect on LinkedInSagnik Bhattacharya

One rule governs the table: nothing in the right-hand column gets delegated, however good the draft looks. The rest of this tutorial fills in each row for one illustrative vacancy.

Subscribe on YouTube@sagnikteaches

The counter vacancy used throughout

The example is an independent paint and hardware shop with six staff, hiring one part-time counter assistant. The owner has no HR department and interviews alongside the shop manager. Before opening an AI tool, write the role brief down in a table. Everything the AI produces gets checked against it.

Three choices in that brief matter for fairness. The must-have list is short, and each line is something the job really requires, which stops you and the AI adding wish-list items that quietly exclude people. The lifting requirement is a task with a number and the equipment provided, where "must be physically fit" is vague enough to deter people who could do the work. And the "not required" row matters as much as the rest, because AI drafts add experience and qualification lines to sound credible unless you say otherwise.

Draft the ad from the brief, then mark it up

Paste the brief into ChatGPT or Claude, ideally inside a Project so the brief stays attached to the conversation (both offer Projects, including on their free plans). Tell the AI to use only what the brief says. Models fill gaps with plausible filler unless you forbid it.

You are helping a small independent shop write a job advert.

Use only the facts in the brief below. Do not add benefits, values,
team descriptions or requirements that are not in the brief. If the
advert needs something the brief doesn't give, list it at the end as
questions for me instead of inventing it.

Write in plain language. Structure: a one-line summary, what the job
involves (bullets), what you need (five bullets at most), what we
offer, how to apply. Describe tasks, not personalities: no words such
as energetic, young, digital native or fast-paced.

Finish with a line inviting applicants to tell us about any
adjustments they need to apply or to interview.

BRIEF:
[paste the table here]

Even with those rules, the first draft comes back with flaws. This one is illustrative:

Counter Assistant (part-time), paint and hardware shop

We're a friendly, well-established shop with a supportive team, and we
are looking for a counter assistant for about 20 hours a week (Thursday
to Saturday, plus occasional weekday cover). Pay is $17 to $19 an hour.

What the job involves
- Serving customers at the counter and on the phone
- Helping people work out what they need, using our product guides
- Running the till and cashing up at close
- Mixing paint and cutting keys (full training given)
- Receiving deliveries and shelving stock in a fast-moving shop

What you need
- Strong communication skills
- Accuracy with money and stock counts
- Reliability on your scheduled shifts, including Saturdays
- Enthusiasm for learning about products
- Able to carry stock up to 15 kg using the trolleys and aids we provide

What we offer
- Two weeks of shadowing, a written product guide and a named mentor
- Opportunities to progress

How to apply
Answer five short questions on our form by [date]. Tell us if you need
any adjustments to apply or to interview.

Questions for you
- What is the closing date?
- Is there a staff discount you want listed?

It is close, and it still needs five edits before it goes anywhere.

1. "Friendly, well-established shop with a supportive team." None of that is in the brief. The model added it because job ads usually contain it. Replace it with a fact you can defend, such as "six of us work on the shop floor", or cut it.
2. "Opportunities to progress." You haven't offered any. Promising progression in an ad creates an expectation you may be held to, so delete it.
3. "Fast-moving shop." Your prompt banned pace words and one still slipped in. If the counter really is busy, say what happens: "Saturdays are our busiest day, with a short queue at the till at peak times."
4. "Strong communication skills." This is the vague version of a real requirement. Say what the communication involves: "explaining products to customers who may not know what they need."
5. The closing questions. The AI did the useful thing and asked for what it lacked. Answer them in the ad: a closing date and, if it is true, the staff discount.

The most important fix is the one where a personality becomes a task. Before: "You'll be a confident, energetic communicator who thrives in a fast-moving environment." After: "You'll explain products to customers who may not know what they need. On Saturdays there is often a short queue at the till." The second version tells an applicant what the day looks like and lets them decide whether they can do it, which is all an ad should do.

Put the pay range in. A growing number of places now require it, and applicants filter by it whether the law demands it or not. For a deeper wording audit, including a second check in a fresh chat, see this tutorial on writing job adverts without biased language.

Five screening questions for the application form

Screening questions go on the application form, so every applicant answers the same five in writing before anyone reads a CV. Each question should test one thing on the scorecard, all five should take under ten minutes to answer in total, and a person with no retail experience must be able to answer every one. If your job board lets you add application questions of your own, use that. Otherwise a plain Google Form or Microsoft Form does the same job.

Here is the role brief for a part-time counter assistant at a paint and
hardware shop. Write five application-form questions, one for each
must-have. Rules: each question must be answerable by someone with no
retail experience; none may ask about age, family, health, nationality,
religion or anything the job doesn't need; keep each under 40 words.
For each question, describe what a strong, an adequate and a weak
answer would contain.

After editing, the set below is what I would send. The "strong answer" column is what you will score against, so write it before the first application arrives.

The raw AI output usually needs three corrections. It tends to include "Why do you want to work here?", which tests nothing on your card and rewards rehearsed answers. It offers "Are you a team player?", which is a leading question anyone can answer yes to. And it often asks "How would you cope with standing all day and lifting?", which can screen out people with disabilities who could do the job with an adjustment. Replace that last one with a plain yes-or-no question on the real task, kept apart from the five scored questions: "The job includes carrying stock up to 15 kg using trolleys and lifting aids. Can you do this, with or without adjustments?"

Some questions should never appear on the form, however the AI words them:

• Date of birth, graduation year or "how long have you been working?"
• Plans for a family, childcare arrangements or marital status (ask about availability instead)
• Health conditions or "any reason you couldn't do physical work?"
• Where someone is originally from, or whether a language is their first
• A photo upload

Only the availability question and, where the law requires you to check it, the question on eligibility to work can knock someone out automatically. Everything else is scored by a person against the anchors in the next section. If you want the answers tidied into one table by candidate number, an AI can do that. It does not score them.

A scorecard with anchors, locked before the first CV

A scorecard is a list of criteria, each with a weight and a description of what a 1, a 3 and a 5 look like. Those descriptions are called anchors, and they stop a "3" meaning one thing to you and another to your manager. Ask the AI to draft anchors from the brief, then edit them until they describe what a good counter assistant in your shop actually does. For the longer method, including probing follow-up questions, see building interview questions and scorecards with AI.

The weights sum to 10, so the maximum weighted total is 50. Helping customers carries the most weight because it is most of the job. Decide that before you meet anyone, not after you've fallen for a particular candidate.

Write the tie rule now too. Mine: if two candidates finish level on total, the higher "Helping customers" score wins; if that is level too, the higher "Accuracy" score wins; if that is level as well, the owner decides and writes the reason against a criterion.

Then lock it. Save the scorecard, weights and tie rule as one file with the date in its name, and send a copy to your co-interviewer. From that moment, any change means re-scoring every candidate you have already seen. That one rule is the heart of a fair process, because it stops the criteria drifting towards whoever you happen to like.

Fairness checks that fit a six-person shop

A shop can't commission a bias audit, so the checks have to be cheap and routine. These seven fit around the working day.

1. Read the ad as an applicant. Ask the AI: "List any phrase in this advert that might make someone with a disability, caring responsibilities, no retail background or over 50 think they can't apply, and say why." Treat the answers as suggestions. Models over-flag (it may object to "stand at a counter"), so you decide which are real.
2. Trace every criterion to the brief. Ask: "For each scorecard criterion, quote the line in the brief that requires it, and list any criterion with no matching line." Criteria such as "presentable", "good fit with the team" or "recent experience" fail this test, and they are the ones that carry bias.
3. Use numbers, not names. Add an ID column to the form export (C01, C02) and hide the name, email, phone, photo and any date-of-birth columns for the first read. Share only the ID and the answers with an AI. Names come back after scoring.
4. Same questions, same order, same length. Give every interviewee the same 25 minutes: the five questions again with follow-ups, then the same counter role-play. Don't score the chat that happens around it. Tell everyone in the invitation that adjustments are available, and mean it.
5. Score alone, straight afterwards. Fill in the sheet before the next candidate walks in and before you speak to your co-interviewer. Next to each score, write one evidence line in the candidate's own words: "said she'd recount, then check refunds, didn't mention telling the manager."
6. Check any AI summary against your notes. The failure here is quiet. A raw note says: "Said he'd count the till twice, didn't mention refunds." An AI tidy-up returns: "Demonstrates a thorough, methodical approach to cash handling." The summary has turned a 3 into a 5. If you let AI tidy notes, compare every line with the original, and keep the original as the record.
7. Test the tidy-up for style bias. Take one real answer and ask a colleague to rewrite it with the same content but rougher wording: shorter sentences, a spelling slip or two, no polish. Run both versions through the same prompt in fresh chats. If one summary makes the candidate sound more capable, the tool is rewarding polish rather than content, which penalises people who write less fluently. Stop using it, or restrict it to copying answers verbatim into the table. The test is crude and catches only blatant effects, but it takes two minutes.

Two habits sit alongside the checks. Write one line of reasons for every "no" at every stage, citing a criterion, and keep it with the scorecard. And look at your own knock-outs: if one must-have, such as being free for the Thursday-to-Saturday pattern, eliminates most of the pile, ask whether it is truly essential or whether a rota change would widen the field.

Scoring five candidates: the sums and the decision

Illustrative numbers: 38 applications arrive. Three can't work any of the Thursday-to-Saturday pattern, which is the one knock-out the owner agreed in advance, so 35 are read in full. At about three minutes each, that is roughly 105 minutes for the owner, spread over two evenings. The owner scores the written answers against the anchors and picks the top five to interview, including anyone tied at the cut-off. After the interviews, each interviewer scores alone and the sheet is merged.

One disagreement shows how the rule works. For candidate C07's "Accuracy" criterion, the owner scored 4 and the manager scored 2. A gap of two or more sends both back to their evidence lines. The owner had written "recount, then refunds, then colleague"; the manager had written "didn't say who she'd tell." Re-reading, the follow-up answer did name the manager, so the agreed score became 4. Neither person won the argument. The notes did.

C07 and C19 tie on 41. The pre-agreed tie rule gives it to C07, whose "Helping customers" score is 5 against 4. Now the person decides. The owner reads the evidence lines for both, checks a reference for the leading candidate with permission, and writes the decision note below. The offer is the owner's, supported by the sheet. The sheet doesn't make it.

DECISION NOTE
Role: Counter assistant. Scorecard version: 2026-10-14.
Offer to: C07. Weighted score 41/50 (tie with C19 broken by the
agreed rule: Helping customers 5 v 4).
Evidence: asked four diagnostic questions in the tap role-play and
checked the fix with a colleague; till answer named every step.
Concerns: Learning products scored 3; plan a written product test
in week two.
Not used: impressions from small talk before or after interview.
Decided by: [owner]. Date: [date].

What the owner did not do is pick C12 because "she reminds me of me". That feeling is real, and it is what the scorecard exists to keep out of the decision. If a gut reaction points somewhere the sheet doesn't, write it down, check whether it maps to a criterion, and if it doesn't, leave it out.

The unsuccessful candidates get a short message. AI can draft the template, and you check each one is true. After the form: "Thank you for applying. We had many strong applications and haven't been able to invite you to interview this time. We'll delete your details by [date] unless you ask us to keep them for future roles." After an interview, offer feedback only if you can base it on the scorecard lines, never on AI-written reasons you haven't checked.

What AI should not touch in this round

• Ranking or scoring applicants. That moves the judgement into a system you can't inspect. This tutorial on screening CVs fairly explains why shortlists drift.
• Reading faces or voices for emotions or personality. Tools that infer emotions are banned at work in the EU (see the next section), and I wouldn't trust any tool that claims to read personality from a video.
• Searching for applicants online and summarising what it finds. It surfaces irrelevant personal detail and sometimes the wrong person.
• Rejecting anyone automatically. That includes auto-reply rules that act on an AI's judgement of an answer.
• Writing reasons for a "no" that aren't in your notes. A fluent invented reason is worse than no reason.
• Setting an individual's pay. Use the range in the ad.

Hiring-law points to check where you recruit

I'm not a lawyer, and the rules differ by place and change often, so treat this as a list of questions to put to an employment adviser or to the official guidance where you recruit.

• Discrimination rules cover the ad, the form, the interview and the decision. The protected characteristics vary by place, but age, sex, race, disability and religion are widely covered, and questions about them are the riskiest part of the form.
• Pay ranges. A growing number of places require a range in the advert.
• Software that screens or scores candidates. Some places require you to tell applicants, and a few require a bias audit. This tutorial on which local laws reach a ten-person employer covers one market in detail.
• If you recruit in the EU, the AI Act lists AI used to place targeted job adverts, analyse and filter applications or evaluate candidates as high-risk (Annex III). The main duties for those systems have been pushed back to December 2027, and AI that infers emotions at work, which includes recruitment, has been banned since February 2025. This tutorial explains the high-risk rules for hiring tools.
• Candidate data is personal data. Tell applicants what you do with it, keep it only as long as you need it, and agree a deletion date with your adviser.

Keeping AI to drafting and tidying, with a person making every decision, stays well away from the uses those rules target. Confirm that with an adviser if you recruit in a regulated market.

Time, cost and how to tell it worked

Here is a rough sum, using my own estimates, so time your first attempt. Writing the ad, questions and scorecard by hand might take an owner three hours. With AI drafting and the owner editing, it is closer to an hour, because most of the time becomes checking. The real saving arrives next time: you keep the locked files, change the brief, and start from a tested kit. The tools can be free. ChatGPT and Claude both have free plans, and their paid individual plans have listed at about $20 a month if you want higher limits. A form and a spreadsheet do the rest.

Check whether the kit worked at two points. Straight after the round, look at your scores: if every candidate got a 3 on a criterion, its anchors are too loose to separate anyone, so tighten them for next time. Ninety days after the start date, ask the manager how the new hire is doing and compare it with their interview scores. One hire proves nothing, but over a few hires you will see which criteria predict a good counter assistant in your shop, and you can reweight the card accordingly. Keep the locked scorecard and the decision note for each round, and read the stage-by-stage hiring process for a small team if you want to extend the approach to more roles.

Questions owners ask before they start

Can I paste applicants' answers into ChatGPT or Claude?

Strip identifying details first. Give each applicant a number, keep names, emails, phone numbers and photos in a separate column, and paste only the numbered answers. Consumer plans have a model-training switch in privacy settings, and business plans don't train on your content by default. Even so, a number instead of a name costs nothing and removes the question.

What if applicants use AI to write their answers?

Expect some will, and don't rely on AI-text detectors. OpenAI withdrew its own in July 2023 because of low accuracy: it caught 26% of AI-written text and wrongly flagged 9% of human writing. Ask for something only the applicant would know, such as a specific moment from their own week, and test the skill live at interview with a short counter role-play.

Do I have to tell applicants I use AI in hiring?

It depends on where you recruit and what the AI does. Some places require notice when software screens or scores candidates, and the EU treats such tools as high-risk. Drafting your own ad or tidying your own notes is different from analysing applicants, but your privacy notice should still say which tools see candidate data. Ask an employment adviser, and when in doubt, say so plainly in the how-to-apply line.

Can I change the scorecard after the first interviews?

Only by starting the scoring again. If you add or reweight a criterion halfway through, early and late candidates are judged on different terms. If something important is missing, write down why, save a new dated version and re-score everyone interviewed so far from your notes. Keep both versions with the decision record.

How many interviewers does a small shop need?

Two is better and one is workable. With two, each person scores alone straight after the interview and you talk only once both sheets are written; a gap of two points or more on any criterion sends you back to the evidence notes. With one interviewer, score immediately, write the evidence line, and ask a colleague to read the notes for your top two.

Further reads

• How to Write Job Descriptions With AI That Attract Good Hires — Longer method for outcomes, pay and must-haves in a job description.
• AI Bias in Small Business Decisions: Hiring, Pricing and Credit — Where bias creeps into hiring, pricing and credit decisions.
• AI for a Hardware Store: Can It Answer "What Do I Need for This Job?" — What AI can and cannot answer at the counter you are hiring for.
• Retail Staff Scheduling: Build a Rota From Hourly Till Sales With AI — Build the new hire's rota around hourly till sales.
• Restaurant Job Descriptions: Can AI Write a Server or Line Cook Ad? — ChatGPT can draft a server or line cook ad in a minute. Here is what a hiring owner must check on shifts, pay, tips and wording before posting.
• AI Interview Software for Video Interviews: A Hiring Checklist for Trades — A hiring checklist for trades firms: which AI interview features to use, which to avoid, and the notice, consent and bias checks to run first.
• 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 text (Annex III, point 4, and the Article 5 ban on inferring emotions at work); OpenAI's notice withdrawing its AI text classifier (July 2023); OpenAI and Anthropic help pages on Projects and privacy; published guidance on structured interviews.