Review an AI resume screener and write the decision memo
Register it, tier it, ask the vendor questions, check fairness and write a go, conditions or stop decision.
What you will have at the end
An inventory record, a risk tier with reasons, a vendor question list, a fairness check on sample numbers, a control plan and a decision memo. It shows a hiring manager that you can run a real review.
What you need
- A document editor
- A spreadsheet
All names, numbers and data in this lab are made up for practice. Never use real customer or personal data. Your ticks are saved in this browser.
The scenario
Northline Foods is a fictional company with 900 staff. The hiring team wants to buy an AI tool that scores resumes and ranks applicants for warehouse jobs.
You are the AI governance lead. You must review the request, decide on the risk tier, ask the vendor the right questions, test fairness on the sample data and write a decision.
This is practice. The vendor, the numbers and the policy are made up. This lab is not legal advice. Say where you would ask a lawyer.
Starter material
Copy these into your own files or chat. Change anything you like.
What the vendor says.
TalentSift: AI that ranks resumes in seconds. - Scores every applicant from 0 to 100 and shows the top 10 percent first. - Trained on millions of resumes. Learns what good looks like from your past hires. - Removes names to reduce bias. 99 percent accurate. Fully automated screening available. - Hosted in the cloud. Data kept as long as needed to improve our models.
The company's short draft AI policy.
1. Approved AI tools are listed by IT. 2. Do not paste customer personal data into unapproved tools. 3. A person must check AI output before it is used for an important decision. 4. High-risk uses need review by the AI governance lead before launch. 5. Report problems to ai-governance@northline.example.
Fake outcomes of the vendor's pilot on 400 past applicants.
GROUP APPLICANTS SHORTLISTED BY TOOL SHORTLIST RATE Men 220 66 30 percent Women 180 36 20 percent Age under 40 240 72 30 percent Age 40 plus 160 30 19 percent (For practice only. Real checks need larger samples and legal guidance.)
Decision memo template.
DECISION MEMO: {{system}} DATE: {{...}} BY: {{...}}
SUMMARY (3 lines): {{what it is, the risk tier, the decision}}
WHAT WE REVIEWED: {{...}}
KEY RISKS: {{privacy, fairness, transparency, vendor, legal}}
CONDITIONS FOR GO: {{numbered list; who owns each; by when}}
DECISION: GO / GO WITH CONDITIONS / STOP DECIDER: {{...}}
REVIEW DATE: {{...}} OPEN QUESTIONS FOR LEGAL: {{...}}The steps
Do them in order. Each step says how you know it is done.
- 1
Register the system
- Fill in an inventory record: purpose, owners, data, who is affected, decisions it shapes, human check.
- Note which claims in the brochure you cannot yet verify.
- Write the questions you would ask the hiring team about how they will use it.
Done when: A complete inventory record and a list of unverified claims.
- 2
Set the risk tier
- Answer the tier questions: does it affect people's jobs, money or rights? Does it use personal data? Can mistakes be undone?
- Choose the tier and write the reasons.
- Say which laws or rules you would ask legal about, such as rules on automated hiring in the regions you hire in.
Done when: A tier with reasons and a list of legal questions.
- 3
Ask the vendor questions
- Choose the ten most important questions from the vendor question list, and add three about this tool: training data, bias testing and appeal.
- For each of the brochure's claims, write the evidence you would ask for.
- Write five contract points you would insist on.
Done when: Thirteen questions, evidence requests and five contract points.
- 4
Run the fairness check
- Compare shortlist rates across the groups in the sample.
- Compute the ratio of the lower rate to the higher rate for each pair, and describe the gap in plain words.
- Write likely causes, what you would check next and what you cannot conclude from small samples.
Done when: A comparison table, a plain-words reading and a list of next checks.
- 5
Design the controls
- Write preventive, detective and corrective controls: for example a person reviews all rejections, a monthly fairness check, and a way for candidates to appeal.
- For each control, add an owner and a way to test it.
- Write what a real human review would need: time, authority, training and records.
Done when: A control table with owners and tests.
- 6
Write the decision memo
- Fill in the memo template. Decide: go, go with conditions or stop.
- Write numbered conditions with owners and dates.
- Write the open questions for legal.
Done when: A one-page memo with a decision and conditions that a team could carry out.
- 7
Explain it to the team
- Write a short message to the hiring team: what you decided, why, and how to move fast on what is allowed.
- Offer a faster path: a lower-risk use such as drafting job ads.
- Write what you would measure after launch to know the governance program works.
Done when: A friendly message that offers a path forward, and three program measures.
Success checklist
How to know the lab is really done
Show it in an interview
- Present the decision in two minutes, including the one claim you did not believe.
- Explain how you would make human review real, not a rubber stamp.
- Say where your role ends and where you hand over to a lawyer.
Common mistakes
- Saying no without offering a path
- Believing the accuracy claim without evidence
- Treating a human in the loop as full protection
- Giving legal advice instead of asking legal
Want a bigger challenge?
- Write the policy line that would have prevented this request from skipping review.
- Design a quarterly fairness report for the live tool.
- Draft a candidate notice that explains the use of AI in plain words.
Next step: practice the interview questions for this role →
Other labs
A small agent that looks up orders and opens tickets, with checks, limits and a test set.
A researcher, a writer and a reviewer pass work to each other. You add limits and test the failures.
Play the forward deployed engineer for a fictional clinic: discover, scope, prototype, test and hand over.