Prepare application evidence for a human shortlist review

Example result.

Application B
1. Coordinate work between teams: "I worked between the warehouse team and customer service to resolve order queries." (Experience, E1)
2. Use a ticketing system: "We used a ticket system for incoming queries; I added delivery updates to tickets." (E2)
3. Deal directly with an external provider: "I chased our courier directly when collections were missed." (Supporting statement, S1)
4. Process change and what happened: "I suggested logging repeat queries…" (S2). Outcome not stated.

The assistant organises text. It must not decide who progresses, score anyone or infer personality.

1. Agree the source and remove what doesn’t belong

Before using real applications, confirm with HR and the data-protection lead what applicants are told, where outputs are stored and for how long. Use the same source set for every applicant; don’t add a social-media search for one person.

Select only the job-relevant text. Leave out health, equality-monitoring information and other sensitive personal circumstances before pasting. A prompt asking the model to ignore that material would still send it. Use an application reference instead of a name where your process allows, remembering that a distinctive work history can still identify someone.

Why this works: the safest way to keep sensitive information out of the answer is to keep it out of the chat.

What you should see: a referenced, job-relevant extract for each application, nothing else.

2. Make the criteria concrete, and fix them first

Use the job description and person specification. “Coordinate order queries between teams” gives the extraction a target. “Be a brilliant communicator” invites a general impression. Don’t change criteria after seeing who applied; if HR finds a problem, decide through the recruitment process how all affected applications are treated.

Why this works: the extraction can only look for what the criteria name, so vague criteria produce vague evidence.

What you should see: a numbered list of criteria, each describing something an application could show.

3. Extract one application at a time

Open a fresh chat. Paste the criteria, then one application’s text with section headings or line references.

Organise this application's evidence under these role criteria. This is clerical preparation for a human reviewer, not an assessment.

Criteria, in this order: [criteria]

For each criterion, reproduce relevant excerpts exactly, with the section or line reference. Keep qualifiers and who did the work. If nothing relevant is found, write "No relevant excerpt found; human source check required". Then list job-relevant lines you couldn't place.

Do not add outside information, infer characteristics or motives, score, rank, compare applicants or recommend progression. A missing excerpt is not evidence that someone lacks the skill.

Application reference: [reference]
Application text: [text]

[criteria] is your list from step 2; [reference] and [text] are one application’s reference and extract.

Why this works: one application at a time stops evidence mixing between people. Exact excerpts keep the difference between “helped with” and “owned”, which a polished summary can quietly erase.

What you should see: every criterion with quoted excerpts and references, or the “human source check required” line, plus a short unplaced list.

4. Check the original, in random order

Take the sheets in a random order: not the order they arrived, and never an order a tool produced. The ICO describes employers whose managers looked properly at “green” candidates and only glanced at “red” ones before rejecting them, and found that this counted as solely automated decision-making. Its point is that meaningful human review “must be applied to every candidate, not just those who score highly”.

Read every quoted excerpt against the source. For every empty criterion, read the relevant original sections: the evidence may be in different words or a different section. “Nothing found” describes the search; it doesn’t establish a lack of ability. If you notice your attention sagging partway down the pile, stop. That’s the mechanism the ICO describes, arriving on schedule.

Why this works: a thin extraction mustn’t quietly receive a thinner review.

What you should see: every empty criterion either filled from the original or confirmed empty by a person.

5. Record the reasoning in your own words

The authorised reviewers assess against the agreed framework themselves, ideally more than one person, as Acas recommends. For each application, write one or two sentences naming the criteria and the evidence relied on, in the normal recruitment record, separate from the AI-prepared sheet:

[Decision]. Criterion 1 evidenced (E1). Criterion 3 evidenced in the supporting statement (S1) rather than the experience section. Criterion 4: a suggestion is described; its outcome isn't stated — ask at interview what happened next.

Weak: “Not a strong fit.” It names no criterion and no evidence, so nobody can check it or explain it to the applicant.

Why this works: a note that names criteria and evidence can be reviewed and challenged; an impression can’t.

What you should see: a short reviewer note per application, with no generated suitability language in it.

Check the result

• Every quoted excerpt matches the application word for word, with its reference.
• Every empty criterion was checked against the original by a person.
• Nothing in the sheet scores, ranks, describes personality or recommends an outcome.
• Each reviewer note names criteria and evidence.

The usual failure: the extraction misses evidence written in different words, and the criterion shows as empty. Spot it by reading the original sections for every empty criterion, and the unplaced lines list.

The application includes this relevant line: [line and reference]. Recheck criterion [criterion] and revise its evidence using exact excerpts. Keep qualifiers. Identify any other affected section. Do not score, recommend or change the criteria.

[line and reference] is the missed line; [criterion] is where it belongs. If the same kind of miss keeps happening, do the extraction by hand. Running the same extraction again isn’t independent validation.

Fictional example.

Agreed criteria:
1. Coordinate work between teams.
2. Use a ticketing system in day-to-day work.
3. Deal directly with an external service provider.
4. Describe a process change and what happened after it.

Application B, Experience:
E1: I worked between the warehouse team and customer service to resolve order queries.
E2: We used a ticket system for incoming queries; I added delivery updates to tickets.

Application B, Supporting statement:
S1: I chased our courier directly when collections were missed.
S2: I suggested logging repeat queries so we could see the pattern.

The first sheet left criterion 3 empty and added, under criterion 4: “Shows strong initiative and excellent stakeholder skills.”

• The failure. It looked for “supplier” or “service provider” and missed “courier”. That’s an extraction failure, not the applicant’s vocabulary. And the criterion 4 line is a judgement the source doesn’t contain.
• The repair. The repair prompt with S1 placed it under criterion 3. The judgement was deleted, and S2 was quoted as a suggestion with its outcome unstated, giving the example result at the top of this page.
• What the sheet doesn’t do. It doesn’t say whether Application B progresses. The reviewer’s note in step 5 does that, in a person’s words.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Check this evidence sheet against the application text. Report only missed relevant excerpts, changed quotations, removed qualifiers and references that can't be found. Do not assess the applicant. Show the exact source for each proposed correction.

Sources

• ICO: Meaningful human involvement (Recruitment rewired)
• Acas: Choosing who to interview
• Acas: Following discrimination law
• Acas: Interviewing job applicants