How to Use AI for Recruitment: CV Screening and Initial Interviews

The average mid-market founder spends 12–15 hours per hire on CV review and first-round screening. Almost none of that time requires human judgment. Most of it just requires consistency; which is exactly what AI is good at.

This is not about buying an ATS or running algorithmic screening that makes your lawyers nervous. It is about using the tools you already pay for; Claude, ChatGPT, whatever is open in your browser today; to cut screening time by 70% without missing good candidates. This article gives you the exact workflow.

Why most companies still screen CVs manually — and what it actually costs

The time math that most founders have never actually run:

That time sits almost entirely with the founder, the COO, or a single HR generalist who is already doing everything else. And it is largely not judgment work. It is pattern-matching against criteria that could be written down: years of experience, specific tools, industry background, red flags in tenure or gaps.

The second cost is consistency. Manual CV review drifts over a 60-CV pile. The founder who reads CVs on Monday morning applies different standards than the same founder reading CVs on Friday afternoon.

AI does not drift.

What AI can and cannot do in your hiring process

Setting accurate expectations before the how-to begins means you do not over-automate and miss good candidates; and do not under-automate and miss the point.

What AI handles well:

• Scoring CVs against an explicit, pre-defined rubric
• Flagging hard disqualifiers: missing required experience, tenure patterns worth questioning, tool gaps
• Summarising each CV against the scoring criteria so the hiring manager reads a one-paragraph brief rather than a two-page document
• Generating a ranked shortlist with a score and rationale for each candidate
• Producing tailored interview question sets based on each shortlisted candidate’s specific background
• Drafting rejection and progression emails in the company’s voice

What AI does not do:

• Assess culture fit or interpersonal dynamics
• Make the hire decision
• Catch the unusual-but-brilliant candidate whose CV does not match the standard pattern (this is why human review of the shortlist matters; not just the top-ranked CVs)
• Replace a conversation that reveals how someone thinks under pressure

The rule of thumb: AI handles the desk work of hiring. Humans handle the room work.

Step 1 — Build the scoring rubric before you touch a single CV

The most common reason AI CV screening fails is not the AI. It is that the criteria were never explicitly defined. The rubric is the foundation of the entire workflow; and it can be built in under 30 minutes.

A scoring rubric has four components:

1. Must-haves (hard disqualifiers)

Binary. If a candidate does not have them, they do not move forward regardless of anything else. Keep this list short; three to five items maximum. Every item should be genuinely non-negotiable, not a preference dressed up as a requirement.

Example for an operations manager role:

• Minimum 3 years in an operations or project management role
• Experience managing a team of at least 5 people
• Based within commutable distance of [location] or open to relocation

2. Strong indicators (weighted positives)

Score them 1–3 based on how much weight they carry.

Example:

• Experience in [your specific industry]: 3 points
• Has managed AI or automation implementations: 2 points
• Has scaled a team during a growth period: 2 points
• Familiarity with [your specific tools; HubSpot, QuickBooks, etc.]: 1 point

3. Yellow flags (not disqualifiers but worth noting)

These do not eliminate a candidate but should be visible in the shortlist summary so the hiring manager can decide how to probe. Examples: multiple short tenures under 12 months; unexplained gaps; managed very small teams only; industry background very different from yours.

4. Red flags (strong disqualifiers that require judgment)

Not automatic rejections, but patterns that significantly reduce confidence without a good explanation. Examples: three or more jobs in two years without a clear progression narrative; vague role descriptions with no specifics; no concrete outcomes mentioned anywhere in a senior-level CV.

Step 2 — Run the CV screen with AI

This section is an operational protocol. Paste-and-use level of specificity.

The screening prompt structure:

You are screening CVs for a [role title] position at [company type/industry].

MUST-HAVES (candidate fails if any of these are missing):
- [Must-have 1]
- [Must-have 2]
- [Must-have 3]

STRONG INDICATORS (score each 1–3 as noted):
- [Indicator 1]: [score]
- [Indicator 2]: [score]
- [Indicator 3]: [score]

YELLOW FLAGS (note but do not eliminate):
- [Flag 1]
- [Flag 2]

RED FLAGS (note clearly):
- [Flag 1]
- [Flag 2]

For each CV I paste, produce:
1. PASS or FAIL on must-haves (with reason if FAIL)
2. Total indicator score out of [max score]
3. Yellow or red flags noted (or "none")
4. One-paragraph summary: what makes this candidate interesting or not for this role
5. Recommended shortlist tier: STRONG / CONSIDER / PASS

How to run it:

• Paste one CV at a time for roles with fewer than 30 applicants
• For larger volumes (50+): paste three to five CVs per session with a consistent separator between them (”---” works)
• Save every output in a simple spreadsheet: candidate name, tier, score, one-line summary, flags

The output is a ranked shortlist with rationale, not just a ranked list. The hiring manager reads the AI summary, not the full CV, until the final shortlist review.

The human review step that matters:

Before confirming the shortlist, the founder or hiring manager should read the CVs of every CONSIDER-tier candidate the AI flagged as borderline. AI screening can be conservative on unusual career paths. The human review catches the outlier the rubric did not account for.

Step 3 — Use AI to prepare for every interview on the shortlist

Once the shortlist is confirmed, the hiring manager typically spends 30–45 minutes per candidate preparing: re-reading the CV, deciding what to probe, writing questions. AI collapses this to under five minutes.

The interview prep prompt:

Here is the CV for [candidate name], who is interviewing for [role title] at our company.

Our company context: [2–3 sentences about the business, what the role does, 
what "good" looks like in the first 90 days]

Based on this CV, generate:
1. Three questions that probe their most relevant experience for this role
2. Two questions about the gaps or yellow flags you notice in this CV
3. One question that tests how they think about [specific capability most 
   important for this role]
4. The one thing in this CV you would most want to understand better before 
   making a decision

Format: numbered list with a one-line note on what each question is trying to reveal.

The output gives the interviewer a tailored question set in under a minute. The questions are grounded in the specific candidate’s background, not generic interview templates.

One note: review the AI-generated questions before the interview. The AI occasionally surfaces a question that is better as a follow-up than an opener. That takes 90 seconds to reorder.

What to do with the rejection and progression emails

Two email types, both handled by AI with minimal input. This is the part that takes founders longer than it should; and is the easiest to hand off.

Rejection emails:

Most companies send generic rejection emails or none at all. AI makes it easy to send a personalised one-paragraph rejection that acknowledges the candidate’s specific background.

Prompt: “Write a brief, respectful rejection email for [name], who applied for [role]. Their background was in [one-line summary]. We are moving forward with candidates who had stronger [specific gap]. Keep it warm, specific, and under 100 words.”

Progression emails:

The AI drafts a progression email that references something specific in the candidate’s background, confirms the next step, and sounds like a human wrote it; because the context is loaded.

The one rule: the founder reads and sends. AI drafts, human approves. Hiring communications carry relationship stakes. They are not automated to send without review.

The bias question — what founders actually need to know

AI screening against an explicit rubric introduces less arbitrary variation than tired-human screening of 80 CVs on a Friday afternoon. But it encodes whatever is in the rubric. If the rubric over-indexes on credentials from specific schools or companies, the AI will reflect that. If the rubric requires tools that only certain demographics have had access to, it will reflect that too.

Three practical rules for rubric integrity:

1. Every must-have should survive this test: “Could a genuinely excellent candidate for this role be missing this and still be the right hire?” If yes, it is a preference, not a must-have. Move it to indicators.

2. Every strong indicator should be directly connected to job performance; not proxy signals for socioeconomic background (school prestige, unpaid internship history, certain hobby activities).

3. Have one other person review the rubric before the first screen runs. Not to approve it; to catch anything that looks like a preference dressed as a requirement.

AI does not introduce bias on its own. It amplifies the bias already in your criteria. The rubric review is the safeguard.

Common questions from founders setting this up

”Do I need a special AI tool for this or does Claude work?”

Claude works. ChatGPT works. The tool matters less than the rubric loaded into it. A well-built scoring rubric in Claude produces better results than a vague brief in a specialist ATS tool. Start with what you have.

”What if we only hire twice a year — is this worth setting up?”

Yes; because the rubric is reusable. You build it once per role archetype and refine it each cycle. By the third hire for a similar role, the rubric is your institutional knowledge on what good looks like; not something you reconstruct from memory each time.

”How do I handle roles where the rubric is hard to define?”

Start with must-haves only. If you cannot define what disqualifies a candidate, you cannot evaluate them consistently; with or without AI. The rubric-building exercise is valuable regardless of whether you use AI; it forces clarity on what you are actually hiring for.

”Can I use this for senior leadership hires?”

Yes, with adjustments. For senior roles, the must-haves list is shorter and the strong indicators list does more work. The interview prep prompt becomes more valuable than the screening rubric; because shortlists for senior roles tend to be small and the prep work is where AI saves the most time.

”What about GDPR and candidate data?”

Do not paste identifying information (full names, contact details, national ID numbers) into a shared AI session if GDPR applies to your business. Use a candidate reference code in the prompt and keep the mapping in a separate document. The CV content itself; role history, skills, qualifications; is what the AI needs and is lower-risk than personal contact data.

”How do I store the rubric so it improves over time?”

Keep a single document per role archetype. After each hiring cycle, add a “lessons learned” section: which must-haves caught good candidates who turned out to be wrong hires; which indicators proved more predictive than their score suggested; which flags were red herrings. The rubric compounds with each use.

Want your hiring workflows documented and running before your next open role?

If you scored 0–2, the work starts with documentation; a context pack, a handful of documented workflows, and a shared workspace to put them in. That is the foundation. Everything else builds on top of it.

If you scored 3–4, you are closer than most. The shared workspace is probably the missing piece. One focused build; with a partner or internal lead who has the time to do it properly; and the system starts compounding.

Path one: build it yourself. Run the diagnostic, identify the three workflows that score highest on frequency and friction, and start documenting. You’ll learn more in 30 days of building than in six months of planning.

Path two: bring in a partner. If you want the Level 2 to Level 3 move done in weeks rather than months; with a shared workspace, documented workflows, and adoption tracking already installed; that is the work Phos does. The fastest way to know if it’s the right fit is a conversation. Thirty minutes, no deck. Start here.