By process

AI in recruiting: faster on the admin, never on the decision

Good candidates take the offer that came first. Most of the days you lose aren't spent on judgment. They go to sorting applications, scheduling and writing up interviews. AI can give those days back, as long as the criteria are written down first and a person makes every call about a person.

Written criteria firstStructured interview kitsScheduling without the back-and-forthConsistent interview summariesNo rejection without a human

Where the days go

Four patterns that slow hiring down without making it better

The pile

Hundreds of applications to read by hand before anyone gets a first call. The strongest candidates are rarely the patient ones.

Scheduling back-and-forth

Seven emails to find thirty minutes across three interviewers’ calendars. Each round costs a day, and the candidate notices every one.

Every interviewer writes differently

Notes arrive in different shapes and lengths, so two candidates for the same role cannot be compared on the same scale.

“Good” lives in a manager’s head

The criteria are spoken, not written. They shift between managers, and sometimes between Monday and Friday.

What we do

The decision about a person stays with a person

Criteria before tools. No rejection without a human. And no scoring trained on who you hired before, because a model that learns from past hires learns their bias along with them. We speed up the admin around the decision; we don't decide who gets hired, and we don't let any system decline a candidate on its own.

Define the criteria before the tool

Each role becomes a scorecard of written, job-related criteria the hiring manager signs off on: must-haves, nice-to-haves, and what evidence counts for each. The same scorecard drives the job description, the interview kit and the summary.

In practice. Hiring manager intake call → the AI drafts criteria and a job description from the notes → a one-page role scorecard → the hiring manager and HR edit and approve it, and HR checks the wording for exclusionary language.

Build screening against written criteria

The AI reads each application against the scorecard and says which criteria are met, quoting the line in the CV that shows it. It does not rank people by how much they resemble past hires. A recruiter reviews every flag, and no candidate is declined unless a person has seen the application.

In practice. New application in Greenhouse, Lever, Workday or Ashby → the AI maps the CV to each written criterion → a criteria checklist with evidence on the candidate record → the recruiter advances, holds or declines, and owns the decision.

Build interview kits, scheduling and summaries

Each stage gets a structured interview kit built from the scorecard: questions, what a strong answer looks like, and a rating guide. Scheduling runs from interviewers’ Google or Outlook calendars. After each interview, notes become a summary in the same structure for every candidate.

In practice. The interviewer’s notes → the AI structures them against the kit → a summary per criterion in the ATS → the interviewer corrects it and submits their own rating. The AI never rates.

Control candidate communication and bias

Status updates, reminders and prep notes go out on time from templates your team approved, and the hiring team gets a daily digest in Slack or Teams. Every quarter, compare the AI’s flags with recruiter decisions and look at pass-through rates by stage.

In practice. Stage change in the ATS → the AI drafts the candidate message → routine updates go out from approved templates, everything else waits in the recruiter’s queue → a person sends every decline.

The rules

Hiring is one of the few places the law names AI directly

This is not legal advice, and your counsel should confirm what applies to you. But these four are the ones that come up most, and the build should respect them from the start.

EU AI Act

Lists AI used for recruitment and selection, including filtering applications and evaluating candidates, among high-risk uses (Annex III). The date those obligations start has moved since the Act passed, so check the current one with counsel. Delayed is not removed: if you hire in the EU, design for it now.

NYC Local Law 144

If you use an automated employment decision tool for candidates or employees in New York City, it needs a bias audit within the year before use, a public summary of the results, and advance notice to candidates.

GDPR Article 22

People have the right not to be subject to a decision based solely on automated processing that significantly affects them. A rejection made by software alone is the textbook case. The safeguard is a person who actually reviews, not one who clicks approve.

Your own policy

Candidate data is personal data. What goes into which tool, how long it is kept and who sees it belongs in your AI usage policy, not in one recruiter’s habits.

Who does what

Every task has an AI role and a human decision

The split we start from. Your version will move a line or two, but the right-hand columns never go empty.

Task AI role Human decision Risk to watch
Job description Drafts from the scorecard Hiring manager approves Exclusionary or inflated requirements
Application screening Maps the CV to written criteria Recruiter advances or declines Proxy bias, so never auto-reject
Interview kit Drafts questions and a rating guide Hiring team edits Questions drifting off the criteria
Scheduling Finds slots, sends invites None for routine slots Calendar access scope
Interview summary Structures notes per criterion Interviewer submits own rating Summary standing in for judgment
Candidate messages Drafts updates and reminders Recruiter sends every decline Wrong message to the wrong person

How we work

What your hiring team holds at the end

Most teams start with an AI implementation sprint: one sprint from the first scorecards to screening, scheduling and summaries running in your ATS, with recruiters trained to run them. The connections use your ATS's own integrations or MCP connectors. For a longer program across HR and the rest of the business, the AI Partner engagement puts forward deployed engineers with your teams, with a working result every two weeks.

Role scorecards

Written criteria for each open role, approved by the hiring manager, reused across the job description, interviews and summaries.

An ATS workflow

Screening against criteria, scheduling and candidate messages running in Greenhouse, Lever, Workday or Ashby, with a person at every decision.

Interview kits and a summary template

Structured questions and rating guides per stage, and summaries that let you compare candidates on the same scale.

An audit trail and a quarterly review

Who decided what, on which evidence, and a regular check of AI flags against human decisions. Useful for any audit, and for your own peace of mind.

Policy comes with the build, not after it. On our AI training and engineering enablement engagement with a PE-backed cybersecurity company, an AI Acceptable Use Policy and responsible-use guardrails, including confidential-data handling, were set up alongside the rollout. Candidate data deserves the same care.

When this fits

If two of these sound familiar, it's worth a call

This is the pain of companies that hire steadily and lose good people to slow days.

  • Good candidates drop out halfway through the process
  • One recruiter reads hundreds of CVs by hand
  • Two interviewers’ notes on one candidate can’t be compared
  • What a good candidate looks like changes between managers

FAQ

AI in recruiting, answered

Won’t this introduce bias?

It is the central risk, which is why we start with criteria, not tools. A system that scores candidates on patterns from past hires reproduces past bias, faster. We build screening against written, job-related requirements, require evidence for every flag, keep every rejection with a person, and review the results each quarter.

Can the AI reject candidates automatically?

Not in anything we build. The AI can say a CV does not show evidence for a criterion. A recruiter decides what that means. Fully automated rejection is where the legal exposure sits, under GDPR Article 22 and the EU AI Act, and it is also where good candidates get lost for bad reasons.

Which applicant tracking systems do you work with?

Greenhouse, Lever, Workday and Ashby are the most common, and each has an API. We connect through the ATS’s own integrations or through MCP connectors that give a Claude, ChatGPT, Copilot or Gemini agent scoped access to the candidate record, with the permissions your admins set.

What about candidate privacy?

Candidate data is personal data in every sense. We define what enters each tool, how long it is kept and who can see it, use business plans with admin controls rather than personal accounts, and write it into your AI usage policy rather than leaving it to individual recruiters.

Do we have to tell candidates we use AI?

Often, yes. NYC Local Law 144 requires notice when an automated employment decision tool is used, GDPR requires you to tell people how their data is processed, and it is good practice regardless. We draft the notice language with your HR or legal team, and they approve it.

Can we record and transcribe interviews?

Only with consent and where local law allows it, and some places require every party to agree. Many teams start with the interviewer’s own notes instead, which gets most of the value and avoids the question.

Do you understand the roles we hire for?

No, your hiring managers do. We bring the AI and the process. They bring the judgment about what good looks like in the role, which is why the scorecard is theirs and every decision about a person stays with them.

Where do your hiring days actually go?

A short call, no deck. We'll look at your funnel, where the delays sit, and say honestly whether we can add anything.