Generative AI: Recruitment and HR Tech

Generative AI in recruitment and HR tech

Recruitment is one of the few areas where regulators have named AI hiring tools specifically. The EU AI Act treats them as high risk and New York City requires bias audits. That does not make generative AI off limits. It means the system has to be designed so recruiters decide, candidates are treated fairly, and every recommendation can be explained. We build CV parsing, matching, screening agents and HR helpdesks that way from the first sprint.

Get a free assessment Recruitment and HR tech software

Key highlights

We build CV parsing, explainable candidate matching, candidate engagement and scheduling agents, interview summaries and HR policy helpdesks.
Recruiters and hiring managers make every shortlist and rejection decision. The AI ranks, explains and drafts.
Matching uses hard requirement filters plus embeddings, with an LLM rationale so recruiters see why each candidate was surfaced.
We build in impact ratio reporting for bias audits, candidate notices and GDPR rights, so you are ready for NYC Local Law 144 and the EU AI Act.
A measured prototype usually takes about 4 weeks, and a production feature inside your ATS 10 to 16 weeks.

Why recruitment AI projects stall

Recruitment teams are keen on AI because the volume is real. Projects stall on fairness, explainability and candidate data rules, or because the feature lives outside the ATS recruiters actually use.

Keyword search that misses good candidates. Boolean search returns hundreds of CVs that mention the right words and misses the candidate who described the same skill differently. Recruiters fall back on their personal shortlist.

Black box scores nobody can defend. A vendor tool gives each candidate a number. When a candidate or regulator asks why they were rejected, nobody can explain it, so legal blocks the rollout.

Bias risk that no one has measured. Models trained on past hiring decisions can repeat past patterns. Without impact ratio testing by sex and race or ethnicity, nobody knows whether the tool is fair until a complaint arrives.

CV data scattered across tools. Candidate data is copied into a matching tool, a chatbot and a spreadsheet. A subject access or erasure request now touches four systems and nobody knows all of them.

Screening and scheduling eat the day. Recruiters spend hours on availability, right to work questions and interview scheduling, so high volume roles go unanswered and good candidates accept other offers.

HR teams answering the same questions. Leave, payroll and policy questions arrive by email and chat all day. Answers vary by who replies, and the handbook nobody reads is already out of date.

Use cases

Generative AI use cases we build for recruitment and HR

Each use case supports a recruiter, hiring manager or HR adviser, with the decision and the accountability staying with them.

CV parsing and enrichment

LLM extraction of CVs in any layout into a typed candidate schema with skills, roles, dates and qualifications normalised to your taxonomy, de duplicated against your ATS such as Bullhorn or RecruiterFlow, with protected characteristics excluded from matching fields.

Explainable candidate matching

Hard requirement filters first, then embeddings on pgvector to rank by relevance, then an LLM rationale that cites the CV lines behind each match. Recruiters see why a candidate was surfaced and make the shortlist themselves.

Candidate engagement and scheduling agents

Chat, SMS, WhatsApp or voice agents that answer role questions from the job pack, collect availability and right to work basics, and book interviews into calendars. Knock out questions are deterministic and any rejection is a recruiter decision.

Interview notes and summaries

With candidate consent, interviews are transcribed and summarised against the structured scorecard, with evidence quotes per competency. Interviewers score, the AI only drafts the notes.

Job descriptions and outreach

Drafts of job adverts, inclusive language checks and personalised outreach from the role brief and the candidate profile, written in your brand voice and sent by the recruiter.

HR policy helpdesk

RAG over your handbook, policies and benefits documents, answering employee questions in Slack, Teams or your HR portal with the source cited, and handing sensitive topics such as grievances or health straight to an HR adviser.

How we deliver

From one desk to a governed rollout

01
Use case discovery and data readiness
We map the hiring workflow, ATS data quality, jurisdictions involved and which steps could count as automated decisions. Output: a written scope, a risk assessment per use case and acceptance criteria including fairness thresholds.
02
Evaluation harness and prototype
A golden set of real, anonymised CVs and roles is labelled with your recruiters. Two or three models are tested for relevance, extraction accuracy, cost and impact ratios across groups before anything reaches a live candidate.
03
Guarded pilot with recruiters in the loop
A small desk uses the feature on live roles alongside its current process. Every ranking, override and candidate interaction is logged, and candidate notices and opt outs are switched on from day one.
04
Production rollout with monitoring
Rollout inside your ATS with dashboards for match acceptance, recruiter overrides, impact ratios and cost per hire. Bias testing is rerun on a schedule and whenever the model or data changes.

AI governance for recruitment and HR

Hiring is a regulated use of AI in several jurisdictions. We design so that your compliance evidence is produced by the system, not assembled before an audit.

EU AI Act (employment high risk) NYC Local Law 144 bias audits UK GDPR Article 22 ICO AI in recruitment guidance Equality Act 2010 (UK) Title VII and EEOC (US) Illinois AI Video Interview Act CCPA and CPRA

Hiring AI is high risk in the EU. The EU AI Act lists AI used to place job adverts, filter applications and evaluate candidates as high risk, with duties on risk management, data governance, logging, transparency and human oversight. We build those controls into the product and document them for your conformity work.

Bias is measured, not assumed. NYC Local Law 144 requires an independent bias audit and candidate notice before an automated employment decision tool is used for New York City roles. We build impact ratio reporting by sex and race or ethnicity categories into the system so your independent auditor has the data, and we rerun fairness tests whenever the model changes.

People decide, and candidates can ask why. UK and EU GDPR restrict decisions based solely on automated processing, and the ICO has published recommendations for AI recruitment tools. Recruiters make every decision, matching excludes protected characteristics, candidates are told how AI is used, and access, correction and erasure requests flow through every system holding their data.

Models and tooling we use for recruitment AI

Strong extraction and reasoning models, embeddings you can explain, and integrations with the ATS, calendar and messaging tools recruiters already use.

Claude GPT Gemini Llama (self hosted) Embedding models and pgvector LangGraph Bullhorn and ATS APIs Microsoft Graph and Google Calendar Twilio and WhatsApp Business Whisper transcription Microsoft Presidio PII redaction Langfuse tracing

FAQ

Common questions

How much does AI candidate matching or a recruiting agent cost?

It depends on the number of use cases and integrations, so we start with a fixed price prototype and price production from a written scope. Model costs are passed through at cost and reported per candidate processed or per conversation, which makes them easy to compare with recruiter time. You get a written estimate within 48 hours, and our pricing page explains how engagements are structured.

Is it legal to use AI to screen candidates?

Yes, in most places, with the right controls. The EU AI Act treats hiring AI as high risk, NYC Local Law 144 requires bias audits and notices, and UK and EU GDPR restrict solely automated decisions. We design so a recruiter makes every decision, candidates are informed, and the logging and fairness data your compliance and audit teams need are built in. We are engineers, not lawyers, so we work alongside your legal advisers.

How do you keep candidate data private?

Candidate data goes only to model endpoints with retention disabled and no training, or to self hosted models, and personal details are redacted where the task allows. Data stays in the systems you control, retention is configurable, and subject access and erasure requests reach every place a CV is stored.

Can AI matching be accurate without being biased?

It can be measured and improved, which is the honest answer. We filter on stated requirements, rank on skills evidence rather than proxies such as names or schools, exclude protected characteristics, and test impact ratios on a labelled set before go live and on a schedule after. If a model fails the fairness thresholds, it does not ship.

How long does it take to add generative AI to our ATS?

Usually 10 to 16 weeks for a production feature. A measured prototype takes about 4 weeks, followed by a pilot with one desk on live roles and then rollout inside your ATS. CV parsing or a scheduling agent alone can be faster.

Should we build AI matching or buy a recruiting AI tool?

Buy if a tool fits your ATS, your markets and your compliance needs. Build when you are an ATS or HR tech vendor making AI part of your own product, when your roles and taxonomies are specialised, or when you need full control over bias testing and candidate data. We can also wrap bought components with your own explainability and audit layer.

Keep exploring

Related services and industries

Free consultation, no commitment

Ready to give recruiters their day back?

Tell us where your hiring workflow slows down. Written scope, timeline and cost estimate within 48 hours.

Book a free consultation All industries