Can You Sell Recruitment Data for AI Training?

Partly. Three buyers of company records name recruitment or staffing among the industries they take, though none names the records a recruiter keeps; the part that carries the value is the record of a recruiter's judgment (the job description, who was put forward or turned down and why, the interview feedback and what became of each candidate). A database of CVs on its own has no buyer that names it. Either sells only as part of the firm's working record, and only once the candidates in it can no longer be identified.

The software is rarely what stands in the way. The terms of the applicant tracking systems we read leave the candidate records with the recruiting business and give it an export, and none of the clauses we read bars the business from licensing its own copy. What decides a sale is whether the records show decisions and their outcomes, and what the law asks of a business that wants to reuse candidates' personal data for a purpose they were never told about.

This guide is for owners and directors of recruitment agencies, staffing firms, executive search practices and recruitment process outsourcing (RPO) providers, and for in-house talent teams thinking about their own hiring records. It covers what recruitment data is, which buyers name it, why the shortlist matters more than the CV, who already trains hiring models and why they are not your buyers, what your applicant tracking system's terms say, what the records are worth, the candidates and others in them, the rules that follow the data to the buyer, how to take the records out and prepare them, and two worked examples. It sits under what AI labs buy, beside the guides to CRM data, email archives and employee emails and chats.

Contents

• What recruitment data is
• Which buyers name it
• Why the shortlist matters more than the CV
• Who already trains hiring models, and why they are not your buyers
• What your applicant tracking system's terms say
• What recruitment data is worth
• The people in the records
• The rules that follow the data to the buyer
• Taking the records out and preparing them
• Two worked examples
• Questions people ask
• What this means for you

What recruitment data is

A recruitment business keeps two kinds of record, and buyers treat them very differently. The first is records about people: candidate profiles, CVs, contact details, work histories, the database a firm builds over years of sourcing. The second is records of decisions: a client's brief and the job description written from it, the search a recruiter ran, who was screened in and who was screened out, the notes on why, the candidates submitted to the client, the interview feedback, the offer, the placement, and how long the hire stayed. The first describes the market. The second describes skilled work being done, step by step, with an outcome at the end.

Most firms hold both in the same system. An applicant tracking system (ATS, the software that moves candidates through a hiring process) or a recruitment CRM links each candidate to the jobs they were considered for and records every stage change. Around it sit the firm's email with candidates and clients, its call notes, its interview scorecards, its terms of business with clients and, for a staffing firm, timesheets, assignments and pay rates for temporary workers.

The decision records are the ones buyers name, and they are also the hardest for anyone else to recreate: a CV can be found again, but the reason a recruiter turned a candidate down for a role in 2019, and whether the person hired instead was still there two years later, exists only in that firm's files.

Which buyers name it

Buyers of company records publish what they want on their own pages, and we read them for recruitment, staffing, hiring, candidate and HR records on 5 November 2026. The result: three buyers name recruitment or staffing, and none of them names the records a recruiter keeps.

Scale AI lists staffing and recruiting among its industries of interest, and names Workday among the business platforms whose HR records it takes. Centific lists staffing and recruitment among the industries on its form, and describes a people record that shows how employees are hired, onboarded and supported. Grepped's intake form lists recruiting and staffing among the industries it asks about.

Two more buyers touch the edge of the field. Polyshares takes HR and payroll systems such as Gusto, Rippling, ADP and BambooHR, asks only for record counts at first and de-identifies payroll detail before anything moves, and Troveo names the records behind HR tickets and approvals. Those are a company's records about its own staff, not a recruiter's records about candidates.

Nobody else we read names recruitment at all. Mercor publishes listings by kind of company, and none of them is for a recruitment or staffing firm, nor is the field among its target industries; the pages of Nyne, micro1 and several others name neither. That is not a rule that such records cannot sell. It means a recruitment firm is judged, by most buyers, as a professional services company whose whole working record (email, documents, CRM, finance) is what they price, with the recruitment records as the part that shows its skill.

The size lines matter as much as the names. Scale AI prefers 40 or more employees, and most other buyers that publish a floor start at 20 or more full-time employees and three years. A two-person search boutique falls under every floor we have read, and the guide to buyer programmes sets the floors side by side.

Why the shortlist matters more than the CV

AI companies are increasingly training models to do the work that happens inside companies, and recruiting is that kind of work: read a brief, search, judge, persuade, decide, follow up. A model learns that work from records that show each step and how it turned out. SuperAnnotate says it of company records in general: the reasoning matters more than the finished output. Centific's people record asks for the same thing in HR terms, the decisions, approvals, exceptions and documentation behind each case.

A CV is a finished output. On its own it says what a person did, not what a recruiter made of it. The value is in the join: this brief, these candidates considered, these screened out for these reasons, these submitted, this feedback from the client, this hire, still in post after a year. The last step, the outcome, is the one most firms cannot show, because their system records the placement but not what happened after. A staffing firm that knows which temporary workers were asked back, or an agency that tracks guarantee-period refunds, holds the outcome by another name.

Three properties raise the value of a recruitment record, and they are the same ones buyers name for company records generally:

1. History. Years of searches show how roles, skills and pay moved. Troveo puts it in one line: value scales with your systems, team size, and years of history.
2. Specialisation. A firm that has placed nurses, actuaries or chip designers for a decade holds judgments about those roles that few others have. Scale AI asks for real work, not demos or synthetic data.
3. Written by people. Job descriptions and candidate summaries written with a chatbot since 2023 are worth less than the ones written before, because models trained on generated text degrade unless real data stays in the mix. The guide to AI-generated data explains how buyers tell them apart.

What this means in practice: a firm should describe its records to a buyer as a decision record with outcomes, not as a candidate database. The second description invites the question no buyer wants to answer, which is whether it is buying personal data about millions of people.

Who already trains hiring models, and why they are not your buyers

A recruitment firm's first thought is often to sell to the companies that build recruitment AI. We read what those firms say about their training data, and they build on their own customers' records and on public profiles, and none we read says it licenses a recruitment firm's records from outside.

Eightfold, a talent intelligence platform, lists past applicants, hiring outcomes, stage definitions and ATS data among the enterprise data its platform uses, beside resumes, career pages and social profiles, and says the platform runs on a global data set of more than 1.6 billion careers. The enterprise data there is its customers' own, used under its contracts with them. The applicant tracking systems do the same under their terms: Greenhouse may train its models on customers' inputs and outputs in de-identified form only, and never on personal data; Workable uses and owns aggregated data that identifies no customer or person; and an agency that turns on Bullhorn's AI Assistant lets Bullhorn use its data in that feature to develop and improve it.

LinkedIn trains on its own members' data. Ireland's regulator said in November 2025 that it had not approved LinkedIn's use of members' data for AI training, but saw no need to intervene further after LinkedIn narrowed the data and added opt-outs. Profiles copied from LinkedIn into a recruiter's database are a different matter: when a firm that scraped LinkedIn settled with it, the judgment required it to permanently delete all the member profile data it held. A recruitment firm whose database was filled by scraping holds records it may be unable to license at all; the guide to scraping law covers why buyers avoid them.

The practical lesson is that the paying buyers for a recruitment firm's records are the AI data buyers above, which want working records to teach models how work is done, not the recruitment software firms, which already have a supply under their own contracts. A buyer that is building a hiring model of its own will still ask the questions a hiring model raises, which the rules section below sets out.

What your applicant tracking system's terms say

Before a firm offers its records, it should know what its software vendor's terms say about who holds them, what the vendor may do with them, and how they come out. We read the published terms of five systems that recruitment firms use. The finding: the records stay with the business, the business can export them, and the vendor keeps only de-identified or aggregated data for itself.

Three things follow. First, no clause we read bars a recruitment firm from licensing its own copy of its records to a buyer. That is a statement about these published clauses, not about every document a firm signs: an order form, a data processing agreement or an enterprise contract can say more, and Bullhorn's main agreement is not public, so a Bullhorn customer should read its own. Second, the vendor's aggregate is the vendor's: a firm can grant a buyer exclusive rights over its records, but not over what Greenhouse or Workable already derived from them. Third, the export window closes: a firm that is changing systems or winding down should export through the API while its subscription runs, because after it ends the records are deleted.

What recruitment data is worth

No buyer, listing or disclosed sale we found prices recruitment records on their own. Every buyer that names them prices the company's whole record, and the recruitment history is one part of it. The figures that apply are therefore whole-company figures.

Across the sales of company records on record, the median is $72.5K per company a company, and published offers to holders run at a median of $277K per company. The buyers' own ranges are wide: Handshake AI says it has paid $100K to $4M+ per partnership, Mode (BizBuyData) $100K to $5M a company, Nyne commonly $100K or more, Scale AI $10K to $1M+, and micro1 sets out tiers per partnership. None of those figures is for a recruitment firm, and none of the buyers says what a recruitment firm's records would add.

The one staffing company sale we hold a court record for was not a data sale. ShiftPixy, a staffing technology company, sold its software platform, app and brand in bankruptcy for $1,850,000, to a buyer that bought it to run, not to train models. It shows that a recruitment business's assets can find a buyer in a closing; it is not a price for candidate records.

A recruitment firm should therefore value its whole record, with the recruitment system as one of its systems, and see what that record shows. What raises the figure is what the buyers name: years of history, several systems that show the same work, a specialism, and records written by people. What lowers it is what they avoid: records about people with no decisions in them, records that cannot be cleaned of personal data, and records whose origin the firm cannot explain. The guide to what company data is worth sets out each factor, and the valuation applies them to what a firm holds.

The people in the records

Every recruitment record is about a person, and most are about people who never dealt with the buyer and never imagined their CV would train a model. A licence of recruitment records is a licence of personal data until those people are taken out. Four groups appear in them.

Candidates

Candidates are the largest group and the one the law protects most closely here, because they gave their data for one purpose: to be considered for jobs.

In the UK and the EU, licensing their records for AI training is a new purpose, and a business may reuse personal data for a new purpose only if it is compatible with the first. The UK now has the test in its own law: a business must weigh the link between the purposes, the context of collection, the nature of the data, the consequences and the safeguards, the same factors as the EU's Article 6(4). Commercial licensing is not among the purposes the UK treats as compatible automatically. The UK regulator's own reading is that a new purpose is likely to be incompatible if it is very different, unexpected or has an unjustified impact, and training an AI model on a job applicant's CV is hard to call expected. Where candidates gave their details directly, the firm must tell them before it processes their data for the new purpose; where it sourced them from a job board or a profile site, the exception for telling people applies only where telling them is impossible or out of proportion.

CVs carry sensitive data. A CV can mention a disability, a health condition, a religion, a trade union role or a nationality that reveals ethnic origin, and equal opportunities forms collect it on purpose. That data is prohibited from processing unless a narrow exception applies, and a firm should take it out before anything is licensed rather than rely on one.

The route that works in practice is de-identification: a lot from which no candidate can be identified is no longer personal data. The step of de-identifying is itself processing, and the UK regulator says it needs a lawful basis and a stated purpose, though it is likely to be fair. A firm relying on its legitimate interests has to show the processing is necessary, which a licence that could be done with de-identified records fails if it sends identifiable ones. The guide to GDPR and licensing data for AI training and the guide for UK holders go further.

In the US, California's privacy law does not exempt job applicants. Its regulator states that employees and job applicants are consumers with rights under the law, and the carve-out that once covered them has expired. Licensing their personal information to an AI buyer for money is a sale for a business over the law's thresholds, and a business that sells must first tell consumers and honour their opt-outs. Records that are properly deidentified fall outside those rules. Video interviews add another layer where a tool has analysed them: in Illinois no business may sell or otherwise profit from a person's biometric identifiers, such as a scan of face geometry or a voiceprint. And a promise in a candidate privacy notice not to share their data is one a firm has to keep: breaking it can be an unfair or deceptive practice.

In Canada, a paid licence is a commercial activity that the federal privacy law reaches, and the privacy regulator expects consent for new purposes and disclosures to third parties that use the data for their own purposes. The guide for Canadian holders covers it.

Clients and their hiring managers

A client's brief often holds more than the job: a salary band, a planned reorganisation, the name of the person being replaced, the reason a previous hire failed. Much of it was given in confidence under the firm's terms of business with that client, and the hiring managers named in it are people too. A firm should read its client terms for confidentiality and for who owns the materials produced in a search before it puts briefs in a lot, and strip client names and identifying detail, so that no buyer can read one client's plans in another's training data.

Referees and third parties

References name the referee and say things about the candidate a third party chose to share. Notes record what a candidate said about a former employer or a colleague. These are people the firm never dealt with in the role it now wants to reuse their words for, so they come out with the candidates.

The firm's own recruiters

Screening notes and feedback are written by the firm's own staff, and they are personal data about those staff as well. In the EU, consent from employees is unlikely to be freely given, so a firm should not rely on asking its recruiters to agree; it codes their names instead. Member states can add their own rules for employee data, and the guide to employee emails and chats sets out what an employer may do with its staff's records.

The rules that follow the data to the buyer

Some rules bind the buyer rather than the firm, and they shape what a buyer will ask for. The most important is the EU's AI Act. An AI system intended for the recruitment or selection of people, to filter applications or to evaluate candidates is listed in Annex III, and the Act says the systems in that annex shall be considered high-risk. The maker of a high-risk system must govern its training data with records of the origin of the data and, for personal data, the original purpose of its collection. Those duties apply to such systems from 2 December 2027.

The consequence for a recruitment firm is direct. A buyer that might use the records for anything touching hiring will ask where they came from, why they were collected and what was done to them, and it will need written answers it can keep. A firm that can state its systems, its years, the privacy notices in force, what it removed and why, is the firm whose records can be used. A firm that cannot is the one buyers walk away from.

In New York City, Local Law 144 already applies to the use of hiring tools. It prohibits employers and employment agencies from using an automated employment decision tool unless it has had a bias audit within one year, the results are public and candidates have been told. It binds whoever uses such a tool, not the firm that licensed records to its maker. A recruitment firm that uses AI screening in New York is covered as an employment agency in its own right.

Historical hiring records also carry the biases of the decisions in them. A buyer training a hiring model has to find and correct them, and a firm that records how its decisions were made (the criteria, the client's instructions, which stage each candidate reached) gives it what it needs to do so. That is one more reason the decision record is worth more than the CV.

Taking the records out and preparing them

Take the records out through the system's own API or export, while the subscription runs. Greenhouse gives an API key that exports Customer Data at any time during the term, at no cost, Lever tools to pull it during the term, and Workable tools to export or retrieve it until termination. After the agreement ends the window closes: Lever's data is promptly deleted and Greenhouse's is queued for deletion. A firm that is closing should export before it cancels, and the guide to how long to keep business records after closing covers what it must keep.

Before any recruitment records go into a lot:

1. Name each system and its years. Which ATS or recruitment CRM held the records, and from when to when. A firm that moved systems has two exports with different fields, and the join between them is part of the work.
2. Read the terms in force for those years, for who holds the records, the vendor's licence and aggregate, and any AI feature that was switched on. Note what the vendor already holds, so a buyer hears it from you.
3. Join the records into decisions. Each job to its brief, its candidates, every stage change, the notes and feedback at each stage, and the outcome. The joined record is the one buyers name; a flat export of candidates is not.
4. Find the outcomes. Placements, guarantee-period refunds, temporary workers asked back, hires still in post. Where the system does not record them, finance and email often do.
5. Leave out what you cannot stand behind. Equal opportunities and health data, video interviews from Illinois residents, profiles scraped from sites whose terms barred it, and anything the candidate privacy notice for those years did not cover.
6. De-identify the people. Candidates, referees, clients' managers and recruiters: names, contact details, addresses, dates of birth, photos, and the employer and school names that make a career unique come out or are generalised; recruiters and candidates get stable codes so the decision trail still holds together. The de-identification standard sets out how records are cleaned before they leave.
7. Describe it in a datasheet: the systems, the years, the number of jobs and candidates, what each record holds, what was removed, what was left out and why, and the privacy notices in force. It is the document an AI Act buyer will ask for.

The guide to preparing data for sale covers the rest of the company's systems, the rights check asks the questions above in order, and the guide to proving quality without sharing the data shows how to answer a buyer's questions before anything leaves.

Two worked examples

These examples are illustrative.

A US IT staffing firm. A firm with about 70 staff has placed software engineers and IT contractors for eleven years, on Bullhorn, with its email, contracts and finance records beside it. Its records show every requirement a client sent, the contractors submitted, the client's feedback, the placements, the timesheets and which contractors were extended or asked back: a decision record with outcomes. It clears the size lines of the buyers that publish one, and Scale AI and Centific name its industry. Before offering it, the firm reads its own Bullhorn agreement (the main one is not published) and checks whether AI Assistant was switched on. Its California contractors are consumers under that state's law, so it removes them or de-identifies them; it takes out names, contacts, client names and rates that identify a contract, and codes its recruiters. Offered as one record with its email and finance, it is valued as a professional services company whose recruitment history is the part that shows its skill.

An executive search boutique in the Netherlands. A practice with six partners has run senior searches since the early 2000s. Its files are deep (long briefs, candidate assessments, references, the reasons a shortlist was cut) and very sensitive. Six people fall under the size line of every buyer that publishes one. Its candidates gave their data to be considered for roles, so licensing it for AI training is a new purpose under the compatibility test, and senior careers are easy to re-identify from a few facts. The honest answer is that its records are unlikely to be licensable as they stand. If it wanted to, it would have to tell its candidates first, strip far more than names, and accept that what remains may be too thin to interest a buyer. Its briefs and its methods (how it scopes a search, how it assesses) may be worth more than its candidate files, and they hold far less personal data.

Questions people ask

Do AI companies buy recruitment data? Some name the industry: Scale AI, Centific and Grepped list it, and none names the records a recruiter keeps. Most buyers do not name it and would judge a recruitment firm's records as part of its whole company record.

Can a recruitment agency sell its candidate database? Not as a database of people. No buyer we read names CV databases, and in the UK and the EU licensing candidates' data for AI training is a new purpose that has to pass the compatibility test. What can sell is the de-identified record of the decisions the firm made, with outcomes.

Can I license job descriptions paired with the candidates recruiters selected? That pairing is the part of a recruitment record that shows judgment. It sells once the candidates in it cannot be identified, with the client's confidential detail taken out, and with a datasheet that says where it came from.

Does my ATS own my candidate data? Not in the terms we read. Lever's customer retains all rights in its Customer Data, Workable's all rights in its Customer Content, and Greenhouse's licence is to provide its service. The vendors keep de-identified or aggregated data for their own use, and that part is theirs.

Is AI trained on hiring data high-risk under the EU AI Act? A system used to recruit or select people, filter applications or evaluate candidates is listed as high-risk, and its maker must document where its training data came from and why it was first collected. That binds the buyer, which is why it will ask the firm for those answers in writing.

What is a recruitment firm's data worth? No one publishes a price for recruitment records alone. Whole-company sales on record have a median of $72.5K per company, and the buyers' own ranges are wide. A valuation of the firm's whole record is the way to a figure.

Do recruitment AI companies buy agency data? None of the firms we read says it does. Eightfold builds on its customers' applicant records and hiring outcomes, beside public profiles, and the ATS vendors train on their own customers' data under their terms. The buyers to approach are the AI data buyers that name the field.

What this means for you

A recruitment firm's records sell when they show the work of recruiting, and its software rarely stops it licensing them. What decides a sale is whether the records show decisions and outcomes, and whether the people in them can be taken out:

1. Name what you hold. Decision records (briefs, screening notes, feedback, outcomes) or a database of profiles. The first has named buyers; the second has none.
2. Read your system's terms for who holds the records, the vendor's aggregate and any AI feature switched on, and export through the API while the subscription runs.
3. Find the outcomes, in the ATS, the finance records or the email, because they are what buyers name last and value most.
4. Take the people out: candidates, referees, clients' managers and your own recruiters, with sensitive data, video interviews and scraped profiles left out entirely.
5. Write down the origin of every system: years, privacy notices, what was removed. It is what a buyer training anything near hiring will ask for.
6. Value your data with your recruitment system among your systems, and check the buyer board for the buyers that name your field.

AIData.Broker introduces the holder to the buyers that take what it holds, free for the holder because the buyers pay us, or, where a sale we run earns more, acts as its agent for a fee fixed in the mandate and due only if the sale closes; either way the holder itself licenses its data to the buyer. We run sales for holders based in the United States, the United Kingdom and Canada. For the wider picture, see how to sell company data to AI companies and what a data licensing agreement contains.

This guide is research from public sources, not legal advice.