Key takeaways
- The regulated activity is screening. The moment AI influences who is rejected, a body of employment law attaches.
- Almost all of the available time saving sits before and after screening, not inside it.
- As a staffing firm you may carry liability for the client's decision as well as your own.
- Job advert language is the one place AI can make bias worse if nobody edits the output.
Recruitment is a volume business built on written communication, which makes it one of the most naturally suited industries for this technology. It is also one of the most regulated in its specific application, because the thing recruiters do that AI appears best at, sorting large numbers of candidates quickly, is precisely the thing that legislators and enforcement agencies have been paying attention to.
The resolution is less restrictive than it sounds. Draw the line at screening, take everything on both sides of it, and you keep most of the benefit with none of the exposure.
What the law actually covers
The regulatory picture is fragmented but the direction is consistent. A few reference points that matter to a US staffing firm:
- New York City Local Law 144 requires an annual independent bias audit, publication of the results, and advance notice to candidates for automated employment decision tools used to substantially assist hiring or promotion decisions for roles in the city.
- Illinois regulates the use of AI to analyse video interviews, with notice, explanation and consent requirements.
- Colorado has enacted legislation covering developers and deployers of AI systems making consequential decisions, employment among them, with duties around risk management and disclosure.
- Federal enforcement agencies have been explicit that existing discrimination and disability law applies to algorithmic tools, and that using a vendor's product does not transfer responsibility for a discriminatory outcome.
Two things follow from this for a staffing firm. First, it is jurisdiction-specific and you place candidates in many jurisdictions, so the strictest rule effectively governs your process unless you are willing to run different processes by location. Second, an employment agency can carry its own liability under discrimination law, separately from the client. Being the intermediary does not put you outside it.
The practical position most firms should take: AI does not decide, rank, score or filter candidates. Full stop, written down, trained on. That one rule keeps you out of the automated employment decision tool definition in most frameworks, and the cost of it is smaller than it appears because screening was never where the hours were.
Where the hours actually are
Ask a recruiter to account for a week and the answer is rarely screening. It is outreach, formatting, chasing, note writing and client communication. All of it is high volume, low judgement, and outside the regulated zone.
| Task | Fit | Regulated |
|---|---|---|
| Turning a client brief into a job description and advert | Strong | No, but review for coded language |
| Building search strings and sourcing queries | Strong | No, it searches rather than decides |
| Personalising outreach at volume from a candidate's own profile | Strong | No |
| Summarising interview notes into a structured writeup | Strong | No, provided it does not recommend |
| Reformatting a CV into the client's submission template | Strong | No, formatting only |
| Drafting client updates, market maps and pipeline reports | Strong | No |
| Preparing candidates with role and company briefings | Good | No, verify company facts |
| Drafting rejection and feedback correspondence | Good | No, but a human decides and reviews |
| Scoring or ranking candidates against a role | Avoid | Yes |
| Filtering an applicant pool before human review | Avoid | Yes |
| Analysing recorded or video interviews | Avoid | Yes, explicitly in some states |
The first six rows are where a recruitment desk loses its week. A consultant who writes forty personalised outreach messages, five job adverts, twelve interview writeups and a dozen client updates is spending most of their productive time on composition rather than on judgement or relationships. That is the recoverable part, and it is entirely unregulated.
The job advert problem, in both directions
This deserves its own treatment because it is the one place where AI can actively make things worse without anyone noticing.
Historical job adverts contain a great deal of gendered and age-coded language, and models trained on them reproduce it. Ask for an energetic advert for a sales role and you will often get exactly the vocabulary that decades of research has shown narrows the applicant pool: aggressive, competitive, dominant, ninja, recent graduate energy, digital native. None of it is intended and all of it has an effect.
Used the other way, the same capability is a genuinely good control. Asking a model to review a draft advert for language that might discourage applicants on the basis of gender, age or disability, and to suggest neutral alternatives, catches things human writers miss. It works because reviewing is a comparison task rather than a generation task.
So the rule for adverts is that a person writes or heavily edits, and the tool reviews. Reversing that order is the version that causes harm.
Worth adding to your standard process: run every job advert through an inclusive language check before it publishes, as a required step rather than an optional one. It takes fifteen seconds, it improves applicant volume, and it demonstrates a good faith process if you are ever asked about it.
Candidate data is personal data
A recruitment firm holds an unusually large quantity of personal information about people who are not its clients and often have no relationship with it. State privacy laws increasingly give those people rights over that information, and candidates are more aware of it than they used to be.
The relevant configuration is the same as everywhere else and matters more here because of the volume:
- One firm account on a business tier with training on your inputs disabled, so candidate CVs are not passing through personal logins.
- Retention set deliberately and consistently with how long you keep candidate records elsewhere.
- Administrator-managed access, which matters in an industry with high consultant turnover and a strong tradition of taking your desk with you.
- A clear internal rule that candidate material goes into the firm account only.
The turnover point is worth emphasising. In most industries an offboarding gap is a compliance annoyance. In recruitment, a departing consultant with continued access to a tool containing candidate and client material is a commercial problem as well as a privacy one.
What clients are starting to ask
Larger employers are beginning to include AI questions in their supplier agreements with staffing firms, driven by their own compliance obligations. The questions are predictable: do you use AI in candidate selection, what tools, have they been audited, and will you warrant that you do not use automated decision tools in filling our roles.
A firm that has taken the position described here answers all four in a paragraph. A firm that has quietly let consultants use whatever they found is in a worse position than it realises, because the honest answer is that it does not know.
Frequently asked questions
Can we use AI to screen candidates?
It is the regulated activity and most firms should stay out of it. New York City requires bias audits and candidate notice for automated employment decision tools, Illinois regulates video interview analysis, Colorado covers consequential employment decisions, and federal agencies apply existing discrimination law regardless of the tool. The screening step was never where the hours were.
What can we safely use it for?
Job descriptions and adverts, search strings, outreach personalisation, interview note writeups, submission formatting, client updates and candidate briefings. None of these decide who progresses, and together they cover most of a consultant's composition time.
Does AI reduce bias in job adverts?
Only when used as a reviewer. Generating adverts from scratch tends to reproduce the gendered and age-coded language common in historical postings. Having a person write and the tool review for exclusionary phrasing is genuinely effective and worth making a required step.
Are we liable if a client's AI rejects someone unfairly?
Employment agencies can carry their own exposure under discrimination law, and being the intermediary does not automatically put you outside it. Larger clients are also starting to ask staffing suppliers to warrant that they do not use automated decision tools. Both reasons point the same way: keep the decision human and be able to say so.
Give your consultants their week back
We choose the provider, configure retention and access properly for candidate data, cap the spend, and write task guidance covering outreach, adverts, interview writeups and client reporting, with the screening boundary made explicit. Fixed price, live in 30 days or less.