This piece was first written in April 2024, when the industry had just found the term Large Action Model and expected it to rewrite hiring inside a year. Two years on, it is worth being honest about what happened: the term did not survive, but most of the capability did. What follows is an updated account of what AI recruitment assistants actually do now, what they still cannot do, what has to sit underneath one before it works at all, and the compliance regime that arrived while everyone was arguing about definitions.

What actually happened to the Large Action Model

The phrase was popularised by Rabbit's R1 device at CES in January 2024. Rabbit defined a LAM as a model that could "use any software it runs into, takes actions, and gets better over time" by watching how people operate interfaces and then driving those interfaces the same way a human would. Applied to recruiting, the promise was obvious: an assistant that could just use your ATS, your LinkedIn, and your inbox, the way a junior recruiter would.

The shipped reality was narrower. Rabbit's own newsroom page listed the LAM-powered categories available on day one: music, generative AI, rideshare and food. Travel was demonstrated and then held back. Rabbit was also candid in the same post about the underlying limitation: "LLMs, the underlying frameworks for understanding intentions, are still at their early stages. They hallucinate, make mistakes, and are not fast enough yet." Reviewers were blunter, and The Verge called the device unfinished at launch.

The term quietly disappeared. What replaced it was less magical and more useful: agentic AI and computer use, where a model is given a defined set of tools and API calls rather than being expected to learn any interface on sight. That distinction is the whole story for recruiting teams. The assistants that work today are not one general model that teaches itself your stack. They are narrow agents wired to specific integrations, with a human gate on the decisions that matter.

So the 2024 framing was right about the direction and wrong about the mechanism and the timeline. Keep that in mind for every claim below, including the ones this article made two years ago.

How AI recruitment assistants actually operate

An AI recruitment assistant earns its place by doing complex, repetitive top-of-funnel work autonomously and a lot faster than the manual alternative. Here is the honest version of each stage: the autonomous capability, the decision support, and the handoff to human recruiters.

1. Autonomous candidate sourcing and engagement

An assistant can source candidates autonomously by searching profiles, job boards and public platforms, then ranking who matches a role brief. Once candidates are identified, it can draft and send personalised outreach based on their background and public work.

• Example: for a backend engineering role, the assistant identifies people with the right language stack and a history of meaningful open-source contributions, then opens with a specific reference to that work before introducing the role. This beats a templated InMail because it is evidently about the candidate rather than about the vacancy.

The limit worth naming: an assistant can only search the sources it is actually connected to. "Scanning the whole internet" is not a thing. If your agent has a LinkedIn integration and a GitHub integration, it sees LinkedIn and GitHub. Ask any vendor which sources are live, and treat "we search everywhere" as a red flag rather than a feature.

2. Screening and decision support

The second job is deciding who moves forward, based on qualifications, trajectory and fit against the brief. This is where the 2024 version of this article was wrong in a way that is worth correcting rather than deleting.

The original claim was that the assistant learns from each interaction, spots that (say) candidates with cross-functional experience succeed at your company, and starts prioritising that trait automatically. Two things are now clear. First, most production recruiting tools do not continuously retrain on your hiring outcomes: the "learning" is usually you editing a prompt or a scorecard, and vendors that imply otherwise are usually describing a roadmap. Second, and more importantly, that exact mechanism is the one regulators now treat as high risk. An agent that mines your past hires for success patterns and optimises toward them will faithfully reproduce whatever bias your past hiring contained, and it will do it at machine scale while looking objective. A proxy like "cross-functional experience" is harmless. A proxy that correlates with age, gender or ethnicity is a disparate-impact claim waiting to happen, and the agent cannot tell the difference.

The practical guidance: use the assistant to rank and summarise against criteria a human wrote, not to infer its own criteria from your history.

3. The handoff to human recruiters

This is the stage the original article promised and never delivered, and it turns out to be the one that decides whether a deployment works.

A good assistant does not replace the recruiter's judgement, it compresses the evidence that judgement needs. The output should be a short, sourced summary per candidate: what they have built, why they match the brief, what the assistant is unsure about, and what it already said to them. The candidate-facing side matters too. AI assistants can hold a consistent identity and a coherent conversation thread, which is what keeps the experience from feeling like a broadcast.

Set the gate explicitly. Decide in advance which actions the agent may take alone (search, rank, draft, follow up, schedule) and which require a human (rejecting someone, making an offer, anything a candidate would want to appeal). Under the EU AI Act that human oversight is not a nice-to-have, it is Article 14. More on that below.

What an AI recruitment assistant needs underneath it

The part the 2024 article hand-waved: "the AI can initiate contact." How, exactly? An autonomous assistant is only as good as four layers beneath it, and teams usually discover them in the wrong order.

A source of profiles. Whatever the agent is connected to. See the limit above.

A contact channel. This is where most deployments actually stall, and it is almost never in the demo. A GitHub profile is not an email address. A LinkedIn profile is not an email address either. The agent identifies a perfect candidate and then has nowhere to send the message it just wrote. You need a contact-data layer, and Apollo.io is the cheapest published route to one for most recruiting teams.

A system of record. The agent has to write what it did somewhere durable, or you have an unauditable black box. This matters more than it used to, for reasons in the next section.

A human gate. Covered above. Skip it and the other three layers become a liability rather than an asset.

The regulation caught up

In April 2024 this was a footnote. It is now the main constraint on how autonomous your assistant is allowed to be.

The EU AI Act classes AI used for recruitment and candidate selection as high risk (Annex III). That brings real obligations: risk management, data governance and bias controls, meaningful human oversight, logging and auditability, technical documentation, transparency toward candidates, and demonstrated accuracy and robustness. The deadline moved. High-risk obligations were due to apply from 2 August 2026, but the Digital Omnibus package pushed Annex III systems to 2 December 2027 as a fixed date, endorsed by the European Parliament on 16 June 2026 and given final approval by the Council on 29 June 2026. That is a delay, not a repeal, and the tooling you buy now is the tooling you will have to document then.

NYC Local Law 144 has been live and enforceable since 2023, and it applies today. If an automated tool substantially assists an employment decision for a candidate in New York City, you owe an independent bias audit, a public summary of its results on your careers site, and advance notice to candidates. Penalties are $500 for a first violation and between $500 and $1,500 for each subsequent one, and each day of use in violation counts as a separate violation.

Enforcement has been weak, and that is changing. A New York State Comptroller audit covering July 2023 to June 2025 found that 75% of test calls to 311 about automated hiring tools were misrouted and never reached the enforcing agency, and that when the DCWP reviewed 32 published bias audits it flagged a single compliance issue where the Comptroller's own reviewers found at least 17. The agency has committed to tightening its approach. Reading that as "nobody is checking" is a short-term bet.

Where this leaves you

The 2024 thesis holds up better than its vocabulary. AI assistants, HeroHunt.ai included, really do identify, engage, and evaluate candidates at a speed no manual team matches, and the personalised, evidence-led outreach described two years ago is now ordinary. What was wrong was the mechanism (no general model learns your stack by watching), the autonomy (the interesting decisions are gated, increasingly by law), and the assumption that the hard part was intelligence. The hard part is plumbing: connected sources, deliverable contact data, a system of record, and a human gate you designed on purpose.

If you are standing one up, work backwards from the plumbing rather than forwards from the demo. Start by checking whether you can even reach the candidates your agent finds, because that is the failure that shows up in week two.