Recruitment Agencies

Recruitment automation that keeps your CRM and your consultants. Helix agents drive LinkedIn Recruiter and Bullhorn in a real browser, source and draft overnight, and hand the work to a human every morning.

• No API? Use the screenAgents work LinkedIn Recruiter in a real browserEach agent gets its own Linux desktop. It runs the search, reads the profiles and builds the project in the same interface your consultants use.
• The night shiftSourcing continues after your team goes homeThe consultant hands the Recruiter seat over at the end of the day. By morning the project is populated and outreach drafts are waiting for review.
• Your stack staysBullhorn remains the system of recordResults are written back through your CRM's API. No migration, no new tool for consultants to learn, and a person approves every message.

Recruitment automation stops at the tools that have no API

Every agency has automated the parts that were easy to automate. Job posts go out to multiple boards. The CRM fires sequences. Resume parsing pulls structured fields out of a CV. Then it stops, because the two systems your consultants actually live in all day are the two that resist it.

LinkedIn Recruiter has no public API. There is nothing to integrate with. Every search, every profile opened, every InMail drafted, every candidate added to a project is a human moving a mouse. A delivery consultant building out a search spends hours scrolling, and none of that time produces anything a workflow tool can pick up.

Your CRM has an API, so the integration exists, but it is a one-way street. It records what already happened. Bullhorn will store the candidate your consultant found on LinkedIn. It will not go and find the next one.

This is why recruitment automation software tends to disappoint agencies. The category has settled on the workflow layer, which was never the bottleneck. The bottleneck is the browser.

Agents that use the same screen your consultants do

Helix gives each AI agent its own Linux desktop with a real browser, running on your infrastructure. The agent does not call an API. It logs in, runs the search, reads the profiles, and clicks the buttons, in the same interface your team uses.

An agent with a desktop can work in any tool a human can work in, including the ones that were never designed to be automated. The architecture behind this is described in Virtual Desktops for AI Agents.

The desktop is shared, not hidden. You connect to it through your browser and watch the agent work, or take the mouse and do a step yourself. A consultant can sit with an agent for an afternoon, doing a search the way they would normally do it, and the agent learns the shape of that search from watching and from being told what matters.

The night shift

Recruitment automation has an unusual constraint that most software categories do not: LinkedIn Recruiter reacts badly when one seat is used from two places at once. The agent and the human cannot both be working the account. So they take turns.

At the end of the working day the consultant hands their session to the agent desktop. Overnight the agent sources against the live searches, reviews profiles, and writes outreach drafts. In the morning the candidates are sitting in a LinkedIn Recruiter project and the drafts are waiting for a human to read, change, and send.

The shift pattern is the same one distributed engineering teams use to pass work across time zones, described in Follow the Sun. In an agency the handover is between a person and an agent, not between two offices. The requirement is identical: context has to survive the handover, or the next shift spends its first hour rebuilding it.

Results are written back to your CRM through its API, so Bullhorn stays the system of record. Nothing migrates. The full technical write-up is in Automating LinkedIn Recruiter when there is no API.

What agencies point agents at

• DeliverySourcing and long listsThe work that occupies delivery consultants full time. An agent runs the search, screens against the brief, and builds the project. A human decides who is worth a conversation.
• DeliveryOutreach drafts in your consultant's voiceWrite the first few messages yourself. The agent continues in that register, adjusting each one to the person it is addressed to. You approve every message before it is sent.
• SalesBusiness development researchFind companies hiring for the roles you place, work out who owns the requirement, and pull the relevant case study from your own site. For PE-backed clients, the same research runs across a fund's portfolio.
• ContractContract and interim desksAvailability checks, compliance chasing, timesheet nudges and rate benchmarking are routine enough to delegate and consequential enough that a human still signs off.
• InternalYour own hiringInternal roles lose every prioritisation contest against billable ones. An agent working the internal desk does not have that conflict.
• EverythingA person approves the outputCandidates and clients can recognise a language model. An agency that sends generated outreach at volume trains its market to ignore it, so nothing goes out unread.

The result is a busier team, not a smaller one

The first feedback from an agency running this pattern was that it was keeping them busier. The pitch for AI in recruitment is usually the opposite.

An agent that sources overnight produces more qualified conversations to have the next day. The constraint moves from finding people to talking to them, which is the part your consultants are paid for and the part that cannot be delegated.

There is a real cost to the approach. A human reviews and approves the output, so the throughput ceiling is your team's review capacity, not the agent's capacity. Use the model below to see where that ceiling falls for one of your desks.

If your model is volume outreach with minimal human contact, this will not help you, and there are cheaper tools that will. Agencies that want a fully autonomous pipeline with nobody in the loop should not buy this.

How this compares with other recruitment automation tools

Running it on your own infrastructure

Candidate data is personal data. Under UK GDPR you are a controller for it, and sending CVs, contact details and interview notes to a third-party model provider is a processing decision you have to be able to defend.

• Where it runsYour Kubernetes, your office, or a MacOn your own cluster, on a Sovereign Server in your office, or on a Mac. Nothing about the workload requires a vendor cloud.
• Where data goesModels run locallyCVs, contact details and notes never leave your perimeter. Clients in financial services and the public sector audit this as part of supplier onboarding.
• Who did whatEvery action is logged and attributableEach agent action is recorded against the desktop it ran on. Automated decisions about people carry obligations under UK GDPR Article 22 and the EU AI Act, and a human stays on each one.

Helix runs on your infrastructure, including a Sovereign Server in your office. Agencies placing into financial services or the public sector have clients who audit their data handling as part of the supplier process. There, where the data lives is often the difference between a project that gets approved and one that does not.

The EU AI Act classifies employment and worker-selection systems as high risk. Keeping a human on every consequential decision is the design here, and it is also what those regimes expect. Take your own legal advice on your jurisdiction and your process.

Getting started

Most agencies are five to fifty people with a CRM they are not going to replace and no appetite for another tool that needs learning. So the first engagement is scoped as one workflow, usually overnight sourcing on one desk, run alongside what your team already does.

1. 1
   Pick one desk and one open briefUsually the delivery desk where sourcing is the constraint. One consultant, one Recruiter seat, one live search.
2. 2
   Sit with the agent for an afternoonThe consultant runs the search on the shared desktop, talking through what a good profile looks like. The agent learns the shape of the search from watching.
3. 3
   Hand the seat over at 18:00The agent sources, screens and drafts overnight. Records go to Bullhorn through its API. A Slack message lands when a batch is ready.
4. 4
   Review the project in the morningThe consultant reads the shortlist, edits the drafts and sends the ones worth sending. If there is no usable project by the end of week one, we stop.

We work with your consultants to set it up rather than handing over a login, and the use cases expand from there once the first one is holding.

Frequently asked questions