Field notes
AI replaced the admin surface of recruiting. The job got harder.
13 July 2026
On this page
Twelve months ago, the Perplexity CEO said AI would replace recruiters within six months. It is July 2026. Recruiters are still here. What is interesting is not that the prediction was wrong. It is what it got right.
The administrative surface of recruiting has been replaced. Sourcing lists, outreach drafts, follow-up sequences, scheduling, first-round screening questions. All of it is now faster, cheaper, and largely automated. If your competitive advantage was sitting in a boolean search for two hours or drafting a speculative email from a blank page, that advantage is gone.
What remains is the job. And the job, stripped of its administrative cover, turns out to be almost entirely judgment. That is harder, not easier.
Picture a mid-size contingency desk, eight consultants, mostly perm. In early 2025 they adopted an AI sourcing layer, a sequencing tool with GPT-generated personalisation, and an automated scheduling integration. By Q3 2025 the admin overhead per role had dropped by roughly 60%. Consultants were getting to first conversation faster. Pipeline volume went up. The MD was pleased.
By Q1 2026 something had shifted. Conversion from first conversation to placement had not improved. In some months it had got worse. The volume of activity was higher. The output was flat.
The MD pulled the data. Time-to-first-contact: down. Candidate response rate: up slightly. Offer-to-acceptance: unchanged. Placement-to-fee: down four points year-on-year.
The admin surface had been removed. What was left was the judgment layer, and it turned out the team had been spending most of their cognitive energy on the admin. Not because they were lazy. Because the admin had always been the comfortable part of the job.
The MD ran three placement reviews from that quarter. In each case the breakdown was not sourcing. The candidate list was good. The outreach had landed. The interviews had been booked.
The breakdown was earlier, and quieter.
The brief. In two of the three cases, the consultant had taken a role brief over a thirty-minute call, fed the notes into an AI to generate a job spec and a target profile, and started sourcing. The AI produced a clean, confident output. Nobody had pushed back on the brief itself. One of those roles was a replacement hire for someone who had been managed out. The hiring manager had not told the consultant that. The team dynamic was the actual constraint. No amount of sourcing accuracy was going to fix that.
The client read. In the third case, the role was marked urgent. It had been marked urgent for eleven weeks. The consultant had kept the pipeline moving because the system flagged it as a priority. What the system could not flag: the budget had not been signed off internally. The urgency was the hiring manager's urgency, not the business's. A more experienced consultant would have heard that in week two and either qualified it out or had a direct conversation about timeline. The AI-assisted process had made it easy to keep the activity going without asking the uncomfortable question.
The admin surface being gone had not created these problems. The problems had always been there. The admin surface had just been absorbing time that should have been spent on them.
The same dynamic plays out on business development. AI can score which prospects deserve your morning. It can draft an opener in your voice, reference a recent hire on their LinkedIn, and time the send for when their sector tends to be active. The mechanics are largely solved.
The question that still requires a human is: what is this prospect actually trying to do, and is this the right moment to be in front of them?
A system can identify that a fintech scale-up posted three engineering roles in the last fortnight and draft a message that references the hiring signal. It cannot know that the CTO you spoke to at an event six months ago is under pressure from the board to cut headcount by year-end, and that the hiring activity is a last push before a freeze. That context lives in a conversation you had, a tone you picked up on, a pattern you recognised across similar clients at similar stages.
The right design here is human approval before anything sends. Not because the AI output is bad. Because the decision about whether to send at all, and what the real ask is, belongs to the consultant. That is a design decision. Treating it as a limitation of the tool is how you end up with high-volume outreach that damages relationships you spent years building.
I have written about this intake problem from a different angle in the post on brief quality and AI agency workflow. The brief is where the judgment should be concentrated. The automation is everything downstream of a good brief.
| Stage | Before AI layer | After AI layer | What changed |
|---|---|---|---|
| Role brief | Consultant notes, manual spec write-up | AI-generated spec from call notes | Speed up. Quality of brief-taking unchanged. |
| Sourcing | Boolean search, 1-2 hours per role | AI-assisted shortlist, 20-30 minutes | Speed up. Accuracy similar. |
| Outreach drafting | Consultant writes per candidate | AI draft, consultant edits | Speed up. Voice consistency variable. |
| Scheduling | Back-and-forth email/phone | Automated booking links | Speed up. |
| First-round screen | Consultant call or structured form | AI screening questions, async video | Speed up. Judgment on fit: unchanged. |
| Brief interrogation | Depended on consultant seniority | Still depends on consultant seniority | No change. |
| Client read / politics | Depended on relationship depth | Still depends on relationship depth | No change. |
| Offer management | Consultant-led | Consultant-led | No change. |
The table is blunt but the point is clear. Everything that got faster was already a process. Everything that stayed the same was always judgment.
The prediction was not that AI would replace the judgment. It was that AI would replace the job as it was being done. That part landed. The job as it was being done in 2024 had a lot of admin in it. That admin is gone, or going.
What the prediction underestimated is that the admin had been providing cover. Consultants who were technically busy but not particularly sharp at client reading could fill their days with sourcing and outreach and still look productive. The AI layer has removed that cover. Judgment is now the whole job, and it is visible.
This is not comfortable for everyone. Some consultants who were good at the admin are finding the stripped-down version of the role harder. Some who were always strong on relationships and reading rooms are finding it liberating. The distribution of performance on a desk is getting wider, not narrower.
If you are running a recruitment business right now, the question is not whether to adopt the tooling. That decision is largely made. The question is whether your team has the judgment skills to operate at the level the tooling now exposes. Training, coaching, and structured brief-taking matter more than they did two years ago. The AI workflow audit I run with agencies is often less about the tools and more about whether the human decision points in the workflow are designed deliberately or inherited from a process that no longer exists.
The admin surface is gone. The job is still the job. It is just harder to hide from now.