An AI Coworker Just Went Live at 27,000 Desks. Look at What It Was Given to Do.

Agentic AI in sales just shipped at real scale, and the job list is the story

On September 15, the Adecco Group announced it is rolling Salesforce's Agentforce Coworker out across 40-plus countries, into the daily workflows of 27,000 employees, after a pilot in the UK and France. This is one of the largest named, in-production deployments of agentic AI in sales and recruitment anyone has put a press release behind, and it is worth reading closely — not for the headcount number, but for the task list.

Here is what the agent was given to do, in Adecco's own words: sales professionals and recruiters can "find priority prospects, prepare sales briefs, enrich prospect data and view the status of leads across teams."

Find. Prepare. Enrich. View status.

Not send. Not sequence. Not book the meeting autonomously. The largest agentic GTM rollout of the year handed the machine the research half of the job and left the conversation half with the human. That is the single most useful fact published in go-to-market this week, and it lines up exactly with what the outbound data has been saying for two years: the bottleneck in B2B selling is not message throughput. It is knowing who to talk to and what to say.

What Adecco actually shipped

The details matter, because most agentic AI announcements are demos with a logo on them. This one is not.

• Scope: Agentforce Coworker across 40-plus countries, into the workflows of 27,000 employees, following a pilot in the UK and France.
• Prior footprint: Adecco says it had already deployed agentic AI across recruitment workflows in ten countries, representing 50% of group revenues. This is an expansion, not a first attempt.
• Context layer: Coworker draws on more than 2.5 million agent-candidate interactions logged since April 2025. The agent is not reasoning from a blank slate — it sits on a proprietary interaction history.
• Consolidation, not addition: the stated benefit is "a single point of access to data, systems and organizational knowledge that was previously sat across dozens of tools."
• The model underneath: Salesforce's enterprise AI teammate, powered by Anthropic's Claude.

CEO Denis Machuel's framing in the release is worth quoting because it is unusually restrained for this genre: the point is to free teams "to focus on what only humans can do: richer engagement with candidates and clients." Reporting from the Dreamforce keynote, AI News noted Machuel citing 35% to 40% of recruiter time saved through agentic AI, and 20,000 more people placed so far this year, which he attributed to better conversations.

Hold onto that last clause. It is the whole argument.

Why this one matters more than the hundreds that never shipped

Most agent pilots die before production. The figures vary by source and none of them are flattering: AI News, summarising Deloitte's 2026 technology trends research, puts the pilot-to-production failure rate for AI agents at 89%, with a Teradata survey finding 78% of enterprises running at least one agent pilot but only 14% scaling one organisation-wide. Treat the exact numbers as directional — they come from vendor-adjacent research — but the shape is consistent with what every RevOps leader has watched happen in their own building.

The GTM-specific version is equally blunt. Salesloft's 2026 US Revenue Benchmark Report, based on 500 US sales and revenue decision-makers at companies with 200+ employees, found AI in use somewhere in the revenue process at effectively every organisation surveyed — and only 20.6% describing their AI strategy as production-ready with measurable outcomes. Roughly one in five. Everyone has the tool. Almost nobody has the operating model.

So when a 27,000-person deployment does clear the gap, the interesting question is not "did AI work." It is what, specifically, did they let it do?

The pattern across the deployments that stick

Look at the successful rollouts of the last eighteen months and the same shape keeps appearing:

• The agent operates on bounded, verifiable tasks — summarise this account, enrich this record, rank this list — where a wrong answer is visible immediately and cheap to correct.
• The agent sits on proprietary context the model could not have guessed. Adecco's 2.5 million interaction history is the moat, not the model.
• The agent consolidates tools rather than adding a new surface. Every "one more tab" AI product fights the rep's habits and loses.
• The irreversible step stays human. Nothing that touches a prospect's inbox, a candidate's reputation or a brand's standing fires without a person.

And the mirror image: deployments that fail almost always handed the agent the send button first, because that is the demo that looks most impressive on stage and costs the least to build.

Research is the expensive part. Sending never was.

This is where the Adecco task list stops being a piece of corporate news and starts being a directive for anyone running outbound.

Sending has been free for a decade. Any SDR with a sequencer can push a thousand touches a week. What sending is not, in 2026, is safe — LinkedIn's connection limits and behavioural scoring, plus enforced sender requirements at Google and Microsoft, mean incremental volume now costs you domain reputation and account standing rather than buying you incremental replies. Volume stopped being the lever a while ago. Most teams just kept pulling it.

Research is the part that was always expensive, because it never scaled. A genuinely well-targeted message needs someone to know that this account just posted a role that implies your problem, that this VP complained about your competitor's onboarding three weeks ago, that the champion who bought you at their last company started here in July. That work took a human twenty minutes per prospect. At twenty minutes a head, no team can do it at pipeline-relevant volume. So teams skipped it and sent generic messages instead, and reply rates went where reply rates went.

An agent that does research, enrichment and prioritisation attacks the constraint that actually binds. An agent that does sending attacks a constraint that was removed in 2014.

Time saved is not pipeline

Here is the conversion every AI-in-sales business case waves through, and the one that decides whether your rollout looks like Adecco's or like the 80% that stall.

Suppose the agent really does give a rep 35% of their week back. That is roughly 14 hours. Those hours become revenue only if they are spent on conversations that were not going to happen otherwise, with people who were going to buy something. Spend them on a bigger cold list and you have built a faster machine for producing ignored messages. The efficiency is real; the output is not.

Salesloft's benchmark data makes the size of the problem concrete: teams average 35.2 touches to create a single qualified opportunity, average quota attainment sits around 62%, and 68.4% of leaders report higher pipeline quotas this year. The top 10% of sellers produce 47.4% of closed-won revenue. That is not a throughput problem. A team needing 35 touches per qualified opportunity does not get healthier by making touch number 36 cheaper. It gets healthier by making touches one through five land on people with a reason to reply this month.

Note what Machuel attributed the 20,000 extra placements to. Not more outreach. Better conversations. The time saved was the input; the conversation quality was the mechanism; the placements were the output. Skip the middle term and the business case collapses.

The buyer side has already voted

There is a second reason to be careful about pointing reclaimed hours at volume: buyers have been getting steadily more allergic to it.

Gartner's 2026 sales survey found 67% of B2B buyers now prefer a rep-free buying experience, up from 61% a year earlier, with 45% saying they used AI during a recent purchase. Meanwhile G2's 2026 research found roughly half of B2B software buyers now start vendor research in an AI chatbot more often than in Google. The front door moved, and it moved away from your sequence.

But the same Gartner work contains the opening: 69% of buyers return to a sales rep to validate what the AI told them. Buyers do not want to be pitched. They do want a person who can confirm, correct and contextualise what they have already researched alone. That is a high-value conversation, and it is only available to reps who arrive with something the buyer does not already have. Which is, again, a research problem.

What to do with this in the next 30 days

You do not need a 40-country enterprise agreement to copy the structure. The structure is the transferable part.

1. Audit your AI spend against the task list. Write out every AI tool in your GTM stack and mark each one as research-side or send-side. If most of your budget sits on the send side, you have automated the cheap half of the job. That allocation is the single most common reason AI shows up in the tool list and never in the pipeline number.
2. Pick one signal and instrument it properly. Not twelve. One. Profile views, competitor complaints, a hiring pattern that implies your problem, engagement on a specific kind of post. Get it flowing daily, with the research attached, before you add a second. Signal-based outbound is what Updately is built around for exactly this reason: the capture, enrichment, scoring and per-prospect research have to be one motion, or the signal arrives without the context that makes it actionable and the rep ignores it.
3. Set a per-prospect research floor and enforce it. Decide what a rep must know before a first touch — the trigger, one specific detail about the account, one reason now rather than last quarter. If the agent cannot supply all three, the prospect does not get worked. This one rule does more for reply rates than any template library.
4. Keep the send button human, and say so out loud. Graduated autonomy with a human gate on anything irreversible is not timidity; it is what the successful deployments have in common. Draft with AI, approve with a person.
5. Rewrite the success metric before the rollout, not after. If you launch on "hours saved," you will hit it and learn nothing. See the table below.

Measure the conversion, not the input

That third row is the experiment worth running this quarter, and it is cheap: hold your list and your messaging constant, run one cohort with full agent research attached and one without, and read the reply-rate difference after three weeks. Whatever the gap is, that is the real, measured value of agentic AI in sales inside your motion — not a vendor's slide.

Takeaways

• The largest agentic GTM rollout of the year gave the agent research, enrichment, briefing and prioritisation — and kept the conversation with humans. Copy the division of labour, not the vendor.
• Sending was never the constraint. It has been free for a decade and is now actively risky at volume. Research was the expensive part, and it is the part agents genuinely fix.
• Time saved is an input, not an outcome. The chain is reclaimed hours to better-targeted conversations to pipeline. Adecco's own framing credits better conversations, not more of them.
• Buyers want fewer, better-informed reps. Two thirds prefer a rep-free experience, and most still come back to a human to validate what AI told them. Show up with something they could not get from a chatbot or do not show up.
• Most AI GTM programmes stall at measurement. Only about one in five revenue teams call their AI production-ready with measurable outcomes. Define the metric that proves the conversion before you scale the tool.

The agent is now a real coworker at 27,000 desks, doing the research half of a sales job in 40 countries. The question for your team is not whether to hire one. It is whether you will give yours the job that was actually hard, or the job that was already easy.

Sources: Adecco Group press release, September 15, 2026 · AI News on the Adecco rollout · AI News on enterprise agent pilot failure rates · Salesloft 2026 US Revenue Benchmark Report · Gartner 2026 sales survey · G2 2026 AI search research