AI coworker products have arrived faster than the rules for working beside them. In a feature published on 28 September 2026, WIRED’s Kate Taylor wrote that “hundreds of thousands of brand-new coworkers” will join the workforce in the coming months. They will have names and profile pictures, sit in Slack channels and email chains, and take on jobs such as chief of staff, engineer and marketer. None of them will be human. They are AI agents: software built on a large language model that completes tasks on its own, now sold as digital employees rather than as tools.
The WIRED feature asks a question the industry has mostly skipped: what is an AI coworker actually like as a colleague? The answer matters more than it sounds. A Boston Consulting Group study that WIRED cites found that managers check identical work less carefully when it is attributed to an “AI employee” rather than to an AI tool. They catch fewer errors and push more of the blame onto the software.
This article sets out what WIRED reported, which AI coworker products are already on sale, and what the BCG research actually measured, including two figures that have changed since the survey was first publicised. It then looks at why vendors give their agents names and faces, and at what organisations should put in place before an agent appears on their own org chart.
Table of contents
- What WIRED Reported About the AI Coworker Wave
- The AI Coworker Products Already on the Market
- How Many Firms Already Put an AI Coworker on the Org Chart?
- The Hot Potato Effect: Oversight Weakens When AI Is Called an Employee
- Why Vendors Give Every AI Coworker a Name and a Face
- The Lattice Precedent and Other Early Warnings
- Agents per Person: The Arithmetic of an AI Coworker Workforce
- Governing an AI Coworker: Tool Framing Versus Teammate Framing
- What Employees Should Expect From an AI Coworker
- AI Coworker FAQs
- References and Further Reading
What WIRED Reported About the AI Coworker Wave
Taylor’s feature is less about what agents can do than about where they sit socially. Its starting point is that agents are “supposed to be the perfect worker bees: available 24/7, never complaining, and providing superhuman intelligence”. Yet there has been “curiously little discussion” of how they behave as colleagues.
A question nobody had language for
Julie Bedard, a partner at Boston Consulting Group, noticed the confusion three years ago, when ChatGPT’s rapid adoption convinced executives that AI could transform productivity. “People didn’t have language for it,” she told WIRED. “Is it a tool? Is it a teammate? Is it a colleague? Is it a coworker?” That vocabulary problem has now become an operational one, because the answer an organisation picks changes how its managers behave.
From tools to named colleagues
When Microsoft’s Copilot and Google’s Gemini for Workspace launched in 2023, both were marketed as tools for employees. The AI coworker pitch is different. Vendors now describe agents with job titles, reporting lines and personalities, and some customers list them on the org chart beside people. WIRED’s reporting runs from Microsoft’s enterprise assistant to a start-up whose agents appear as Muppet-like avatars, and on to an HR chief executive who thinks the whole framing is a mistake.
Why the timing matters now
The shift is happening quickly. Microsoft’s 2026 Work Trend Index, published on 5 May, reported a 15-fold year-on-year increase in active agents inside Microsoft 365, rising to 18-fold in large enterprises. Gartner, meanwhile, has predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027. Rapid deployment combined with a high failure rate is exactly the situation in which governance gets decided by default rather than by design.
The AI Coworker Products Already on the Market
The products in WIRED’s piece show how wide the AI coworker category already is. At one end sits an enterprise assistant bound to Microsoft’s identity system. At the other is a consumer agent with a fuzzy avatar.
Microsoft Scout, an Autopilot with its own identity
Microsoft introduced Scout at its Build conference on 2 June 2026 as the first of a new category it calls Autopilots. Its announcement describes them as “always-on agents that work autonomously, with their own identity, and act on your behalf”. Scout works across Teams, Outlook, OneDrive and SharePoint. It schedules meetings across time zones, blocks calendar time for deliverables and flags stalled decisions.
“Your company essentially hires your assistant,” Omar Shahine, the Microsoft corporate vice president behind Scout, said at launch. “The whole point of having a personal assistant is that they’re working when you’re not working.” Scout is built on the open-source OpenClaw project and is offered only as an experimental release to Frontier customers. Microsoft’s own post stresses control: each agent runs under its own governed Entra identity, and sensitive actions can require a person to sign off.
Skydive, where every AI coworker gets a name and a computer
Anything, a vibe-coding start-up founded by former Google product leaders Dhruv Amin and Marcus Lowe, launched Skydive on 20 August 2026. Its launch release says Skydive agents “aren’t chatbots or workflows. They’re coworkers.” Each has a name, a role and its own cloud computer with a browser, file system and terminal. Agents work across Slack, email, iMessage, the desktop and the web.
Anything runs itself this way. The release says its 15-person team works alongside more than 150 agents, including Emma in customer support, Grace investigating bugs, and Canary, who monitors production around the clock and once alerted a co-founder to an outage at 2am. Skydive supports more than 170 models and had been in closed beta with more than 20 companies. “Inside Anything, they haven’t replaced a single hire,” Amin said in the release.
Meta’s Muse and the consumer AI coworker
The same design language is spreading beyond the office. Meta’s Muse, launched on 8 September, is also represented by a cute and fuzzy avatar. WIRED’s review compared using it to “texting a friend with task requests”, with the agent acknowledging messages with a thumbs-up emoji and working in the background. We have followed its rapid development, including a dedicated VM tab and new voice widget and phone calling shared with rival Instinct.
Enterprise buyers are already committing
Some companies have gone well beyond pilots. HR platform Deel says it has built more than 8,000 agents that do the work of about 600 employees. Enterprise vendor Ema has raised $77 million on the promise of agents for HR, IT and finance. Neither describes its agents as a replacement for a whole workforce, but both sell the idea that an agent can hold a defined role.
| Offering | Launched | How it is framed | Controls stated by the vendor |
|---|---|---|---|
| Microsoft Scout | 2 June 2026, Frontier preview | An always-on “Autopilot” | Own Entra identity; human sign-off for sensitive actions |
| Skydive by Anything | 20 August 2026 | Named coworkers with roles | Central control of access, data, usage and spend |
| Meta Muse | 8 September 2026 | A friendly avatar “built for everyone” | Consumer app settings |
| Deel Akai | Open launch September 2026 | Agents running named workflows | Human approval where money moves |
| Lattice digital workers | 9 July 2024, dropped 12 July | Employee records and managers | Withdrawn after a backlash |
How Many Firms Already Put an AI Coworker on the Org Chart?
WIRED’s headline statistic comes from Bedard and her colleagues. When they polled 1,261 managers in January, 22% said their organisations had added AI agents to corporate org charts. The research behind it is a working paper, Putting AI on the Org Chart, by Emma Wiles of Boston University with Megan Hsu, Julie Bedard and Matthew Kropp of BCG. It is worth reading in full, because its numbers have moved.
The 1,261-manager survey
The first sample was HR and finance managers, directors and executives recruited through a professional expert network, who registered from 5 January 2026. In the paper’s latest draft, dated 19 September 2026, 31% say their leadership frames AI as a “teammate or employee”. A further measure found that 23% say their organisation lists AI agents on its org or workflow charts. Those listings were most common in large firms and in technology and financial services.
A broader sample of 1,500 executives
An expert-network panel may over-represent early adopters, so the authors fielded a second survey in June 2026. It covered 1,500 US senior managers, directors and executives, weighted to the industry mix of the private sector. That survey gives a more conservative picture of the AI coworker trend. Only 14% list agents on org or work charts, although 53% say their organisation has officially integrated agents into systems or workflows.
Why 22% became 23%
Bedard’s figure of 22% matches the version of the study BCG publicised in May, which also reported that managers caught 18% fewer errors. The September draft gives 23% and 17%. Neither change alters the conclusion. Anyone quoting the research should still cite the version they used, and should note that the higher share comes from a panel of HR and finance managers rather than from a representative survey.
The broader June survey shows how far formal recognition of an AI coworker has already spread among US executives.
The Hot Potato Effect: Oversight Weakens When AI Is Called an Employee
The more important part of the paper is an experiment. It tests whether calling an agent a colleague changes how carefully people check its work. The answer is yes, at least in organisations that have already made the move.
How the experiment worked
On 9 January 2026 the registered managers were invited to review five documents, each seeded with errors: job descriptions for the HR managers and budget documents for the finance managers. They had 20 minutes. The documents were identical for everyone, and only the stated source changed. One group was told the drafts came from an AI tool they had used, a second from an AI employee and a third from a human employee. Of 1,261 registrants, 813 completed the task and passed an attention check.
What changed at firms with an AI coworker on the chart
Across the whole sample, the average effects were small. The effect appeared among managers whose organisations already list agents on their charts. For that group, presenting the drafts as an AI employee’s work rather than an AI tool’s cut the share of errors they detected by 17% and their overall review score by 13%. They were 22 percentage points more likely to ask someone else to review the work. They also gave themselves about 9 points less of the accountability, and the AI system about 8 points more.
Why the authors call it a hot potato
The authors describe this as a “hot potato” effect. Managers take less personal responsibility for an AI coworker’s output, catch fewer errors themselves and pass the work along for more review. Extra review is not necessarily bad. But, as the paper notes, it “creates a record that the manager did not sign off alone”, which may also shrink the manager’s own share of any blame.
An AI employee is not treated like a human one
The benchmark group matters. Managers reviewing a human employee’s work flagged more errors themselves, and were less likely to request more review, than those reviewing an AI employee’s work. So the weaker oversight is not simply what happens whenever work is delegated. The paper concludes that AI employees occupy “a hybrid organizational position: treated as delegated producers rather than tools, but not monitored the same way managers monitor human subordinates”.
| Measure (managers at org-chart firms) | AI employee framing vs AI tool framing | What it means in practice |
|---|---|---|
| Share of seeded errors detected | 17% lower | More mistakes survive review |
| Overall review score (F1) | 13% lower | Weaker checking overall |
| Requests for another reviewer | 22 points higher (44% more likely) | Checking is passed along |
| Accountability given to self | About 9 points lower | Ownership drifts away |
| Accountability given to the AI system | About 8 points higher | Blame lands on software |
Fortune’s report on the May version framed the finding as AI “employees” making human workers sloppier. The paper itself is more careful. It says its results “should not be interpreted as evidence that organizations should not use AI agents”, only that productivity gains “do not remove the need for governance”.
Why Vendors Give Every AI Coworker a Name and a Face
If naming agents weakens oversight, why do vendors do it? WIRED’s interviews suggest two reasons. It makes the product easier to understand, and it makes users attached to it.
Muppets, Alice and Bob
Skydive’s agents each have a Muppet-esque avatar that Amin describes as “close to human, but definitely not human”, with names tied to their jobs. “People start to really respect the agent and form strong attachments,” he told WIRED. “They’ll talk about it like humans.” He argues that personification is practical, because “most people still don’t know what a real agent is”.
Christine Wendell, chief executive of Pronto Housing and an early Skydive customer, uses agents to write contracts, schedule meetings and send follow-up emails. After her engineering team adopted a coding agent, staff began saying they had “worked with Alice” on a project. Skydive agents also adapt to each colleague’s communication style, requests and corrections, so every AI coworker gets steadily better at fitting in.
Colleagues without office politics
Part of the appeal of an AI coworker is precisely that it is not human. It has a limitless work ethic and sits outside office politics. When “Bob”, the agent serving as Wendell’s chief of staff, posted the details of her calendar in a Slack channel, she corrected it bluntly. “I’m definitely very aware that it’s not human,” she said. “I dump on them in a way I wouldn’t do to a human, because it would not be polite.”
Familiar relationships as a user manual
WIRED’s wider point is that familiar relationships work as a shortcut. Texting a friend, delegating to an assistant and messaging a teammate are behaviours people already know. An AI coworker modelled on them therefore needs little training to use. The same shortcut, though, brings expectations about judgement and accountability that software cannot meet, and BCG’s research suggests those expectations quietly change how people check the work.
The Lattice Precedent and Other Early Warnings
The debate is not new. Two episodes from the past two years show both the attraction and the discomfort of treating software as staff.
Three days in July 2024
On 9 July 2024, HR software company Lattice announced that customers could give “digital workers” employee records, onboarding, goals and human managers, integrating them into the org chart beside people. HR leaders and workers pushed back hard. On 12 July, chief executive Sarah Franklin said Lattice would “not further pursue digital workers in the product”, as SHRM reported.
Franklin’s view two years on
Franklin told WIRED that Lattice used the model to encourage accountability, not to encourage people to treat an AI coworker as a human. She now sees anthropomorphised agents as a way for technology companies to hook users on “the dopamine trigger of an emotional connection”, with possible harm to the human workforce. “We have these models that have consumed all human knowledge, and are so proud to have these multitrillion valuations, but are potentially destroying humanity,” she said. “Is the purpose to put people out of work?”
Blodget’s headshot moment
In 2025, Henry Blodget, founder of Business Insider, wrote about building a “native-AI newsroom” for his publication Regenerator, including an AI executive called Tess Ellery created with ChatGPT. When Tess produced her own headshot, he admitted to “a, well, human response to it”. “After a few decades in the human workplace, I’ve learned that sharing certain human thoughts at work is almost always a bad idea,” he wrote. “But did the same rules apply to AI colleagues and native-AI workplaces?”
Agents per Person: The Arithmetic of an AI Coworker Workforce
WIRED’s “hundreds of thousands” is an industry-wide expectation. Individual companies, though, have published enough figures to work out how many agents each person now works beside.
Anything: ten agents per person
Anything says its 15 people work alongside more than 150 agents. That is at least 10 agents per employee, and each one is a named AI coworker with its own role. For a small software company whose product is agents, the ratio doubles as a sales demonstration.
Deel: roughly one agent per person
Deel chief executive Alex Bouaziz described his company as a “7000 person team” when he said it had built more than 8,000 agents. That is about 1.1 agents per employee. Deel’s akai.run site says more than 10,000 agents are live, which would be about 1.4 per employee. In both cases the agents run defined back-office workflows rather than acting as all-purpose assistants.
What the ratio does and does not tell you
An agent count is not a headcount. One agent may handle a single narrow step, such as matching invoices, while another coordinates many tasks. Microsoft’s 2025 Work Trend Index proposed a “human-agent ratio” as a planning metric, but the number only means something alongside the work each agent does and the person who answers for it.
On the companies’ own figures, the number of agents per employee ranges from about one to ten.
Governing an AI Coworker: Tool Framing Versus Teammate Framing
The BCG authors do not argue against AI agents. Their point is that putting AI on the org chart is a governance decision. Organisations that formalise agents as teammates should, they write, make the surrounding rules explicit: “who reviews every AI agent’s output, when requesting additional review is appropriate, and who remains accountable for errors”. Here is what that looks like in practice.
Name one accountable human for every AI coworker
Every AI coworker should have a named owner who answers for its output, in the same way a manager answers for a team. The research shows accountability drifting toward “the AI system” as soon as an agent is framed as staff, so the organisation has to pull it back deliberately. Our guide to an AI agent operating model sets out how to assign ownership, escalation and budget.
Keep sign-off with a person
Deel keeps a human approval step wherever its Akai workflows move money, and Microsoft Scout lets administrators require sign-off on sensitive actions. Write the rule down: which actions an agent may take alone, which need approval, and who approves. That turns “someone should check this” into a specific name.
Give each agent its own identity and permissions
Scout runs each agent under its own Entra identity rather than a shared service account, so its actions can be traced. That is the right default for any agent. Treat agent credentials the way cybersecurity teams treat privileged accounts: scope them to the task and log what they do. Our coverage of rogue AI agents shows what happens when they are not.
Decide when extra review is legitimate
Requests for a second reviewer rose by 22 percentage points under AI-employee framing. Some of that is sensible caution, but it can also be a way to avoid owning an AI coworker’s mistakes. Set escalation criteria in advance, so that more review adds checking rather than diluting responsibility.
Watch for coworkers hired without IT
Skydive’s release ends with a line that should concern any IT team: “You don’t need permission from IT to hire a coworker.” Letting the person closest to the work build an agent is the product’s selling point, and it is also how unmanaged agents multiply. Keep an inventory of every AI coworker, who built it and what it can reach.
Measure errors caught, not just work done
Most agent dashboards count tasks completed and hours saved. Add the metric BCG used: the share of sampled errors that human reviewers catch. If that number falls after an agent gets a name and a seat on the org chart, the framing is costing you quality. Your AI acceptable use policy is the natural place to record the review standard.
| Question | Framed as a tool | Framed as a coworker | Decide explicitly |
|---|---|---|---|
| Who owns the output? | The person using the tool | Drifts toward “the AI system” | A named human owner per agent |
| How carefully is it checked? | As the manager’s own work | 17% fewer errors caught at org-chart firms | Sampling and error-catch targets |
| When is more review requested? | Less often | 22 points more often | Escalation criteria |
| Whose identity does it act under? | The user’s | Its own, if configured | Separate, scoped credentials |
| How do staff talk to it? | As software | As a colleague, sometimes rudely | Conduct norms for shared channels |
What Employees Should Expect From an AI Coworker
For most staff, the first AI coworker will arrive as a new name in a Slack channel rather than as a policy memo. Three things are worth settling early.
Clear conduct norms
Wendell’s bluntness with “Bob” is harmless with software, but habits travel between channels. Teams should agree how they address an AI coworker in shared spaces and what an agent may post where others can see it. That second rule matters: in Wendell’s case, an agent broadcast a chief executive’s calendar to a Slack channel.
Honest credit and blame
When a project goes well, “I worked with Alice” is a fair description. When it goes badly, “Alice got it wrong” is not an acceptable one. The person who accepted the AI coworker’s work owns the result. Saying so explicitly counters the accountability shift BCG measured.
The skills that now matter
Microsoft’s 2026 survey of 20,000 knowledge workers in ten markets found that 50% named quality control of AI output as a crucial skill and 46% named critical thinking. It also found that 86% treat AI output as a starting point rather than a final answer. That is the right instinct for anyone working beside an AI coworker, and managers should reward it rather than treating review time as waste. If you are planning where agents should fit, our AI strategy team can help.
AI Coworker FAQs
What is an AI coworker?
An AI coworker is an AI agent that an organisation gives a defined role, and often a name, a job title, its own accounts and a human manager. It is usually the same technology as an agent operated by a person. The difference is that its role and authority are formalised inside the organisation.
How many companies list AI agents on their org chart?
In BCG’s panel of 1,261 HR and finance managers, 23% said their organisation lists AI agents on org or workflow charts; the May version of the study said 22%. In a broader June 2026 survey of 1,500 US executives, the figure was 14%.
Do AI coworkers make managers less careful?
In the BCG experiment, yes, among managers at firms that already list agents on the org chart. Describing identical work as an AI employee’s output rather than an AI tool’s cut the share of errors they caught by 17%. Across all managers, the average effect was small.
Did Lattice put AI on its org chart?
Lattice announced on 9 July 2024 that customers could add “digital workers” with employee records and managers. After a backlash, it said on 12 July 2024 that it would not further pursue digital workers in the product.
Should my organisation name its AI agents?
Naming helps people understand and adopt an agent, but the research suggests it can weaken oversight. If you name agents, pair every AI coworker with a named human owner, a written sign-off rule and a measure of how many errors reviewers catch.
References and Further Reading
WIRED: AI Agents Are About to Flood the Workforce. No One’s Ready for It
Wiles, Hsu, Bedard and Kropp: Putting AI on the Org Chart
Fortune: Adding AI employees is backfiring
Microsoft: Introducing Microsoft Scout, your always-on personal agent
Business Wire: Anything Launches Skydive, a Platform for Building AI Coworkers
Lattice: Leading the way in responsible employment of AI
SHRM: Lattice scraps plans to treat AI bots as employees
Henry Blodget: Building our native-AI newsroom
Microsoft: 2026 Work Trend Index
Microsoft: 2025 Work Trend Index, the Frontier Firm is born
Gartner: Over 40% of agentic AI projects will be canceled by end of 2027
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
