How is AI use at work evolving from individual workflow to shared, agentic systems?
AI use at work started as a private, one-to-one habit and is now splitting into three layers: individual use, where real judgement gets built; agentic AI, which forces organisations to redesign work and accountability; and shared "multiplayer" systems, still early, where colleagues and agents work together in one session. This shift is not really about tools. It is about organisation design and governance, and most companies are further behind on both than their AI adoption headlines suggest.
How did people first start using AI at work?
Even at its simplest, one person working with one AI (a dyad), there was never a single approach. A large field study of BCG consultants found two distinct styles emerging almost immediately (Dell'Acqua et al., 2026). Some people worked as "centaurs", dividing the task cleanly between what they did and what they handed to the AI. Others worked as "cyborgs", blending continuously with the model throughout, drafting and refining together. A follow-up study of 244 consultants refined this further into three modes, adding "self-automators" who largely handed a task over and accepted the output with little review (Randazzo et al., 2026).
The skilling consequences were not equal:
Mode | Approach | Skill outcome |
Centaur | Divides the task cleanly between human and AI | Keeps building domain expertise |
Cyborg | Blends continuously with AI throughout the task | Builds AI fluency, not necessarily domain depth |
Self-automator | Hands the task over with little review | Builds neither |
Anthropic's own 4D Framework, developed with Professors Joseph Feller and Rick Dakan, arrives at a similar conclusion from a different angle (Anthropic, Feller & Dakan). It describes four competencies that every AI interaction requires at once: Delegation, Description, Discernment and Diligence. What is most useful about this framing is where it puts the human: not as a prompter typing requests, but as an editor-in-chief, accountable for judgement and risk even when the drafting itself is delegated.
Interestingly, Anthropic's internal research backs this up with real company data rather than a lab experiment (Anthropic, 2025). Surveying 132 of its own engineers and researchers, alongside 53 interviews and actual usage data, the company found staff becoming more "full-stack" and faster to iterate. However, the same research surfaced genuine anxiety sitting alongside the gains: some employees worried about losing depth in their own craft, or becoming less able to properly supervise what the AI produced. Skill and outcome, across all three pieces of research, depend on how the individual works with the tool, not on whether they have access to it at all.
What happens when AI use stays hidden inside individual silos?
Working with AI one-to-one is, by design, invisible to everyone else. That privacy can feel like safety. Nevertheless, it also breeds a particular kind of isolation: whatever one person learns about working well with AI rarely spreads to the colleague sitting next to them, because there is no natural mechanism for it to do so.
I felt this myself two years ago, while using AI to build a very specific qualitative data analysis tool (I can't go into more detail than that). I was doing the work in the office, surrounded by colleagues. However, it often felt like being locked away in a lab on my own. It was hard to get a genuine reaction when I was excited about a breakthrough, or genuinely deflated when progress stalled, because no one around me really understood what I was doing. From the outside it just looked like furious typing. To what end, nobody could tell.
This siloing has a name: BYOAI, short for "Bring Your Own AI", where employees use personal AI tools and accounts for work tasks, often without IT visibility or a manager ever finding out (MIT Sloan). It is not a fringe habit. Survey data cited by DevRev found that close to half of employees admit to hiding their AI use from colleagues, and more than half of C-suite leaders do the same (DevRev, 2026). The reason is rarely laziness. It is fear, whether of being seen as cutting corners, of a policy that hasn't caught up with practice, or simply of being the person who has to explain themselves first.
This creates a genuinely awkward problem for the organisations trying hardest to adopt AI properly. It is often not the laggards who resist a company's official AI rollout, but some of the most capable users, the ones who have already built fluent, idiosyncratic workflows across two or three different platforms and see little reason to compress that back down into whichever single tool procurement has approved. Their expertise, ironically, becomes a quiet form of sabotage: the people best placed to model good AI use for everyone else are often the ones least visible to the rollout meant to encourage it. No amount of governance policy reaches a workflow nobody admits to using.
Why do organisations need to redesign work, not just add AI to it?
This is where agentic AI changes the calculation. A chatbot answers what you ask. An agent can be handed a goal and left to plan, act, and adjust across multiple steps, sometimes touching real systems, without a person approving each move along the way. Delegation used to mean handing over a subtask and reviewing the result. With agentic AI, it can mean handing over a goal and only finding out what happened afterwards, if at all.
Deloitte's 2026 Tech Trends report puts a name to what this demands of organisations: a "silicon-based workforce" that complements human talent rather than simply automating what already exists (Deloitte, 2026). Its central finding is worth sitting with:
Only 14% of organisations have deployable agentic AI solutions
Just 11% are actually running them in production
Most enterprises are trying to automate processes designed around humans, rather than redesigning the work itself for an agent-native environment
The report has a name for what happens when this goes wrong: "workslop", where a poorly designed agentic application increases operational burden rather than reducing it
That finding reframes the ambitious headline numbers coming from some of the most AI-forward companies. McKinsey's CEO Bob Sternfels has spoken about the firm treating AI agents as part of its effective headcount, with an ambition that every human employee should be enabled by at least one agent (Sternfels, 2026). Nvidia's Jensen Huang has floated a long-run ratio in the region of one hundred agents for every human employee (Fortune, 2026). At the more extreme end, some organisations have built entire engineering workflows around explicit rules that no human writes or reviews the code an agent produces at all (Deloitte, 2026).
These are genuine signals of where things are heading. However, Deloitte's research suggests most organisations attempting anything like this without first redesigning how work and accountability actually flow are more likely to produce workslop than a functioning silicon workforce. Salesforce's 2026 benchmark backs this up in a smaller way: the average organisation runs around a dozen AI agents, with half operating in complete isolation from one another (Belitsoft, 2026). Headcount ambition has outpaced organisational design almost everywhere.
Who is accountable when an AI agent acts on its own?
Redesigning work for agents also means redesigning who is accountable when something goes wrong, and this is where policy is scrambling to catch up.
Singapore's Model AI Governance Framework for Agentic AI, launched January 2026, is the world's first comprehensive attempt at this: it requires every agent to carry a verifiable digital identity and an audit trail of which agent acted under whose authorisation
NIST launched a parallel AI Agent Standards Initiative in February 2026, built on the recognition that agents are too often treated as generic service accounts with no dedicated identity or accountability controls attached to them at all
The EU AI Act's human oversight provisions already apply to autonomous agents operating in high-risk domains, placing the duty of oversight squarely on the organisation deploying the agent, not the tool provider
The practical thread running through all of this regulatory activity is the same one BYOAI already exposed at the individual level: accountability depends on visibility. A shared session in which several colleagues can redirect a long-running agent capable of acting on real systems creates a governance problem the moment that agent touches anything consequential. Watching an agent work is not the same as controlling it, and an audit trail that only exists after the fact is not oversight, it is an autopsy.
What comes after individual AI use?
Some of the most AI-forward companies are now moving past private, individual AI use towards shared, persistent systems that several colleagues, and sometimes other agents, work with inside the same live session. Gabriel Hubert, CEO of Dust, draws the contrast well: single-player AI means one person learns a better way of doing something, but that improvement rarely spreads; a shared system means the agent participates in the same workflow as everyone else, carries context between people, and lets one team's progress become the next team's starting point (Hubert, in Sifted, 2026).
A pre-registered field experiment of 776 professionals at Procter & Gamble, run by researchers at Harvard, Wharton and Warwick, found that individuals working with AI matched the performance of two-person teams working without it, and that AI collaboration also broke down functional silos between departments (Dell'Acqua et al., 2025). That is one study, not yet a large body of proof, and it deserves to be treated that way. It is nonetheless a genuine early signal that AI-assisted collaboration can reshape not just speed, but how teams share expertise across boundaries.
What should organisations actually do about this?
Individual, one-to-one AI use is not about to disappear, and organisations that treat shared, agentic systems as a straightforward replacement for it are likely to be disappointed. The evidence is fairly consistent that individual use is where judgement, discernment and depth of expertise actually get built. Shared and agentic systems, by contrast, have their strongest evidence so far in execution and cross-team coordination, not in developing any one person's skill.
A more realistic sequence for organisations to work through looks like this:
Make individual AI use visible first. BYOAI is not a discipline problem to police away; it is a signal that trust, not tooling, is the actual constraint. Fix the reason people are hiding their workflow before asking them to share it in a multiplayer system.
Redesign the process before adding the agent. Deloitte's research is clear that automating an existing, human-shaped workflow produces workslop. Agentic AI needs work redesigned around it, not bolted onto it.
Assign accountability before granting autonomy. Decide who is answerable for an agent's actions, with an audit trail that exists in real time, before letting it act on real systems unsupervised.
Move to shared systems once the above is solid, not before. Multiplayer AI's evidence base is genuine but still early, and it inherits every governance problem an organisation hasn't yet solved at the individual level.
This sequencing sits close to the ground ThinkWorkPlay's AI Transformation work is built on: helping organisations redesign how they work before they scale the tooling, rather than after. It is also, on a more personal note, why I am currently studying towards the AIGP certification myself, since the accountability questions raised above are quickly becoming a governance discipline in their own right, not just a technology one.
An organisation that rushes towards shared, agentic systems before it has redesigned its own processes, clarified who is accountable for an agent's actions, and addressed why capable people are quietly working around its official tools, is not skipping a step. It is building a silicon workforce on top of exactly the problems Deloitte's research warns against, and the accountability gaps regulators are only now starting to close.
Frequently asked questions
What is the difference between a centaur and a cyborg approach to AI at work?
A centaur keeps a clear division of labour, deciding upfront what they do themselves and what they hand to the AI. A cyborg blends continuously with the AI throughout the task, drafting and refining together rather than handing off discrete pieces. Both outperform simply accepting AI output unreviewed, but they build different skills: centaurs strengthen their domain expertise, cyborgs strengthen their AI fluency.
Does using AI at work make people worse at their jobs over time?
Not inherently. The research shows the outcome depends entirely on how someone uses AI, not on whether they use it. Centaurs and cyborgs both continue to build real skill. Self-automators, who hand a task over with little review, are the group who build neither domain expertise nor AI fluency.
What is multiplayer AI?
Multiplayer AI describes a shared, persistent AI system that several colleagues, and sometimes other agents, work with inside the same live session, carrying context between people rather than starting fresh with each individual chat.
How many AI agents does the average company use?
Salesforce's 2026 Connectivity Benchmark found the average organisation runs around 12 AI agents, but roughly half of them operate in complete isolation from one another rather than as part of a coordinated system.
What is BYOAI?
BYOAI stands for "Bring Your Own AI": employees using personal AI tools or accounts for work tasks, often without their IT department or manager ever finding out.
If your organisation is somewhere in the middle of this shift, individual AI habits well established but the move to agentic or shared systems still undefined, that gap is exactly where ThinkWorkPlay's AI Transformation work sits. Get in touch if you want a second opinion on where your organisation actually stands.
This is the first in the FrAImeworks series looking at how AI is reshaping the way people and teams actually work, not just the tools they use to do it.
Sources
Dell'Acqua, F., McFowland, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., Lakhani, K. "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality." Organization Science, 37(2), 403-423, 2026. (Originally released as Harvard Business School Working Paper No. 24-013, September 2023.)
Randazzo, S., Lifshitz-Assaf, H., Kellogg, K., Dell'Acqua, F., Mollick, E., Candelon, F., Lakhani, K. "Cyborgs, Centaurs and Self-Automators." Harvard Business School Working Paper 26-036.
Anthropic, with Prof. Joseph Feller (University College Cork) and Prof. Rick Dakan (Ringling College of Art and Design). "AI Fluency: Framework & Foundations" / The 4D Framework.
Anthropic. "How AI Is Transforming Work at Anthropic." December 2025.
MIT Sloan Management Review / MIT Center for Information Systems Research. "What Is BYOAI?"
Sternfels, B., CEO, McKinsey & Company. Comments on AI agent headcount, reported January 2026. Source.
Huang, J., CEO, Nvidia. GTC keynote, reported in Fortune, March 2026.
Belitsoft. "2026 AI Agent Trends" report, reported via Barchart.
Infocomm Media Development Authority (IMDA), Singapore. "New Model AI Governance Framework for Agentic AI." 22 January 2026.
National Institute of Standards and Technology (NIST), Center for AI Standards and Innovation (CAISI). "Announcing the AI Agent Standards Initiative." 17 February 2026.
European Union. "Artificial Intelligence Act, Article 14: Human Oversight."
Hubert, G., CEO, Dust. Interviewed in Sifted, "From chatbots to 'digital teammates': The shift towards multiplayer AI." July 2026.
Dell'Acqua, F., Ayoubi, C., Lifshitz-Assaf, H., Sadun, R., Mollick, E., Mollick, L., Han, Y., Goldman, J., Nair, H., Taub, S., Lakhani, K. "The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise." Harvard Business School Working Paper 25-043, March 2025.
Comments