Most people still think of AI in organisations as a better assistant: something you prompt, something that helps, something that saves a bit of time. But the real shift underway is much bigger. We are moving from prompting AI to truly delegating work to AI - and the path to digital coworkers is closer than most organisations realise. Microsoft’s new Copilot Cowork is a strong signal of that direction, positioning itself as an execution layer for M365 that can turn intent into action and still keep the user in control.
What is it?
What is changing is not just the quality of AI output. It is the relationship between people and AI at work.
I see the path as four phases, and it maps neatly onto the maturity model I use with organisations, of Assistants, Copilots, Agents, and Autonomous Agents.
Phase 1: Assistants (Prompting) This is where most organisations started. Humans drive the interaction directly through prompts into assistants like Microsoft Copilot or ChatGPT. The AI helps with drafting, summarising, searching, or analysing, but the human is still steering every step.
Phase 2: Copilots (Guided workflow assistance) This is where AI starts acting inside the application and using your work context and data to help move things along. Think of features that can build a first draft of an artefact from an existing file or edit within a file (PowerPoints ‘Create PowerPoint from file’ or Excel Agents mode/Edit with Copilot), but still clearly user-led.
(These first two phases are where most organisations seem to be.)
Phase 3: Agents (Delegation with checkpoints) This is where Copilot Cowork becomes important. Microsoft describes Cowork as a way to delegate work: it can ground a request in your emails, meetings, messages, files, and data, turn that into a plan, keep the work moving in the background, and return for clarification, checkpoints, and approval before actions are applied. That is a meaningful shift from helping with tasks to carrying work forward.
Phase 4: Autonomous Agents (Organisational digital coworkers) This is the logical next step: persistent, context-aware, policy-bounded digital coworkers operating across the enterprise environment. Not free-roaming bots, but governed autonomous agents working within permissions, role boundaries, audit trails, and escalation rules. (Think OpenClaw for the enterprise - the next logical step for Microsoft.)
That is why this matters. What looks like a set of product enhancements is actually a progression from AI helping with tasks to AI carrying work forward.
What does it mean from a business perspective?
This is not just a user experience shift. It is a work model shift. It also explains why the path we are on is not yet fully appreciated.
- Risk changes shape as maturity increases. In the early phases, the main risks are poor answers, hallucinations, weak prompts, or people misusing the tools. As you move into delegation, the risk shifts toward execution risk: acting on the wrong data, in the wrong system, with the wrong permissions, or without the right approval. That is one reason Microsoft is emphasizing user control, permissions, compliance - boundaries around Copilot Cowork.
- Human control becomes more important, not less. The more capable the AI becomes, the more important checkpoints, approvals, escalation paths, and auditability become. Enterprise AI will scale through governed delegation, not uncontrolled autonomy.
- Workforce reaction will shape adoption. If employees experience AI as something being done to them rather than with them, trust will drop and adoption will slow. McKinsey’s 2025 workplace AI research argues that the biggest barrier to scaling AI is often leadership and change management, not employee readiness. Deloitte’s 2025 human capital research also found that 54% of workers and leaders are concerned about blurred lines between human work and technology work.
- This becomes a change management issue, not just a technology issue. Leaders will need to explain where AI is assisting, where it is acting, where human judgment remains essential, and how roles will evolve. Without that clarity, even strong technology will run into understandable resistance.
- The change is easy to underestimate. The interface still looks familiar, so many people see this as a series of useful features. But underneath that, responsibility for moving work forward is starting to shift from human-only execution to mostly supervised AI execution.
- This is bigger than better prompting. Once AI can plan, coordinate, and act with checkpoints, the conversation moves beyond productivity tips. It starts to touch operating models, role design, management practices, and how knowledge work is structured.
- Trust becomes an architectural issue. Organisations do not just need capable AI. They need AI that is permission-aware, auditable, policy-bounded, and able to escalate when judgment is required.
- Digital coworkers are not as far away as many think. If the path from prompting to delegation is already visible in mainstream productivity software, then the move toward organisational digital coworkers is not science fiction, it simply an emerging design pattern . (Again, think OpenClaw for the Enterprise.)
What do I do with it?
The right response is preparation.
- Work out where you are on the path. Are you mainly experimenting with assistants? Using copilots with access to your data? Starting to test agents with human initiation and oversight? Knowing your current phase helps clarify your next step.
- Update your mental model. Stop thinking about AI only as an assistant that waits for prompts. Start thinking about where work can be delegated safely and where human judgment still needs to sit.
- Design control points early. If the future is governed delegation, then approvals, escalation paths, auditability, and role boundaries cannot be left until later. They need to be built in from the start. (Think Risk vs. Friction models.)
- Treat risk differently at each phase. The controls that are enough for an assistant are not enough for an agent. As capability rises, governance needs to mature with it and be an enabler - human control is not a brake on this future. It is the reassurance layer that makes it viable at enterprise scale (use systems thinking - see Sometimes Friction is Good).
- Focus on workflows, not just tools. The bigger opportunity is not simply trying a new feature in Word, Excel, or PowerPoint. It is identifying repeatable work that could move from manual effort to supervised delegation.
- Bring people with you. Training, communications, and lightweight change management matter. People need to understand not just how to use the tools, but how their role changes as AI moves from assistant to coworker.
- Help leaders see the scale of the shift. Many organisations still think they are adopting assistants. In reality, they are starting down the path toward digital coworkers. That is a much bigger conversation.
The real question is no longer just what AI can generate. It is what work AI can be trusted to carry forward, under what controls, and with whose approval.
Further Reading
Copilot Cowork: A new way of getting work done (Microsoft)
Superagency in the Workplace (McKinsey & Company)
Deloitte’s 2025 Global Human Capital Trends (Deloitte)