Five Phases of AI That Every Business Leader Must Understand
Each phase changed what organisations needed to do, not just what the technology could do. Most leadership teams are one phase behind.
This newsletter draws on the five-phase framework introduced by Jordan Wilson in Episode 723 of the Everyday AI podcast, “From AI Chatbot to Autonomous Coworkers” (February 27, 2026). The analysis and management implications that follow are my own.
Most large organisations are now paying for AI. Microsoft Copilot licences run at thirty dollars per user per month. At five hundred users, that is one hundred and eighty thousand dollars a year before implementation, training, or change management. Similar numbers apply to Google Gemini, Salesforce Einstein, and the growing roster of enterprise AI subscriptions that have found their way onto IT budgets in the past two years.
The question being asked in boardrooms right now is a reasonable one: where is the return?
In most cases, the return is disappointing. Adoption is lower than projected, use is shallow. A McKinsey survey found that fewer than one in three employees who have access to generative AI tools use them daily, and the most popular use case, consistently, is writing emails and planning holidays. Meanwhile, the AI agents, workflow automations, and knowledge systems the tools were designed to enable sit largely untouched. Leadership escalates to the vendor, the vendor recommends more training. Generic onboarding follows, the licence renewal conversation becomes uncomfortable.
The problem is not training. The problem is the wrong kind of training. Most onboarding teaches people how to open a tool but what organisations actually need is structured capability building that teaches people how to work differently, how to redesign workflows, how to delegate to machines, how to define what good output looks like. That is a different intervention entirely, and it is the one most organisations have not yet made.
Last week I came across a podcast episode that gave me the clearest explanation I have heard for why this keeps happening. Jordan Wilson of Everyday AI was mapping how consumer AI has evolved since 2022, phase by phase, and something in the sequence made me stop and listen from the beginning again. His framework surfaces a problem that is not about the tools at all. AI has moved through five distinct phases in three years. Each phase changed what organisations needed to do, not just what the technology could do. Most organisations are equipped for phase two, while the tools they have bought operate in phase four. The gap between those two positions is where the investment disappears. This is not a technology problem, it is a management one. And it has a specific, solvable shape.
Think about what happens when a leader deploys the right capability at the wrong moment of organisational readiness. They invest in automation before their processes are legible enough to automate, buy delegation tools before their managers know how to define what they are delegating, licence AI that connects to internal knowledge before that knowledge has been structured well enough to be useful. The organisation spends, trains, and waits for results that the underlying conditions cannot yet support. That is the situation in most organisations that are currently underperforming on their AI investment.
Read what follows as a diagnostic, map your organisation against each phase. The point is not to catch up on three years of AI history. The point is to locate, precisely, the gap between where your organisation is and where your tools assume it to be. That gap is where your management attention belongs.
Summary
AI has moved through five distinct phases in three years. Each one demanded something new from organisations, not just from technology departments.
Most leadership teams are managing today’s phase with mental models built for the previous one.
The organisations pulling ahead in 2026 have rebuilt how knowledge flows, how workflows are structured, and how managers define and evaluate work. The tools they use matter far less than those decisions.
Your organisation can now operate continuously, around the clock. Whether that means anything depends entirely on the foundational work done in the phases before.
Phase One: The Individual Experiment (Late 2022 to Early 2023)
You recognise this phase if...
People in your organisation started using ChatGPT on their own, at home and then at work, without asking permission. The IT department was asked for a policy, and that policy was usually: proceed with caution, or do not use at all.
When ChatGPT launched in November 2022, it reached one hundred million users in sixty days. Nothing in the history of software had moved that fast. But what arrived was a consumer product, not an enterprise tool. It had no access to company data, no memory between sessions, and a confident tendency to invent facts. Evaluated as a business tool, it fell short. Most organisations concluded: not yet.
That conclusion was understandable. The mistake was treating it as a product assessment rather than a trajectory signal. The relevant question was not whether ChatGPT could do the job in November 2022. The relevant question was: if this technology is improving this fast, what will it require of us in eighteen months? Organisations that asked the second question arrived at every subsequent phase with a head start that no competitor could buy from a vendor.
The organisational demand in Phase One was awareness and a position. Neither required large investment, both required a decision at the leadership level about whether AI was something to monitor seriously or to manage as a compliance risk. Most organisations chose the latter and a few the former. And that choice is still visible in their performance today.
The cost of staying curious in 2022 was low. The cost of having stayed incurious is now high.
What this means for your organisation now
Phase One is behind us, but its consequences are not. If AI scepticism is structurally embedded in your leadership culture, if there are no internal champions who built genuine understanding early, if your team still evaluates AI on what it cannot yet do rather than on how fast it is improving, you are carrying Phase One inertia into a Phase Four world. The remedy is not a tool. It is a deliberate investment in building leadership understanding.
Where to focus
Assess honestly whether your senior team’s understanding of AI reflects where it is today, or where it was two years ago. The gap between those two positions is a strategic liability.
Designate someone at a senior level to track AI developments continuously. Not to deploy them. Not to build business cases. To ensure leadership is never surprised by what is coming next.
Invest in structured AI literacy at the leadership level. Not tool training. Conceptual understanding of what AI can and cannot do, and how fast that boundary is moving.
Phase Two: The Enterprise Rollout (2023 to 2024)
You recognise this phase if...
Your organisation purchased enterprise AI licences, most likely Microsoft Copilot, Google Gemini for Workspace, or Salesforce Einstein. IT rolled them out broadly, adoption is lower than projected. Most employees use the tool occasionally, for drafting emails and, as one survey memorably captured it, planning holidays. The ROI conversation is uncomfortable.
Phase Two is where most large organisations are today, and it is where most of the money has gone. Microsoft Copilot alone costs approximately thirty dollars per user per month. At five hundred users, that is one hundred and eighty thousand dollars per year before implementation or training. The premise was reasonable: embed AI in the tools people already use, and productivity will follow. It has not followed in the way or at the pace anticipated.
The reason is structural. Phase Two tools are designed to work on top of an organisation’s knowledge, its documents, emails, shared drives, and internal systems. When that knowledge is well-structured, current, and consistently maintained, the tools perform well. In most organisations, that foundation does not exist. Documents contradict one another, critical expertise lives in the heads of people who have since left, folders have not been reviewed in years. The AI surfaces this dysfunction immediately and at scale. It is not creating the problem. It is making a pre-existing problem undeniable at the exact moment leadership is paying attention.
The organisations that are extracting real value from Phase Two tools have done two things:
Treated knowledge governance as a leadership responsibility, not an IT task
Invested in genuine capability building, not generic onboarding. Teaching someone to open Copilot takes an hour. Teaching them to restructure how they work around it takes a programme.
A better AI tool cannot fix a knowledge infrastructure problem. A management decision can.
What this means for your organisation now
If your Phase Two deployment is underperforming, resist the instinct to escalate to the vendor or run another training session on tool features - diagnose the actual bottleneck first. In most cases it is one of two things: the knowledge the tool is drawing on is not structured well enough to be useful, or employees have not been given a clear picture of how AI changes the way their specific job should be done. Both are solvable, but neither is solved by a better model.
Where to focus
Audit your knowledge infrastructure before expanding AI access. Identify where the most-used information actually lives, how current it is, and who owns it. This audit will explain most of your adoption problems.
Replace generic tool training with role-specific capability building. The question employees need answered is not how do I use this tool but how does this tool change the way I do my job. Those are different questions with different answers.
Assign explicit ownership for knowledge governance at a senior level. AI tools amplify the quality of the knowledge they draw on. Ownership of that quality belongs in the leadership team, not in IT.
Phase Three: The Workflow Agent (2024)
You recognise this phase if...
Your organisation has begun experimenting with automating specific workflows using AI. Some experiments work, many do not, and the team cannot clearly explain why. The more capable the AI becomes, the more it exposes how poorly documented your actual processes are.
In 2024, AI models gained the ability to reason across multiple steps. Rather than answering a single question, they could accept a goal, break it into a sequence of tasks, make decisions at each step, and deliver a finished output. For the first time, you could hand AI a process rather than a prompt. Research a competitor, draft a summary, flag the three most important implications, and format it for the leadership team. A single instruction, a complete output.
The limitation that emerged was not technical, it was organisational. Processes that appeared structured turned out to depend on informal judgment: the experienced analyst who knew which source to trust, the manager who recognised when a summary missed the point, the institutional knowledge that had never been written down because it had never needed to be. Workflow agents cannot operate on unspoken understanding. They execute what they are told, precisely and without interpretation. Where human judgment had been silently filling gaps for years, the gaps became visible.
The organisational demand in Phase Three is process discipline. Specifically, the ability to document not just what a workflow produces, but what governs the decisions within it. Most organisations discovered this discipline was significantly underdeveloped and that discovery is valuable. It is also the reason most Phase Three experiments stall at the pilot stage.
You cannot automate a process you cannot describe. Most organisations cannot describe as much as they think.
What this means for your organisation now
Workflow agents are available, affordable, and in use at your competitors. The limiting factor is process documentation and the organisational willingness to invest in it, it is not technological. Organisations that build this capability now will compound the advantage across every AI development that follows. Those that treat each workflow automation as a one-off technical project will restart from scratch every time.
Where to focus
Select two or three workflows that are genuinely repetitive, rule-based, and currently consuming disproportionate time from skilled employees. Document them fully: inputs, outputs, and the decision logic in between.
Treat the first automation attempts as diagnostics, not deployments. Where the AI fails or produces unexpected results, the failure is telling you something precise about where your process documentation is incomplete.
Build structured capability in workflow design as an organisational skill. The teams that can describe what they do precisely enough for AI to do it will be the most valuable teams in the organisation within two years.
Phase Four: The AI Coworker (2024 to 2025)
You recognise this phase if...
AI is no longer just a tool your employees use, it is beginning to do work alongside them, or instead of them, on sustained tasks: drafting proposals, conducting research, preparing reports, handling correspondence. The question is no longer can AI help with this but who is responsible for what AI produces.
Phase Four represents a genuine shift in the nature of AI in the workplace. Earlier phases involved AI as a tool, something a person operated to do their job more effectively. Phase Four involves AI as a worker, something that accepts a goal, determines how to pursue it, and delivers a result with limited human involvement along the way. Products like Microsoft Copilot Agents, Salesforce Agentforce, and various custom implementations built on models from Anthropic and OpenAI now make this possible at scale in enterprise environments.
The organisational challenge this creates is one that management theory has not previously had to address, because it has never previously been necessary. Delegation to humans works because ambiguity resolves through conversation. A person who receives an unclear instruction asks a clarifying question, uses contextual judgment to fill in gaps, and senses when something feels wrong. An AI agent does none of this, it executes the instruction it received, and the output reflects what was asked, not necessarily what was intended. The manager who has spent a career relying on shared context, on teams that understand what good looks like without being told, finds that AI amplifies their lack of clarity rather than compensating for it.
The organisations extracting the most value from Phase Four are those that have invested in a specific management skill: the ability to define work precisely before handing it off. This means specifying not just the task but the standard against which the output will be evaluated. It sounds elementary. In practice, most managers have never needed to do it at this level of precision, because experienced human teams fill in what is left unsaid. AI does not.
Delegation to a machine requires the same clarity that good management always required, but never before enforced.
What this means for your organisation now
The return on Phase Four investment is directly proportional to the quality of instruction going in. Organisations that train managers to define outcomes, specify quality standards, and build review processes will compound that investment across every AI capability that follows. Organisations that deploy agents without investing in this management layer will find the technology amplifies inconsistency rather than solving it.
Where to focus
Define what a good output looks like before any task is handed to an AI agent. If that definition cannot be written down, the agent cannot reliably produce it, and a human reviewer cannot reliably evaluate it.
Build review processes that evaluate AI output against an explicit standard, the same rigour that would apply to work from a capable but junior team member who is new to the organisation.
Invest in structured training for managers on delegation discipline: how to specify outcomes, how to set constraints, and how to design evaluation criteria. This capability pays back across human teams as well as AI ones.
Phase Five: The Autonomous Organisation (2026 onwards)
You recognise this phase if...
AI in your organisation is beginning to operate without a human actively in the loop. Tasks are scheduled to run overnight, agents access systems, process files, and deliver results before anyone arrives at their desk. The question is no longer how do we use AI but what does the organisation look like when AI works while we sleep.
Phase Five is defined by two interrelated capabilities:
Desktop agency: AI that can operate a local computer, navigate applications, access files, and complete tasks within a user’s own environment, without requiring cloud APIs or IT integration
Scheduled autonomy: AI that can be given a task at the end of a working day and deliver results the next morning, having executed entirely without human oversight. Both capabilities are available today through products including Claude Cowork, OpenAI Operator, and a growing number of enterprise implementations.
The organisational opportunity is significant. Output that previously scaled with headcount and hours can now scale with well-designed instructions. An analyst who leaves a research brief running overnight returns to a completed first draft. A finance team that schedules routine reporting before a weekend returns on Monday to finished documents. The constraint on organisational output is no longer people and time, it is the quality of the instructions and the infrastructure built to evaluate what comes back.
The risk is equally significant. Every governance architecture built over the past two decades assumes that the entity operating inside a system is human, accountable, and correctable in real time. Autonomous agents are none of these things in the traditional sense. A misunderstood instruction, executed at scale overnight, can produce significant consequences before anyone notices. Security, data access, and accountability frameworks designed for human workers do not automatically extend to agents. Organisations that deploy Phase Five capabilities without rebuilding their governance architecture are not moving fast, they are accumulating exposure that will eventually surface as an incident.
The organisation that works while you sleep is only as good as the governance built before it started.
What this means for your organisation now
Phase Five separates the organisations that have done the foundational work from those that have not. Structured knowledge, documented processes, clear management standards, and explicit governance, all of which were recommended in the earlier phases, are now the prerequisites for operating at this level. Organisations that have built them are in a position to extend their working capacity continuously. Organisations that have not will find that autonomous AI produces volume without value: output that arrives in the morning and cannot be trusted without the human review time that was supposed to be saved.
Where to focus
Establish agent-level access controls and activity logging before the first live deployment. The governance architecture must precede the autonomy, not follow the first incident.
Define clearly which systems and data agents are authorised to access, and under what conditions. The boundary between a sandbox and a live environment must be explicit, documented, and enforced technically, not just stated in policy.
Assign explicit human accountability for agent actions at the organisational level. When an autonomous agent makes a consequential error, someone must be responsible. That accountability must be designed in before deployment begins.
The Change That Connects All Five
Underneath every phase is a shift that receives almost no attention. AI systems can be scheduled. Define the task, set the time, return in the morning to finished work. The work happened without you.
Organisational output has always scaled with people and hours. That equation has changed. An organisation that has done the foundational work, defined what good output looks like, structured its workflows, built review processes that evaluate results efficiently, now has access to working capacity that operates continuously.
For organisations that have not done that work, scheduled execution produces volume without value. Volume without value is noise with a cost attached.
Drucker’s formulation is precise: what gets measured gets managed. The corollary is equally precise: what cannot be defined cannot be delegated. The organisations that benefit most from autonomous AI are those that have already answered the definition question precisely enough to evaluate results in the morning without having watched them happen.
The relationship between human and AI has shifted from conversation to management. Whether your organisation can manage in that sense, define outcomes, delegate clearly, and evaluate results, is the most important operational question of 2026.
Map every AI initiative in your organisation against these five phases. If most cluster in Phase One or Two, the constraint is not better tools. The constraint is foundational work on knowledge, workflows, and management clarity that was never done.
That work is what separates the organisations pulling ahead from those watching them do it.