A Four-Phase Model for AI Transformation.

Enterprise AI spending keeps climbing. By any adoption metric, the assistant era has arrived. But adoption isn't transformation, and that's the gap this paper is about.

Most organizations have treated AI as a productivity upgrade to existing workflows rather than an opportunity to fundamentally redesign how work gets done. The result is predictable: individuals feel faster, but the throughput of the business hasn't moved. As one CIO put it after a year of Copilot deployment: "the work went right back to being stitched together by human hands."

The four-phase model presented here is the shared language we use with every client navigating this gap. It maps the deliberate progression from people using AI tools to AI systems doing work, with humans in the roles where humans genuinely add value.

The Gap Today: Individual Speed, Flat Throughput

McKinsey's research on the "agentic organization" confirms a pattern we see constantly: organizations that layer AI onto existing workflows see modest returns. The transformative value comes from redesigning operations around what AI can own, not what it can assist with. Microsoft's maturity models acknowledge the gap between copilot adoption and agent-led autonomy. MIT Sloan argues the management paradigm itself must shift, from supervising human task completion to orchestrating agent-led work.

These frameworks are directionally correct. They are also, in the specific place that matters, silent. They describe the destination without mapping the road. That's the gap the four-phase model fills.

Phase 1 · Human-Led Innovation

In Phase 1, individuals use AI as a collaborative tool to solve real business problems. The human identifies the problem, structures the approach, and uses AI to accelerate execution. The interface is a chat window; the work pattern is call-and-response. This is where every organization starts, and there's nothing wrong with starting here. Phase 1 is where people develop AI intuition, which cannot be skipped.

Phase 1 produces real value. A portfolio company's operations lead who categorizes thousands of customer reviews in minutes instead of days has genuinely accelerated their work. These are meaningful improvements. But Phase 1 has structural limitations that no amount of tool improvement will overcome:

• Knowledge stays locked in individuals. The workflow your supply chain analyst builds lives in their chat history and their head. A colleague with the same problem starts from scratch.
• Work is still human-initiated and human-paced. AI makes people faster, but the process still waits for a person to start it, monitor it, and finish it.
• ROI is nearly impossible to measure. People absorb time savings into other work. Usage metrics don't tell you whether output was used or discarded. Managers intuitively sense gains; quantifying them is elusive.

These aren't failures of Phase 1. They are the natural ceiling of a model where AI assists individuals rather than owning work.

Phase 2 · Skill Codification (the Step No One Talks About)

Phase 2 is the most overlooked step in AI transformation, and the one that determines whether your investment compounds or dissipates. The solutions individuals build in Phase 1 get captured, documented, and packaged as reusable, shareable assets. The terminology varies by platform, Claude calls them skills, OpenAI calls them GPTs, Microsoft calls them plugins, but the concept is consistent: a repeatable, parameterized solution to a defined problem that anyone can invoke without rebuilding it.

When Phase 2 is executed systematically, the output is a strategic asset: the skill catalog, a growing library of portable AI capabilities deployable across teams, business units, and portfolio companies. A customer review mining skill built for one consumer products company can be deployed at every consumer products company in the portfolio. Each deployment refines the skill. Each refinement compounds its value. This is the network effect that turns AI investment into durable competitive advantage.

The most effective approach is crowdsourcing: equip your best operators with the tools and training, then systematically capture the best of what they build. Your top performers already know where the operational leverage is. Phase 2 gives them the framework to codify that knowledge in a form that scales beyond their individual contribution.

Phase 3 · Agent-Led, Human-in-the-Loop Workflows

Phase 3 is the inflection point where AI moves from assistant to participant. Codified skills from Phase 2 become the foundation for autonomous workflows that execute end-to-end, with humans involved only at the points where judgment, approval, or expertise genuinely matter.

The work is no longer human-initiated and human-paced. It's triggered by events, a customer submits an application, a supplier misses a delivery window, a new data set arrives, and executed by agents that coordinate tasks, collect results, apply decision logic, and route exceptions to the right human at the right time.

Human-in-the-loop isn't an afterthought or a safety net. It's a first-class design pattern: the system is architected from the start to know what it can decide autonomously and what requires human judgment. Clear cases get automated decisions. Ambiguous cases get routed to a reviewer with full context and a recommended action. The human's role shifts from doing the work to validating the agent's judgment on the cases that matter.

The ROI Inflection Point

Phase 3 is where the ROI measurement problem disappears. Unlike the diffuse, hard-to-attribute gains of Phases 1 and 2, agent-led workflows produce discrete, countable units of work with calculable costs. You know exactly how many times a workflow executed. You know the compute cost per execution. You know the escalation rate. You can assign a value to each completed unit and a cost to each human intervention.

The math is simply: value of work completed minus cost of agent execution minus cost of human interventions equals ROI. That's the measurement clarity Phases 1 and 2 cannot provide. Early well-designed workflows show 70–85% of cases auto-decided, 15–30% escalated to human review, with timelines of 10–12 weeks to production.

Phase 4 · Agentic Departments

Phase 4 is the natural consequence of Phase 3 at scale. When enough workflows within a business function are agent-led, the function itself can be reconceived as a hybrid department, a team of AI agents and human managers operating together. This isn't a distant theoretical future. It's the logical endpoint of the progression that begins with a single agent-led workflow.

Once an organization has operationalized five or ten workflows within a function, the various steps of a procure-to-pay process, or the multiple verification checks in a compliance function, the aggregate begins to look less like a set of automated tasks and more like a department that happens to be staffed primarily by agents.

MIT Sloan argues management itself must be redefined, from supervising human task completion to orchestrating agent-led work. The agent manager role is analogous to any high-performing team manager: you don't do the work yourself, you ensure the work gets done correctly. Handle escalations. Spot-check quality. Monitor dashboards for anomalies. Tune decision logic based on outcomes. The economics are compelling: agents don't take vacations, don't require benefits, and scale linearly with demand. Early-adopting organizations are reporting 30–50% cost reductions in high-volume functions while maintaining or improving quality.

Where to Start

The four-phase model isn't a rigid waterfall every use case must follow. Some processes, those with well-defined logic, clean data, and established decision criteria, can move directly to Phase 3. Others will spend productive time in Phase 1 as people experiment. The value of the model isn't a mandatory sequence; it's a shared language for understanding where you are and what needs to happen next.

If your organization has deployed AI assistants but hasn't formalized anything beyond individual productivity, the highest-leverage move is to establish Phase 2 discipline. Identify what your best people are building. Evaluate those solutions for reusability. Begin building your skill catalog. This creates the pipeline that feeds everything after.

For organizations ready to operationalize, select a process that is high-volume, rule-driven, and currently bottlenecked by human coordination. The first Phase 3 workflow is the proof point that unlocks everything else.

The organizations that will lead their industries in the next five years aren't the ones with the most AI licenses. They're the ones that figured out how to build systems that own the work.

Phase 1 produces skilled people. Phase 2 produces reusable IP. Phase 3 produces measurable operational leverage. Phase 4 produces a fundamentally different cost structure and capacity model. Each phase compounds the last, and the advantage accelerates with each step.