The AI Value Realisation Pathway: Five Stages to Compounding Returns

In March 2026, we published The AI Value Realisation Pathway — a four-stage framework mapping how organisations move from ad-hoc AI usage to workflow transformation. The framework captured a structural truth: 88% of organisations use AI, but only 6% capture meaningful enterprise-wide value.

Eight months of client engagements and the industry's rapid shift toward agentic AI have revealed what happens after Stage 4. This updated framework adds a fifth stage — and reframes what "workflow transformation" actually means when AI can reason, act, and improve autonomously.

The four-stage model asked: How do you move from prompts to platforms?

The five-stage model asks a harder question: How do you move from automation to intelligent orchestration?

The progression is not about AI sophistication. It is about operational maturity. A Stage 2 organisation and a Stage 5 organisation might use the same foundation models. What differs is how work is structured around those models — whether agents have identity and governance, whether value compounds or resets, and whether the system learns from itself.

The 2026 Inflection

Two shifts have changed the landscape since our original framework.

Agentic AI has moved from experiment to production. McKinsey's 2025 State of AI survey reports that 88% of organisations now use AI in at least one business function — up from 78% a year earlier. Twenty-three percent are already scaling agentic AI systems in their enterprises. Autonomous agents are handling customer support, financial reconciliation, compliance reporting, and fraud detection in production, at scale.

Governance has become a prerequisite, not an afterthought. Gartner forecasts that 40% of enterprise applications will feature AI agents by the end of 2026, up from less than 5% in 2025. But they also warn that more than 40% of agentic AI projects will be canceled by 2027 due to escalating costs, unclear business value, and inadequate risk controls. Governance is no longer slowing organisations down; it's the infrastructure that determines which projects survive.

But here's the structural problem: McKinsey's research shows that only 38% of organisations have scaled AI beyond pilots. The majority remain stuck in experimentation mode — they have AI running but cannot scale it across the organisation. The gap is not technical. It is architectural.

This is what Stage 5 addresses.

Understanding Each Stage in Practice

The overview above gives you orientation. What follows is the operational reality of each stage — what it looks and feels like from inside the organisation, what it costs to stay there, and what it actually takes to move forward.

Select any stage below to explore it in depth.

The movement from Stage 1 to Stage 3 is additive — more tools, better prompts, richer context. Stage 4 is architectural — deterministic logic moves into code, AI handles bounded judgment tasks via APIs. Stage 5 is orchestration — autonomous agents with memory, tool use, and self-correction, governed as a fleet of digital workers.

The majority stuck in experimentation have reached Stage 4 without the infrastructure for Stage 5. They have automation that works but doesn't compound.

The Stage 4 to Stage 5 Transition

The jump from Stage 4 to Stage 5 is where most organisations stall — and it has almost nothing to do with the quality of the models.

Stage 4 solves the prompt problem. Business rules are in code. AI is called via APIs for specific tasks. Value is measurable and real. But it's linear: each new workflow requires explicit engineering, and the system doesn't learn from itself.

Stage 5 solves the orchestration problem. Agents operate end-to-end with memory and tool use. Governance treats them as digital workers with identity, permissions, and audit trails. Human-in-the-loop checkpoints are architecturally defined for high-stakes decisions. Value compounds because the system continuously improves.

The distinction matters because most organisations attempt Stage 5 capabilities (agentic AI) without Stage 5 infrastructure (fleet governance, observability, tiered HITL). The result is what Gartner's 40% cancellation forecast predicts: projects that deploy agents without the operational infrastructure to support them. The gap is not model quality — it's deployment discipline.

The breakthrough is not better agents. It is better governance.

What the 2025 Research Confirms

The patterns we observe working directly inside operational workflows are consistent with what the leading research houses report across thousands of enterprise AI deployments. This is not isolated experience — it is a structural pattern.

BCG's 10-20-70 framework captures the investment reality: 10% of value comes from algorithms, 20% from data and technology, and 70% from people, process, and organisational change. The 70% is what separates Stage 4 (automation) from Stage 5 (intelligent orchestration).

McKinsey's Rewired methodology defines six capabilities that high performers build: strategy tied to value, talent at scale, an operating model that moves at pace, flexible technology, embedded data, and adoption that converts solutions into gains. These capabilities are not Stage 4 prerequisites — they are Stage 5 requirements.

Deloitte's State of AI in the Enterprise draws the line between "Automators" and "Transformers." Automators layer AI onto existing processes. Transformers redesign how work gets done. Only Transformers reach Stage 5.

What It Takes to Move Forward

The jump from automation to intelligent orchestration requires five things operating together. Most organisations attempt two or three. The ones that achieve Stage 5 build all five.

Agent identity and permissions. Treating agents as digital workers with defined scope, access controls, and accountability. Not tools that anyone can spin up — workers that are provisioned, monitored, and governed.

Tiered human-in-the-loop. Architecturally defining where humans must remain in control. Tier 1 decisions can be fully autonomous. Tier 2+ requires checkpoint validation. This is not about slowing down — it's about building the trust infrastructure that enables speed.

Fleet observability. Unified monitoring across all agents. When something fails downstream, you can trace whether it was the agent, an integration, a data quality issue, or a model provider outage. Siloed agents produce siloed failures.

Feedback loops. The system improves itself. Agent outputs are evaluated, edge cases are captured, models are fine-tuned, workflows are optimised. Value compounds because the architecture is designed for it.

Operating model integration. AI is not a capability bolted onto the existing structure — it is part of how the enterprise runs. Roles are redesigned, not just augmented. Career paths account for human-AI collaboration. The 70% that BCG identifies as the value driver.

This is what we do. We embed inside operational workflows, re-engineer them from first principles, build the governance and technology layers around what we find, and then continue to operate it. We live this ourselves — our 120-person team across five regions runs entirely on purpose-built internal platforms that follow exactly this architecture.

The staying is the differentiator. Not because it is clever positioning, but because production agentic AI genuinely requires ongoing operational engineering to sustain and compound its value.

Start with one workflow. Book a conversation to map it, identify where agents can operate end-to-end, and see exactly where the governance checkpoints need to be.

Where Does Your Organisation Stand?

Honest self-assessment is the foundation of effective AI strategy. Most organisations overestimate their maturity — the presence of AI agents is mistaken for agentic operations. The diagnostic below maps your current state across the six dimensions that most reliably predict where value is being realised and where it is being left on the table.

Your result is a starting point, not a ceiling. Most organisations that have reached Stage 5 did not arrive there in a single initiative. They built incrementally — one workflow at a time, one governance checkpoint at a time — until the architecture became self-evident and the compounding returns became visible enough to justify the next step.

---

Sources

1. McKinsey & Company. The State of AI: Global Survey 2025. QuantumBlack, November 2025. Survey of 1,993 respondents across 105 countries. mckinsey.com

2. Boston Consulting Group. AI @ Scale: The 10-20-70 Framework. BCG Artificial Intelligence Practice. bcg.com

3. Boston Consulting Group. From Potential to Profit: Closing the AI Impact Gap. BCG, 2025. bcg.com

4. Gartner. Agentic AI Forecast: 40% of Enterprise Applications by 2026. Gartner Newsroom, 2025. gartner.com

5. Deloitte. The State of AI in the Enterprise. Deloitte AI Institute. deloitte.com

6. Cloud Security Alliance. Agentic AI Governance Maturity Model. CSA Lab Space, 2026. cloudsecurityalliance.org