BCG’s analysis of AI integration depth across large organisations identifies three performance tiers that are unevenly distributed and not converging (BCG, 2025b). Tier 1, roughly 60% of organisations, uses AI reactively: single-step interactions, full human control, no process redesign. Tier 2, approximately 35%, has accelerated existing workflows. Tier 3, 5%, has redesigned the AI operating model: autonomous agents execute multi-step workflows inside designed governance structures, and returns grow with each operating cycle.
In Brief
- Only 5% of organisations have redesigned work around AI capability; 60% remain at the reactive-tool stage.
- The Silo Tax — coordination overhead and learning signals lost at functional boundaries — absorbs the gains that AI investment generates.
- 74% of organisations plan autonomous agent deployment within two years; only 21% have governance designed for it.
- Persistent, cross-functional teams replace project handoffs and functional separation.
- Tier 3 organisations already generate 3.6 times the total shareholder return of laggards.
The AI productivity gap is already measurable
The financial difference is already measurable. Tier 3 organisations generate 1.7 times the revenue growth of laggards, 3.6 times the three-year total shareholder return, and 2.7 times the return on invested capital (BCG, 2025b). Nearly 80% of organisations have deployed generative AI in some form (McKinsey & Company, 2025). Most of those deployments sit at Tier 1. The investment is present. The returns are not.
BCG is specific about what separates these cohorts: Tier 3 organisations have redesigned their operating models. Tier 1 and Tier 2 organisations have purchased the same tools and applied them to unchanged structures. Individual productivity increased. Organisational productivity did not, because the operating model sitting between the individual and the outcome was not part of the redesign.
Silos absorb what AI generates
“The cost of this pattern, including coordination overhead that absorbs capacity, learning signals dying at boundaries, and management attention redirected towards utilisation rather than outcomes, is the Silo Tax. It is the price an organisation pays for being structurally unable to learn in the second mode, and it is why Plan-Build-Run cannot be fixed from within.” (Hodgson, 2026)
In a functionally organised structure, a developer who writes code three times faster still hands that code to another function for testing, then to another for deployment. The speed gain is absorbed at each boundary. Coordination overhead (the meetings, alignment sessions, handoffs, and queued work) does not decrease when individual contributors are faster. It increases, because faster workers generate handoffs faster than downstream functions can absorb them.
AI accelerates work within each function but does nothing to the boundaries between them.
This is not a technology problem. Functional separation was designed to manage scarce specialists and complex cost allocation, not to produce outcomes at speed. When AI is deployed into that structure, it produces more throughput at each stage while doing nothing to the handoffs between them. The Silo Tax scales with the number of contributors, human or AI, because the bottleneck is the routing, not the workers (Hodgson, 2026).
The pattern predates any AI cycle. The telecommunications industry built the infrastructure on which the digital economy runs and, over two decades of continuous investment, produced total shareholder returns far below the cross-sector average. The operating model was high-investment and structurally unchanged: the same combination organisations are repeating now. In our diagnostic work, organisations that have broken functional silos consistently outperform those that have not. The margin shows up in revenue growth, return on capital, and shareholder return. AI is now producing the same dynamic at a faster rate.
Operating model constrains AI returns
74% of organisations plan to adopt agentic AI within two years, yet only 21% have a mature governance model for autonomous agents (Deloitte, 2026b). Those organisations are preparing to deploy autonomous execution into the same structure that is currently failing to convert AI productivity into returns. The structural gap is what determines whether addressing the governance gap changes anything.
Investing in more AI tools, more training programs, and more change management campaigns adds inputs to a system whose outputs are constrained by how the organisation is structured. Workforce upskilling does not reduce coordination overhead: faster individual contributors generate handoffs faster than the downstream function absorbs them. Better AI explainability does not route learning signals to the right decision-makers. Management attention stays fixed on utilisation rather than outcomes, not because leaders fail to understand the priority, but because no mechanism in a functional structure exists to redirect it.
Argyris (1977) named the distinction that matters here. Single-loop learning adjusts actions within the existing logic: the governing assumptions are never examined. Double-loop learning questions the governing assumptions themselves. Technology investment at Tier 1 and Tier 2 is single-loop; it makes the existing structure faster without examining whether the structure is what should change. The organisations at Tier 3 asked a different question.
Three structural moves
Three structural moves determine whether AI investment reaches the organisation or stays at the individual.
Move #1: Team topology aligned with communication structures
The systems an organisation produces will mirror its communication structures (Conway, 1968). A company organised by function produces systems that require cross-functional coordination to operate, and pays coordination overhead on every value-creating activity. A company whose team topology is organised around the product it delivers reduces that coordination at the structural level. Cross-functional, stream-aligned teams that own the full sequence from strategy through to value realised produce a shorter path from AI deployment to outcome, because the team that designs the AI-enhanced workflow is the same team that governs it.
Move #2: Persistent teams replace project handoffs
Project logic hands work off at completion. The knowledge the team accumulated (the constraints discovered in testing, the trade-offs made during development, the dependencies the specification did not anticipate) disperses at project close. The next piece of work starts with no institutional memory of what the previous one learned. In a product operating model, the team that generated the information is the same team that decides what to do next.
The Product Leadership Trinity governs this model: product leadership sets direction, product management translates that direction into an ordered backlog, and the delivery team executes within defined boundaries (Hodgson, 2026). Empowerment without operational management is a structural gap; the Trinity closes it by establishing who ensures operating consistency, who escalates decisions that exceed a team’s authority, and who governs the model itself.
Move #3: Decision boundaries replace approvals
What the executive is accountable for changes under subsidiarity (Hodgson, 2026). It is no longer about deciding on demand; it is about designing the boundary that determines which decisions the team holds and which require escalation. In AI-governed workflows, this distinction is consequential: an executive who defines what agents are authorised to do, what quality standard their output must meet, when human oversight is required, and who is accountable when something goes wrong has built a governance model. One who has not delegated that design to the default behaviour of the systems they purchased.
Redesign multiplies the AI return
Organisations at Tier 3 are generating measurable financial separation from their peers. Their AI initiatives move into production at 62% (compared with 12% at laggards), and those deployments produce returns that grow with each operating cycle, because the operating model was designed to capture them (BCG, 2025b). Every tier has the tool spend. The structural conditions that convert that investment into organisational returns are present only at Tier 3.
AI amplifies what the operating model enables. At Tier 1, AI amplifies execution inside each functional silo, where the coordination overhead still sits between the faster worker and the slower boundary. At Tier 3, AI amplifies execution inside cross-functional teams that own the full sequence, and the structure produces learning rather than absorbing it.
What this means for senior leaders: The three structural moves above are mutually reinforcing. Each reduces the coordination overhead the other two still carry. Organisations that have made all three are at Tier 3, generating returns that reflect the structure they built rather than the tools they purchased. The investment decision that matters is not an AI budget decision; it is an operating model decision. The executive who makes it now is working with a different organisation in eighteen months than the one who inherits the structure that was not designed around what AI can actually do.
References
- Argyris, C. (1977). Double loop learning in organizations. Harvard Business Review, 55(5), 115–125. https://hbr.org/1977/09/double-loop-learning-in-organizations
- BCG. (2025b). AI at scale: Building the future-built enterprise. Boston Consulting Group.
- Conway, M. E. (1968). How do committees invent? Datamation, 14(4), 28–31. http://www.melconway.com/Home/Conways_Law.html
- Deloitte. (2026b). Enterprise AI survey 2026: Governance and readiness. Deloitte Insights. https://www.deloitte.com/global/en/issues/generative-ai/state-of-ai-in-enterprise.html
- Hodgson, M. (2026). Evolve: The Operating Model AI Demands. Zen Ex Machina.
- McKinsey & Company. (2025). The state of AI 2025. McKinsey Global Institute. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
