As many businesses struggle to realise tangible returns from their artificial intelligence investments, Matthew Skelton, co-author of Team Topologies, suggested at QCon London 2026 that the primary obstacles are often organisational rather than technical.
It was at this same conference over a decade ago that the core ideas for the Team Topologies framework first surfaced, and Skelton used this anniversary to note that as many as 80% of firms report no tangible benefit from AI adoption, often because they lack the organisational maturity to govern delegated agency.
Skelton argued that the principles found in the original Team Topologies framework provide the necessary infrastructure for agency. This starts with trust founded on bounded agency, a concept where the authority to act is intentionally constrained by rules and guardrails to ensure delegated initiatives remain governable. A focus on bounded agency clarifies mission and focus across the organisation, answering the fundamental question of what value is being provided and to whom.
By focusing on the ongoing stewardship of long-lived services through stream-aligned teams, companies can create an environment where both humans and AI agents operate within clear boundaries. He noted that organisations already structured for bounded agency in humans will find the transition to agentic systems significantly more straightforward.
A critical hurdle in many implementations is the tendency to grant AI tools unbounded access to data resources, a vulnerability formalised as Excessive Agency (LLM06) in the OWASP Top 10 for LLM Applications. Skelton questioned why a business would grant an agentic AI write access to any data store across the organisation when they would never permit a human to do the same.
Industry research supports this concern; for instance, a report by data security firm Metomic found that 86% of files in collaborative environments like Google Drive go untouched for 90 days, yet often remain indexed by AI agents, creating a massive surface for accidental exposure. Maintaining strict security boundaries is essential not just for safety, but for protecting domain fidelity. When this fidelity is lost, software ceases to reflect the original business intent and instead becomes a convoluted machine that is difficult to govern or understand.
The talk also highlighted the direct mapping between human cognitive load and AI context windows. Just as humans struggle when their mental capacity is exceeded, AI agents begin to lose coherence or hallucinate when they operate outside their defined boundaries.
Managing this load ensures that teams can act as effective stewards rather than just owners. Skelton suggested that stewardship is a more productive framing, as it encourages looking after systems for those who come after, rather than merely possessing a codebase or a specific model.
To scale these successes, Skelton introduced the Innovation and Practices Enabling Team. This specialised team type identifies successful patterns within the organisation and shines a spotlight on them to assist other departments. Companies such as Klarna and the Financial Times have established industry benchmarks for these internal learning models, while EBSCO repurposed $9.1 million annually by optimising its delivery through these methodologies.
JP Morgan provided a major proof-of-concept for the talk's core message of knowledge diffusion over mandate, having reduced 60% of dependencies in its Athena platform using an opt-in model. Rather than enforcing top-down rules, the bank utilised a social dynamic dubbed "friendly FOMO" to drive adoption of its LLM Suite. Skelton emphasised that because technology evolves faster than traditional learning structures, this type of active diffusion, building momentum through shared success rather than mandatory compliance, is the only way for large-scale organisations to achieve the cultural shift required for effective AI stewardship.
These insights underpin Skelton’s forthcoming book, Adapt Together, co-authored with Renee Hawkins. The work aims to operationalise value flow much as DevOps transformed software delivery. Skelton concluded that with technology now evolving faster than organisations can learn, active knowledge diffusion and a deep understanding of the systems teams work with are the only viable response to today's cultural and architectural challenges.