The enterpriseafterthe chatbot.

At World Summit AI in Amsterdam this October, the conversation moved well beyond the question of which model could produce the best answer. I moderated discussions on how a global travel business interprets human intention and how the infrastructure behind AI agents copes with execution, latency and scale. On 7 October, on the ADOPT AI stage, I hosted Xavier Amatriain, Chief AI and Data Officer at Expedia Group, for a session titled "Human intention to intelligent itineraries". On 8 October, on the mainstage, I moderated David Casem, CEO and co-founder of Telnyx, in "Beyond the model: Execution, latency and scale in the age of AI agents".[3][1]

The details differed, but the operational problems were recognisable. It is difficult enough for a company to make one intelligent prediction. It is considerably harder to maintain a reliable chain of decisions across customer systems, proprietary data, approvals and changing commercial circumstances. The executive team still owns the outcome, whatever proportion of the work a system performs.

I have spent much of the past decade examining how emerging technologies shape human behaviour, particularly through my work on social media literacy and online safety. The lesson I keep returning to is how quickly organisations and individuals adapt their routines to a new capability before they fully understand its effects. AI is now becoming embedded in the way people learn, write, decide and work. Enterprise leaders carry the additional responsibility of integrating it into shared processes in which an error may affect customers, colleagues, regulatory obligations or the company's finances. Technology procurement will not answer those organisational questions.

This paper is my interpretation of what the summit's published programme and the wider research suggest for senior leaders. It is written for chief executives, boards and the executives who will be asked to make agentic systems work in practice. Where I refer to a session, I refer to its published title and synopsis; I do not attribute claims to speakers beyond that. It is an executive field guide, not a peer-reviewed study, and the interactive tools within it are conceptual aids designed to sharpen judgement.

Dean McCoubrey, Amsterdam, October 2026

The first wave of enterprise generative AI was organised around the answer. A person asked, a model replied, and the person decided what to do with it. Responsibility stayed obviously with the human because nothing happened until they acted. Agentic systems change the unit of work. The system is given an outcome to pursue, and it may take many steps, across hours or days, before anyone reviews the result.

A mission-oriented AI system begins with a specified outcome. It needs to assemble authorised context, break the assignment into well-defined activities, select appropriate tools or models, track intermediate results, identify uncertainty and route certain decisions to a person. Each piece can appear straightforward in isolation; production reliability depends on the whole chain. The relevant evaluation is therefore how often the end-to-end system produces a correct, useful and appropriately authorised outcome under realistic conditions. A polished demonstration tells management little about that. Evaluations such as METR's work on task time horizons show how capability on longer tasks is being studied, and also how much the results depend on the tasks and definitions chosen.[9]

Model selection changes when decisions have consequences. The most capable model is not always the economical option for a narrow task. Some work benefits from smaller or specialised models, deterministic code, conventional search or a human expert. Other tasks justify more expensive frontier reasoning. Leaders need to know where accuracy, latency, privacy and total cost shift the decision, and those points differ by task.

A route that calls a frontier model six times and a verification model twice can have very different economics from a simple single-answer assistant. Both should be tested against actual business tasks before procurement decisions harden into dependency. Open-weight and distilled models offer more control over data residency and cost, at the price of in-house evaluation and operations. Model routing can reduce cost while preserving quality on difficult cases, but it adds another component that must be monitored. Vendor concentration is a strategic risk in its own right; the Stanford AI Index documents how capability and investment have clustered among a small number of developers.[6]

The summit programme gave substantial space to infrastructure, inference and deployment economics.[1] McKinsey's October paper on an "AI operating system" for the enterprise proposes treating these choices as part of how the company is run.[7] That is a consultancy's proposal and should be read as one, yet its central point is sound: model choice is an operating-model question with a recurring cost, and it deserves the same scrutiny as any other long-term supplier commitment.

The economic case for agentic AI is usually presented as time saved. Time saved is a capacity figure, and it becomes money only if the organisation does something with that capacity: handles more volume without hiring, redeploys people to work that earns revenue, or reduces cost deliberately. Each of those is a management decision with its own consequences. Boards should therefore ask for five distinct measures and resist letting one stand in for another: individual productivity, EBIT contribution, revenue, customer outcomes and measured error rates.

The evidence base reflects that gap. McKinsey's respondents widely reported productivity benefits while only a small minority attributed material EBIT impact to AI, and around one in five had constrained some use because of operating costs.[4] KPMG's US chief executives were similarly measured about agents.[5] Operating cost deserves particular attention with agents, because a single task may trigger many model calls, and oversight time is a real cost that pilot business cases often omit.

One consequence for boards is that AI strategy belongs within enterprise strategy, operating-model design and risk management. The board should expect a portfolio of clearly valued opportunities with named business owners, evidence of system reliability, an explanation of the full cost to operate, workforce implications and explicit criteria for scaling or stopping. The first 90 days can establish those habits. The objective is to fund a small number of measurable business improvements and build the company's capacity to make better decisions about what comes next.