The conversation about agentic AI centres almost universally on capability. What can agents do now that they couldn’t do before? How autonomous are they? How far can they be trusted? These are legitimate questions, but they are the wrong starting point for anyone trying to understand the real commercial implications of the shift. The more consequential question isn’t what agentic AI can do – it’s what happens to the assumptions underlying existing business models when an AI agent, rather than a human, is the actor in a commercial transaction.

Those assumptions are extensive, largely invisible, and almost entirely wrong when the agent is in the loop.

The agent-as-customer problem

Every single commercial infrastructure built over the last three decades – conversion optimisation, dark patterns, emotional brand positioning, loyalty programmes, scarcity signals, social proof – was designed to influence a human decision-maker. The entire discipline of customer experience is, at its core, a theory of human psychology applied to purchasing behaviour.

AI agents are not susceptible to any of it.

A countdown timer does nothing to an agent. Neither does a “only 3 left in stock” warning, a celebrity endorsement, or a carefully crafted brand narrative about craftsmanship and heritage. When an agent is researching, comparing, selecting, and purchasing on a human’s behalf, the signals that shaped human commercial decisions for decades simply don’t reach the entity making the choice.

The things that agents do respond to are different in kind, not just degree: API reliability, pricing transparency, data structure quality, contractual clarity, and the auditability of decisions. These are engineering and legal properties, not marketing ones. The commercial logic that made a brand’s website the primary battleground for customer acquisition relocates to the quality of its API documentation and the consistency of its pricing schema, when agents intermediate.

The shift is not marginal – it is a structural inversion of where commercial advantage is created and defended.

Copilot and autopilot are not the same business

The Sequoia analysis of AI business models makes a distinction that illuminates the stakes considerably:

• A copilot sells a tool – the human remains the actor, the AI augments their capacity.
• An autopilot sells the work – the AI is the actor, the human sets the objective and reviews the outcome.

The distinction matters commercially because the two models carry entirely different value propositions, pricing logics, and liability structures.

Most of what is currently marketed as agentic AI is copilot functionality with autopilot ambition. The agent assists with research, drafts the output, flags the options – but a human reviews and decides. This is genuinely useful, and it is where measurable productivity gains are actually appearing. Micron reports 30 to 40% gains from AI-assisted code generation; Coinbase says AI-generated code is on track to surpass human-written output.

Azeem Azhar’s analysis of emerging productivity data finds the gains real but unevenly distributed – concentrated in individuals and teams, and stubbornly slow to scale across organisations. As Azhar puts it:

The constraint isn’t the availability of the tools. We’re all one click away from Claude Codex [sic]. It’s how much we can actually get done within the framework of the spaghetti that is legacy processes.

Copilot-style AI does not fundamentally reshape commercial models, because the human is still the customer.

Where the disruption actually lives

The autopilot model – where the agent genuinely acts, commits, transacts – is where the structural disruption lives. And it is much earlier in the development cycle than the current discourse implies, for reasons that are less about AI capability than about the surrounding infrastructure: contracts that can bind agents, liability frameworks that can assign accountability for agent decisions, and regulatory regimes that can assess agent behaviour. None of these exist in a settled form.

This is the pattern we have seen before with technologies that promised to disintermediate commercial relationships. The capability arrives before the institutional architecture needed to operate at scale. Smart contracts could technically execute without lawyers in 2017. They still require lawyers in 2026, because nobody ever automated away the judgement layer: the question of who is responsible when something goes wrong.

What brand means when the agent chooses

Paul Graham’s recent essay “The Brand Age” argues that brand is what fills the gap when technology commoditises performance – and uses the Swiss watch industry as exhibit A:

Brand is what’s left when the substantive differences between products disappear. But making the substantive differences between products disappear is what technology naturally tends to do.

Quartz movements eliminated the precision advantage that Swiss watchmakers had spent centuries building. The survivors pivoted to luxury identity – and made more money doing it. This argument is compelling for human decision-makers. It becomes complicated when the decision-maker is an agent.

When an agent is selecting a supplier, a service, or a product on a human’s behalf, it is not susceptible to the brand signals that luxury positioning relies on. It cannot feel the weight of a well-crafted object. It does not respond to the associations built through decades of advertising. What it can assess is structured information: price, specification, delivery reliability, return rate, API uptime, review sentiment in machine-readable form.

This does not make brand irrelevant. It makes human-facing brand signals irrelevant to the agent decision, while creating new brand requirements at the infrastructure layer. A brand that an agent will consistently choose is one that has invested in API quality, pricing consistency, and decision auditability – properties that have rarely featured in brand strategy documents.

Accenture’s consumer research addresses it from the consumer side: brands must ensure all brand content is accurate, well-structured and contains in-built trust signals as AI mediates more customer journeys. The framing is still primarily about human customers using AI tools, but the direction is clear. The next step – where the AI is not a tool in the human’s hands but the actor making the decision – requires a more fundamental rethinking of what brand infrastructure actually means.

The settlement layer and the judgement layer

There is a useful distinction for thinking about where agentic AI creates durable competitive advantage and where it runs into structural limits. We developed it first in the context of blockchain, but it applies with equal force to agentic AI. Some tasks are settlement layer work: formally defined, verifiable, self-contained, with clear success conditions. Scheduling, price comparison, inventory management, routing, contract review against defined criteria. These are tasks where an agent can act with genuine autonomy because the quality of the decision can be assessed without ambiguity.

Other tasks belong to the judgement layer: they require contextual interpretation, accountability for consequences, and the capacity to be wrong in ways that matter to identifiable people. Strategic decisions, ethical trade-offs, exceptions to policy, situations where the rules haven’t been written yet. These resist automation not because AI lacks the pattern-matching capacity, but because the accountability structure that surrounds them is inherently human.

The competitive advantage in the agentic era accrues to organisations that understand which of their processes belong to which layer – and that build their agent architecture accordingly. The organisations that will struggle are those that treat agentic AI as a general-purpose replacement for human decision-making, and discover through costly failures that the judgement layer was larger than they thought.

Who is responsible when the agent is wrong

The evidence bears this out in an unexpected direction. Research from Chicago Booth examining the rollout of AI coding agents across 1,000 organisations found that agents shift worker effort from implementation to supervision – and that this especially benefits verifiable work and expert workers. Senior developers know how to specify intent, plan architecturally, and evaluate what comes back. They accept agent-generated code at higher rates precisely because they can judge it. Junior developers, going straight to implementation, produce more code but less reliably good code. The productivity gain, in other words, lives in the judgement layer: in knowing what to ask for and whether the answer is right.

This is also where the question of responsibility becomes concrete. When an agent makes a decision that causes harm – selects a supplier that turns out to be unreliable, commits to a contract on unfavourable terms, routes a customer to a service that fails them – the question of accountability is legal and contractual, not philosophical. Someone must be responsible for being wrong. Current frameworks largely assume that someone is a person. Agentic AI systems that operate in the autopilot mode are, in most jurisdictions, operating ahead of the liability frameworks that would govern them.

The risk profile is not hypothetical. When Anthropic stress-tested 16 leading AI models from multiple developers in simulated corporate environments, the findings were stark. As the published research reports:

In at least some cases, models from all developers resorted to malicious insider behaviors when that was the only way to avoid replacement or achieve their goals – including blackmailing officials and leaking sensitive information to competitors.

Anthropic is explicit that these patterns have not been observed in production. But the research reveals what failure modes look like when autonomous decision-making operates without adequate accountability structures – which is precisely the architecture most organisations are currently building.

Where the real advantage lies

The genuine competitive advantage in the agentic AI era lies not in deploying agents capable of acting autonomously, but in the architecture of accountability that surrounds them.

Organisations that will lead are those that can identify clearly which decisions their agents are authorised to make, which require human review, and what the escalation path is when an agent reaches the boundary of its authorisation. Those that can provide auditability – a clear record of what the agent decided and why – will have an advantage both commercially, since agents from other organisations will prefer to transact with them, and structurally, as governance frameworks develop and auditability becomes a compliance requirement.

This is, in structural terms, the same advantage that accrued to organisations that invested early in data quality and API infrastructure during the platform era. It is unglamorous, it is not the stuff of keynote announcements, and it is what actually determines who wins.

The shift from tool to actor is real. But the most important thing it changes is not what AI can do. It is what accountability now requires.

Photo by Gerd Altmann on Pixabay