Delivered, Not Captured: The Real Question of AI Value

One statistic dominates the debate around AI value: over ninety percent of enterprise AI deployments show no measurable profit impact. Sceptical commentators have glommed onto this as a definitive verdict on agentic AI. Is it really? I would argue it reflects how the technology has been deployed so far. Through this lens, the question of AI value primarily becomes a matter of execution discipline.

We need to distinguish here between the questions "is AI valuable" and "are AI valuations worth it". The former is about value, the latter about price. I will not speculate on the second – valuations are for markets. But the first is answerable, and conflating the two flatters the sceptic and disarms the operator. As we have seen previously, a technology or company can be simultaneously transformative and overpriced.

So let us consider the value question in isolation. Chain-of-thought reasoning has changed the unit of work: an agentic system can now retrieve and analyse information, synthesise and create new outputs, and evaluate conditions to reach a decision – in a goal-directed sequence, not as isolated actions. We should recognise that this represents a lion’s share of modern enterprise labour in many industries.

The knee-jerk objection is complexity and scale. Look at a mature process like product development or go-to-market: hundreds of people, dozens of handoffs, exceptions upon exceptions. How could AI possibly replicate all these intricate steps without years of fine tuning and organisational context? While the challenge is fair, it is shortsighted. We need to stop looking at the process and look at the function.

Consider an enterprise function as a box: inputs go in (information, materials, investment, etc) and outputs come out (decisions, products, revenues, etc). The people and processes are not the function. They are one way of mapping inputs to outputs; built around a constraint – human labour – that in many cases is now obviated. Reframe the function by its inputs and outputs, and the question shifts from "can AI replicate this process" to "can AI take these inputs and produce these outputs."

That said, capability alone does not equal value; it needs to be applied. The leader's role is to know which capabilities unlock value, then ask whether an analogous one, delivered more cheaply or at higher throughput, would contribute the same. Where AI substitutes for a human capability, value is often cost-out. The real prize is the capability with no human analogue: instant issue resolution, real-time personalisation, and complex process orchestration – ideally where it amplifies human ability.

Which returns us to that statistic. In most of those cases the deployments will not have failed for want of AI capability – and even less so today, given the original MIT study is now a year old and the models have moved on considerably since. They failed for one of two reasons: either the capability was never successfully applied, with agents left unembedded in real enterprise processes; or value was delivered but never captured, with the underlying function left unchanged.

Value capture is an old discipline but one feature of this wave is genuinely new. Traditional automation removed discrete chunks: a task, a role, or a team. Agentic AI shaves slivers off these. The saving is real, yet diffuse and hard to capture. It is therefore a distribution problem before it is a change problem. Leaders need to aggregate the savings into something they can target, unlocking value through redeployment or reinvestment. Otherwise, all you end up with are anecdotes.

The potential of AI is not in doubt. Whether organisations do the unglamorous work of capture is. AI technology has answered its half of the question. The other half is up to us.

– Ryan

Cover image by ChatGPT.