If I were a CEO right now
The title of this piece is probably a little presumptuous, as I am not a CEO.
I do, however, spend a great deal of time working with CEOs and their executive teams, and over the last couple of years I have found myself having many of the same conversations.
Some are about operating models, some are about organisational transformation and all of them find their way towards AI.
There appears to be an increasingly accepted view that AI is ready, the opportunities are waiting, and the primary constraint is organizational readiness.
The implication is rarely stated directly, but it is usually there beneath the surface of the conversations with the transformation partners, the hyper-scalers, and the model companies: ‘Come on folks, pull your socks up, there are AI sunlit uplands to inhabit.’
Organisations therefore are told they need better data, better context, better governance, better technology, and greater urgency.
Once those things are in place…disco, AI ROI.
It is a compelling narrative because it is mostly true.
I just don’t think this is the complete picture. What if this was a partial truth but not quite a whole truth?
Yes, organisations are typically nowhere near ready for AI or Agents at scale in their operating models — this is irrefutable — the part that makes the prevailing argument so convincing.
But what if the unspoken assumptive part was only partially true or perhaps not yet true? The part that assumes that the technology and the market are already there — ready and waiting.
It is this part I’d like to kick the tyres of a little. I think there are three reasons that CEOs and their teams do not need to be hitting the panic button just yet regarding AI and Agentic lost opportunity or competitive disadvantage.
Firstly, and the big one, it is almost impossible to price AI and Agents confidently, either today or for tomorrow.
The frontier model providers have moved to token- and consumption-based pricing to try to bolster revenue into their capital-intensive and currently loss-making businesses pre-IPO. What does this mean?
It means that as of today, we do not fully understand the unit economics of an agentic workflow. We lack the financial observability tools to measure the exact ‘cost of goods’ for a single automated business decision.
There is zero transparency and no future confidence regarding how token pricing will evolve over the next 3, 12, or 36 months. While the industry hopes that scale will stabilise and reduce these rates, there is absolutely no future guarantee.
This market and commercial pricing immaturity is precisely why a CFO will balk. Why would any rational CFO replace a department of known human labor costs and predictable productivity volumes with an infrastructure that offers zero cost certainty and completely volatile runtime economics?
This is not an enterprise problem to solve; it is a market maturity one.
The second one is the predictable output of the models that are the engines of the agents.
Yes, LLMs are wonderful, magical and a world-changing technology, but they are still probabilistic and generative by their very mathematical nature.
They are guessing, not reasoning. And yes, better prompt engineering helps, better contextual grounding helps, as does enterprise context graphs and smaller, more focused models (SLMs might be the agentic future). It helps, but it does not wholly fix.
The core tension remains: we are attempting to plug a probabilistic engine that deals in statistical likelihoods into a legacy enterprise architecture that demands strict, determinism. Because they cannot bridge that gap natively, they will still hallucinate (human term for the inevitable statistical mistakes) and still not come close to the predictability of a person.
They are still not Janet in accounts or Steve in the call centre.
My point here is less pessimism or futurology and more ‘pointing out’ that we are going to continue to see massive advances in models which will hit at increasing pace, so we might not quite be there yet, but give it a minute, this one will resolve fast.
Lastly, and probably in some ways the more trifling of the three but also potentially the most consequential, is the regulatory regime. Technology always runs faster than the leaden-footed regulator, but here we have a technology which might not simply have consumer duty concerns but broader societal impacts. Partnering with your industry regulator now to both influence but also understand the likely direction of travel will be critical before you run off and rewire 50% of your business.
Three reasons not to panic about missing the AI race just yet. None of which are wholly in the gift of the enterprise.
So, take a watching brief right?
If organisations cannot directly control the maturity of the market, the pace of technological progress, or the evolution of regulation, then perhaps the strategic question is not how quickly they can deploy AI, but where they should invest while those external conditions continue to evolve?
That feels like a rather more useful consideration because it shifts attention away from prediction and waiting and back towards what the CEO and their team can and should be driving right now.
CEOs cannot decide when frontier models become sufficiently reliable, nor can they decide how inference pricing evolves. They can, however, decide what sort of enterprise they want to become before those questions are resolved.
Get Matt Hopgood’s stories in your inbox
Join Medium for free to get updates from this writer.
And so the killer CEO question would then be — if all the 50–100 AI use cases in your current enterprise AI backlog both deliver and move into production, does this create your desired enterprise target state?
It’s unlikely, right?
Most of these backlogs contain dozens, sometimes hundreds, of opportunities spanning automation, machine learning, copilots, intelligent search, customer experiences, and increasingly agentic systems.
There is nothing inherently wrong with any of those ideas. What strikes me, however, is that many of them eventually encounter exactly the same organisational questions.
How do we govern intelligent systems? Who remains accountable for decisions? How do we control ongoing operating costs? How do we make enterprise knowledge available safely? How do we explain behaviour to regulators, customers, and our own employees?
The individual use cases differ enormously — the enterprise questions are remarkably consistent.
This is because it is typically being approached from a single direction: business unit out, not operating model in. Ask the business to identify the value and then fund their experiments.
But what if we approached this operating model in, rather than asking every business unit for another collection of AI opportunities?
Perhaps organisations should begin by selecting a single business capability and asking a different question altogether:
What would it take to make this part of the business the safest place in the organisation to deploy models and intelligent systems?
If I were leading a bank, perhaps that capability could be Lending. If I were leading an insurer or healthcare organisation, perhaps it could be Claims. Automotive could be dealer management…
Pick somewhere that is value-creating for the business and part of its core proposition, includes colleagues who could benefit from models and agents in respect to effectiveness not only efficiency, choose somewhere with customer proximity, and where there is a regulatory and legal burden on predictable, explainable, and safe outcomes.
The specific capability choice matters rather less than the intent. The objective is no longer to prove that AI works: It is to prove that the enterprise knows how to absorb agents and intelligent systems safely into a real operating environment.
Viewed through that lens, the sponsorship changes almost immediately. AI stops being primarily a technology programme and becomes an enterprise operating model programme owned by the executive team.
The Chief Operating Officer begins designing an operating model where intelligent systems become recognised operational actors alongside people and technology. The Chief Technology Officer focuses on durable control planes, runtime architecture, and sovereign capabilities rather than whichever frontier model happens to lead the market this quarter. The Chief Data or AI Officer begins treating enterprise knowledge as runtime infrastructure. The Chief Financial Officer develops new approaches to forecasting, governing, and attributing runtime AI costs and agent economics. The Chief Risk Officer, Chief People Officer, and Chief Marketing Officer all discover new questions around organisational control, delegated authority, customer trust, and workforce confidence.
Above them all sits the CEO, who must answer perhaps the broadest question of all: what ethical principles should intelligent systems uphold when acting on behalf of the organisation?
The business capability in, not process/person use-case out, will produce value regardless of how quickly AI itself evolves.
Organisations attempting to create genuinely safe operational business capabilities or domains for AI and Agents will inevitably discover that they need new operational capabilities many do not currently possess.
To name (and invent) some of these capabilities:
· Agentic Model Office: A specialized, ring-fenced operational environment — effectively an AI and agent staging environment in the operating model — capable of proving how autonomous agents will perform in runtime, testing their interaction with colleague and customer environments, and establishing production-grade governance before full deployment. Stop testing AI like its technology.
· Bounded Context Management: The architecture required so that enterprise knowledge becomes coherent enough for models and agents to operate reliably. Stop thinking full enterprise ontology and start thinking smaller business capability ontology, semantic layers, and intent boundaries. (Btw, the idea of a single universal enterprise ontology is for the birds). Decomposing the operating model into bounded domain or capability contexts is the next big architecture decision.
· Knowledge Product Management: The practice of treating enterprise documentation as strict operational infrastructure and runtime context and control. This means formatting and managing critical content — like pricing models, regulatory frameworks, and customer segments — as data products with defined version control, distinct ownership, API accessibility, and clear deprecation lifecycles. Think enterprise documentation as runtime control and not regulatory explainability.
· Agentic Cost, Financial, and Value Management: A dedicated FinOps framework for AI and Agents, because runtime economics must be measured, forecast, and governed in ways traditional software rarely required, shifting the focus from predictable upfront capital expenditure (CapEx) to highly dynamic, usage-driven operational expenditure (OpEx).
None of these capabilities exist solely to support AI. They improve the enterprise whether frontier models advance rapidly or slowly.
This is a very different investment proposition from building another fifty disconnected AI use cases — accepting that this is probably already a runaway train.
If frontier models become dramatically cheaper, more reliable, and easier to govern over the next eighteen months, organisations that have invested in enterprise readiness will already possess the foundations required to scale with confidence.
If technology, market economics, and regulation take considerably longer to mature, those same organisations will still have strengthened governance, enterprise knowledge, architecture, financial control, and operating discipline.
Either outcome leaves the organisation in a stronger position than it occupies today. That strikes me as an unusually asymmetric investment.
Which is perhaps why, if I really were a CEO, I would not be asking my organisation for another fifty AI use cases.
I’d be asking my executive team to prove that we know how to safely absorb models, agents, and intelligent systems into one strategically important business capability.
Because my conviction is that scaling AI is not actually about scaling AI.
It’s about building an enterprise that is ready when everything else in the market finally is.
