What Cloud Transformation Taught Me About Agentic Programs.

For the last decade, I have led enterprise cloud transformations across distributed teams in the United States, Mexico, and India. The patterns I have watched play out — across industries, regions, and platform choices — are remarkably consistent.

The hardest part of cloud transformation was never the technology. I expect the same to be true of agentic AI, and I think most of the companies entering this wave are not yet planning for it.

The misleading framing

Almost every CIO I talk to right now describes the present moment as “the biggest technology shift since cloud.” That framing is accurate as a description of the shift, and misleading as a description of the work.

Agentic AI is, yes, the biggest technology shift since cloud. But the work of capturing its value will not be primarily technical. Cloud taught us that the platform was the easy part — the organisational redesign was the work, and it was where most cloud programs underdelivered.

The companies entering the agent era now are about to relearn that lesson the hard way, unless they choose to act on it first.

The number that should be on every board pack

McKinsey estimates that roughly $1 trillion in run-rate EBITDA is up for grabs among Fortune 500 companies from cloud adoption by 2030. “Up for grabs” is the operative phrase. That value is not automatic. It is unrealised, sitting on the table, waiting for the operating-model changes that turn cloud spend into cloud results.

McKinsey names three sources of the leak: unrealised use cases (programs that capture IT savings without unlocking new business value), cloud sprawl (redundant foundations built by uncoordinated teams), and stalled adoption (initiatives that never reach the scale they were funded for). Each of these is an organisational failure mode, not a technology one.

The AI version of the same pattern is already visible. McKinsey’s long-running research on enterprise transformation places the failure rate of major change programs around 70%, and frames the hardest work as organisational rather than purely technical. The people leading AI programs today are running into a problem the industry has already named. The companies that act on that knowledge before their pilots stall will be the ones that outperform.

What to copy from cloud, and what not to copy

Copy the operating discipline: platform standards, shared governance, FinOps-style visibility, security review before scale, and business-unit accountability for outcomes.

Do not copy the timeline. Cloud programs could spend years maturing their operating model because infrastructure shifts moved slower and the regulatory surface was familiar. AI programs do not have that luxury. Model capability, vendor terms, policy exposure, and audit expectations can all change inside a single planning cycle.

The practical move is to compress the cloud operating model into the first ninety days of the AI program: name the business owner, name the governance owner, name the evaluation owner, and name the budget owner before the first production use case is funded.

Three lessons that translate directly

There are three lessons from cloud transformation that I expect to see repeated, with shorter cycles and steeper costs, in agentic AI programs over the next two years.

Lesson 1: The Center of Excellence is a bottleneck unless it includes business-unit ownership.

Every cloud program I have worked on that ran through a centralized Cloud Center of Excellence — without embedded business-unit accountability — hit the same wall. The CoE became a permitting authority rather than a delivery partner. Business units routed around it, standards eroded over time, and the CoE ended up policing a system it no longer had the standing to steer.

AI Labs and AI Centers of Excellence are being stood up right now with the same structural flaw, and the answer is not “more authority for the CoE.” The answer is federated ownership with shared standards: business units accountable for outcomes, the CoE accountable for the platform, the governance, and the patterns that move across teams. A good CoE earns its seat by enabling speed, not by gating it.

Lesson 2: Governance must be designed in from day one. Retrofitting is three to four times more expensive.

Every cloud program I have worked on that deferred governance — security, FinOps, data classification, identity — paid for it later. Without exception. Costs ended up flattening deployments, audit findings stalled production releases, and budgets that were supposed to fund expansion got redirected to retroactive compliance work. The teams that built governance in during week one of their cloud journey absorbed a slower start in exchange for a faster overall path, and ended up paying roughly three to four times less over the life of the program.

AI governance is the same shape of problem on a shorter cycle. Audit trails, model evaluation, risk classification, human-in-the-loop policy — these are not optional, and they do not get cheaper if you wait. Build them in during week one, or pay 3–4× later. I have not seen a third option that has held up at scale.

Lesson 3: The skills gap is sequential, not concurrent.

The cloud programs that tried to train teams during rollout produced fragile systems and burned-out engineers. The programs that worked invested in training before rollout — they accepted a slower start in exchange for a steadier middle and a faster end.

AI skills have a much shorter half-life than cloud skills did. New model capabilities, new orchestration patterns, and new governance requirements are arriving on a quarterly cadence, which makes sequencing more important, not less. The pattern that works is the same one cloud taught: train the team to the pattern first, then ship. Trying to do both at once is what produces the fragile system.

What this means for the next two years

The companies that captured the most cloud value were the ones that redesigned the organisation before they expected returns. The companies whose AI programs will deliver meaningful value in 2027 and 2028 will be the ones that do the same — federated ownership with shared standards, governance built in from week one, skills built before rollout.

None of this is novel. All of it requires the executive sponsorship and the discipline to invest ahead of the visible payoff. That is the part most programs underweight, and it is also the part that determines whether the program clears its return target.

The technology, in the end, will not be the differentiator. The operating model will. That has been true for every enterprise technology wave I have seen up close, and I have no reason to believe this one is different.

When to talk to us

The cloud playbook was paid for in dollars, audit findings, and burned-out teams. The AI playbook does not have to be — but only if the lessons are sequenced into the rollout before the rollout begins. The build-vs-buy decision is one of the choice points where these lessons compound, and the operating-model work is what we run alongside the technical engagement.

Cloud taught us what these lessons cost in years. AI is teaching us what they cost in months.

References

McKinsey Digital, Cloud’s trillion-dollar prize is up for grabs (2021; methodology updated 2024). mckinsey.com
McKinsey & Company, Why do most transformations fail? A conversation with Harry Robinson (2019–2024). mckinsey.com

Request a consultation