The AI Execution Gap: Building Became Nearly Free, Finishing Did Not

“Integrating AI will not be a simple technology rollout but a reimagining of work itself (processes, roles, skills, culture, and metrics) so people, agents, and robots create more value together.”

McKinsey Global Institute, Agents, robots, and us: Skill partnerships in the age of AI, 25 November 2025^★

The demo lands. The room is impressed. Somebody says the word transformative.

Eighteen months later, nobody can tell you who owns it.

That distance has a name. We call it the AI Execution Gap. It is everything standing between an AI strategy and accountability for something live. It is where most AI budgets are quietly going to die.

What actually broke

Building used to be expensive. That expense did something nobody designed it to do. It forced a conversation.

Nine months and a budget line meant a sponsor. A sponsor meant answers. What is this for? Who will run it? What happens when it breaks?

Nobody held those conversations out of discipline. They held them because finance made them.

Strategy clarity. Architecture design. An ownership model. All three arrived free, as a by-product of cost.

Cost has now collapsed. The by-product went with it. Nothing replaced it.

Most AI strategies read the first half of that and stop. Building is cheap, so the constraint is gone. Wrong. The constraint moved. It now sits entirely on the far side of the build.

Three numbers. Three methods. One finding.

MIT worked through 52 executive interviews, 153 leader surveys and 300 public deployments. 95% of enterprise generative AI pilots produced no measurable profit and loss impact.¹

That number travelled further than it should have, and it earns a caveat: success was defined as direct P&L movement inside a short window, which misses the efficiency and pipeline effects that show up later. Fine. The diagnosis still holds. The failures gather around integration and workflow, not model quality. The same study found vendor-built tools succeeded roughly twice as often as internal builds. That is a verdict on integration discipline, not on software.

JLL surveyed more than 2,200 leaders across 21 countries. 15% have reached the optimising stage of AI adoption. The other 85% are still tracking trends or analysing impacts.²

Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027. The stated reasons: escalating costs, unclear business value, inadequate risk controls.³

Read that list again. Cost. Value. Control.

Those are the exact three things the old build cycle settled before anyone wrote a line of code.

Three research houses. Three methods. One answer. The hard part was never making the thing.

The algorithm is a tenth of your problem

BCG splits an AI transformation roughly three ways. 10% algorithms. 20% technology and data. 70% people and processes.⁴

It is a heuristic, not a measurement. It is useful because you remember it.

Here is the uncomfortable version. If your programme is spending its attention on the tenth, nobody is doing the other nine tenths.

One caution before this gets misread. Nobody is claiming the engineering is easy. Making agentic systems dependable enough to trust in regulated professional settings is hard, specialised and slow. The point is narrower and sharper: technology alone has never once been enough. If the 70% does not land, the 10% is irrelevant.

Which raises a staffing question most executives would rather not answer. If seven tenths of the work is organisational, seven tenths of the capability has to be organisational too. You will not find it in a technology function, however good that function is.

Five things that no longer arrive free

Install these deliberately, because cost is no longer installing them for you.

• Strategy clarity. A business outcome, not a capability. “We are using AI in claims” is not a strategy. It is an observation.
• Architecture design. How this connects to the systems nobody wants to touch. Most of the cost that ambushes people lives here.
• Ownership model. A named person in production, with the authority and the diary space to act.
• Evaluation system. How anyone knows the output is still right. Without it, a system that has quietly degraded looks exactly like one that works.
• Governance enforcement. Not the policy. The mechanism that checks it, and the consequence when it fails.

That last distinction is the one organisations collapse. Plenty have written an AI policy in the past eighteen months. Far fewer can name the person empowered to stop a deployment.

All of it reduces to one question

Who does what?

An ownership model is a role definition. An evaluation system is a role definition plus a cadence. Governance enforcement is a role definition plus authority.

Which is why the real work is the total operating model, and inside it the line between work a person performs and work an artificial mind performs.

Leave that line unwritten and the system does not become ungoverned in any visible way. It becomes governed by whoever happens to sit nearest to it. That is worse, because you cannot see it.

The evidence says treat this as the first task, not the last. Kyndryl surveyed 1,100 senior leaders across eight countries. 79% agree the speed of AI will outpace their workforce, governance and operating models. About 9% are pacesetters, and their first distinguishing behaviour is redesigning roles around AI instead of bolting AI onto structures that never changed. Those organisations were 1.5 times more likely to see AI-related revenue growth.⁵

Intent is not the scarce resource. Mercer found 98% of executives planning organisational design changes within two years.⁶

98% intend. 15% are operating at the maturity where it happens. 9% are getting paid for it.

Four questions. Ask them before you build.

• Who carries this once it is live, as opposed to who champions it during the pilot? Pilot sponsorship is enthusiastic and temporary. Production accountability is a named person whose own objectives move. If nobody's objectives move, nobody owns it.
• What is a wrong answer worth, and whose job is spotting it? One question, not two. The cost tells you how much catching to buy. Cheap and loud errors need a glance. Expensive and quiet ones need a designed control and a person doing it.
• Which outcome, by which date, lets you declare this finished? Not what success looks like in theory. The specific result that lets you stop calling this a project. Work without a stop condition never finishes. It just goes quiet and keeps eating attention.
• What are you stopping to start this? The one nobody asks.

Why the fourth question is the whole game

The first three are competent hygiene. Any decent programme office gets near them.

The fourth is different. It is the only one that puts the filter back.

When building was expensive, the trade-off happened automatically. Fund this, and you could not fund that. The organisation felt the constraint whether or not anyone named it.

Make building nearly free and the trade-off stops happening at approval, because approval no longer costs anything. So it happens later, invisibly, as attention drains away from work already underway.

This is why organisations that went all-in on AI feel busier and slower at once. Nothing was cancelled. Everything was added. Work in progress climbed. Completion did not. The difference piled up as commitments nobody has formally abandoned.

Ask what you are stopping, and the trade-off comes back to the moment of decision, where somebody accountable can make it on purpose.

What capacity actually measures

Here is the reframe that changes how executives look at their AI portfolio.

Capacity is not measured by what you can launch. It is measured by what you can carry all the way home with a named owner attached. That figure never depended on build cost, so AI did not move it.

Size your portfolio against finishing capacity, not building capacity.

Almost nobody does. Which is why so many AI portfolios contain more items than any organisation could plausibly land, and why the honest answer to “how is the AI programme going?” is so often a description of activity rather than of outcomes.

The skill nobody is teaching

Closing this gap needs capability. The capability is not prompting.

The Conference Board found 55% of workers using generative AI or AI agents regularly, and only 33% who have had employer-provided training. Where training exists, it teaches AI literacy and basic prompting rather than managing agents or redesigning workflows. Only 48.0% say they get sufficient time during working hours to build the capability at all.⁷

The people who close an execution gap can look at a confident, well-formatted, plausible output and refuse it. They know what a wrong answer costs in their own domain. They escalate instead of absorbing.

That is judgement, it is specific to the work rather than the tool, and almost nobody is building it.

Africa gets one shot at this

The first digital divide was about access to technology. It took twenty years and enormous money to narrow.

A second divide is forming now, and access is not the problem. AI tools are cheap and everywhere. What is scarce is the capability that turns a tool into an outcome: ownership, evaluation, governance. The 70%.

If that capability pools in a handful of firms and economies, African organisations are not excluded from AI. Worse. They consume AI that somebody else designed, owns and governs, while the valuable work happens elsewhere. That is harder to see than exclusion, and harder to argue with.

The local position is not comfortable. A survey of 426 Western Cape businesses found more than 70% struggling to find workers with the right skills, and most firms spending only 1% to 4% of their wage bill on training.⁸

But there is an asset here that most economies do not have.

South Africa has a large, young, English-speaking workforce already doing structured, quality-assured work in business process and contact centre operations. Follow a defined process. Spot the exception. Escalate rather than improvise. Get measured on quality.

Those are the disciplines of AI supervision. Exactly. Building genuine prompt and supervision capability among young people in those environments is not a social programme. It is the fastest route we have to the 70%, and it is a route most countries cannot take.

Where CCG works

The execution gap is the whole of what we do.

• Strategy clarity and governance enforcement. CCG Advisory works with Boards and executives on oversight, decision rights and accountability. Whether a programme is governable at all is decided here.
• The operating model and the role definitions. CCG Consult works at task and decision level to draw the line between person and machine, and writes the escalation paths that follow.
• The supervisory capability. CCG Learn builds judgement, verification and the confidence to override. Not fluency.
• The evaluation system. CCG Analytics tracks readiness, adoption and sentiment, so you can tell a system that is working from one that has quietly stopped.

Building became nearly free. That is real, and it is good news.

Finishing did not.

Until you can name the owner, price the wrong answer, state the end condition and say what you are stopping, more building capacity buys you more unfinished work. Not more value.

The gap will not close on its own, and no model closes it for you.

If you are working on the execution gap in your own organisation, we would welcome the conversation. Get in touch with CCG.