What AI Really Means for Growth Leaders

The most reliable longitudinal source on enterprise adoption is McKinsey's global survey work, which has tracked organisational AI use year on year. Its 2024 and 2025 findings describe the same pattern: adoption of generative AI in at least one business function has risen steeply, while the share of organisations reporting meaningful enterprise-level financial impact remains far smaller.

Stanford's AI Index provides the useful counterweight, documenting rapid improvement in model capability and steep falls in inference cost alongside continued difficulty in measuring organisational productivity effects. Both things are true simultaneously: the technology is improving quickly, and most organisations are not yet capturing it.

The honest interpretation is that we are early. These are self-reported surveys, subject to definitional inconsistency and to the reluctance of executives to report that an announced initiative has not delivered. Treat all such figures, including those cited here, as directional.

Stripped of the general enthusiasm, AI improves commercial performance at a small number of identifiable decision points.

Note how few of these are technology problems. Each one is a decision that already exists, made today by someone with incomplete information. AI improves it only if the decision is redesigned to consume the new information, and only if the person or system making it is permitted to act differently as a result.

Pilots stall for consistent and largely non-technical reasons, and I have watched most of them at close range.

• No decision owner. A model produces an output that nobody is accountable for acting on.
• Fragmented data. The signal exists but sits across systems with no shared customer identity.
• No integration into workflow. Insight arrives somewhere other than where the work happens.
• Unclear economics. The pilot demonstrates capability without establishing what it is worth.
• Weak governance. Legal, risk and compliance concerns surface late and stop deployment.
• Missing skills. Nobody in the commercial team can interpret, challenge or improve the output.
• Cultural resistance, which is usually rational: people who have seen previous tools add work are correct to be sceptical.
• No feedback loop. The model never learns from the outcomes of its own recommendations.

A pilot that proves a model works but changes no decision has proved nothing that matters commercially.

The organisations capturing real value from AI have generally made four changes that have nothing to do with model selection.

They have redesigned specific decisions end to end, naming the decision, the owner, the input, the action and the measure of success, rather than adding a tool alongside the existing process. They have invested in data foundations, particularly shared customer identity, because personalisation and propensity work are impossible without it. They have built governance early, so that deployment is not blocked at the final approval. And they have raised the fluency of commercial leaders, so that the people accountable for outcomes can challenge a model's recommendation intelligently rather than either deferring to it or ignoring it.

Commercial leaders often experience governance as the function that stops things. In AI deployment, the opposite is closer to the truth: the absence of early governance is the most common reason working models never reach production.

The practical requirements are not exotic. Know what data the model uses and whether you are permitted to use it that way. Be able to explain a decision to a customer, a regulator and an internal reviewer. Test for disparate outcomes across protected groups, particularly in pricing, credit and eligibility. Maintain meaningful human oversight where consequences are material, which means a person with the authority and information to overrule, not a person clicking approve. And monitor for drift, because a model that was accurate at launch is not automatically accurate a year later.

This is becoming a compliance requirement in several jurisdictions rather than a matter of preference, and the organisations that built it early are moving faster now, not slower.

Growth leaders do not need to build models. They need enough fluency to ask five questions well: what decision is this improving, what data is it using, how confident is it, what happens when it is wrong, and how will we know whether it worked.

That level of literacy is achievable in weeks and changes the quality of every AI conversation in the organisation, because it moves the discussion from capability to consequence. It also protects against the two opposite failure modes I see most often: uncritical deference to a model's output, and reflexive dismissal of it.

Expectation inflation is the immediate risk. Current enthusiasm implies transformation timelines that historical technology adoption does not support, and boards that budget on those timelines will conclude too early that AI has failed.

Accuracy and hallucination remain genuine constraints in customer-facing applications. A system that is right most of the time is not acceptable where the exceptions are visible, regulated or harmful, and the cost of verification is often underestimated in the business case.

Bias and fairness require active testing rather than good intentions. Models trained on historical commercial data will reproduce historical patterns of exclusion unless specifically examined for it, and in pricing and eligibility decisions this carries real regulatory and ethical consequence.

Workforce impact should be discussed openly rather than euphemistically. Some roles will change substantially. Organisations that are honest about this, and that invest in transition, retain trust. Those that describe it purely as augmentation while reducing headcount do not.

There is also a credible counterargument to the entire framing here: that the measurable productivity effects of AI at the firm level remain weak, that much of the reported impact is self-reported and unverified, and that a period of disillusionment is likely before any durable gains appear. I find this plausible, and it is a reason for disciplined measurement rather than for inaction. The organisations that build data foundations and decision discipline during a plateau are the ones positioned when the capability matures.

Finally, competitive advantage from AI may prove more transient than expected. Where capability is purchased rather than built, competitors can purchase it too. The durable advantage sits in proprietary data, in decision design and in the organisational ability to act on insight quickly, none of which is available from a vendor.