Our Take
Most companies buy AI tools; the difference between users and builders is organisational, not technical.
Why it matters
If your team treats AI as a feature purchase rather than a core operating model, you're already behind. The gap isn't capability—it's how you structure decisions, hiring, and workflow.
Do this week
Product and ops leads: audit your three highest-friction workflows this week and map where AI could eliminate a decision step, not just speed one up.
McKinsey's seven operating truths
McKinsey surveyed leaders at 15 AI-savvy companies and identified seven core principles that separate organisations that have genuinely embedded AI from those that have simply acquired AI tools. The research treats AI adoption as an operational discipline, not a technology question. The gap, according to the analysis, is that most organisations still misalign their structure, incentives, and workflows to AI's actual affordances.
The excerpt does not detail all seven principles, but the framing is clear: organisations that 'really know how to use' AI operate differently at a fundamental level. This is not about model choice or compute budget. It's about how decisions flow, who owns outcomes, and where humans and machines split the work.
The operating model is the bottleneck, not the model
Most AI deployments fail not because the technology doesn't work, but because organisations try to paste it into existing workflows and incentive structures. A tool that generates drafts is worthless if the approval chain hasn't changed. A model that spots anomalies is dead weight if no one owns the response.
The 15 companies McKinsey studied have figured this out. They've redesigned roles, hiring criteria, and decision rights around AI output. They've separated the teams that build AI systems from the teams that use them—and given them explicit handoff protocols. Most organisations haven't.
This matters now because the window to reset operating models is closing. In 12 months, the cost of retraining teams and rearchitecting workflows will exceed the cost of doing it today. Companies that delay are committing to inefficiency at scale.
Start with one workflow, not one model
Practitioners should resist the urge to chase the latest LLM release and instead pick one internal workflow—hiring, code review, customer triage, whatever is most painful—and redesign it end-to-end around AI assumptions. That means deciding who validates output, what the SLA is, when human judgment overrides the model, and how errors feed back into training.
The seven principles McKinsey identified are almost certainly variations on themes: clarity of ownership, feedback loops, explicit fallback paths, and metrics tied to business outcomes rather than model accuracy. Build all four into your first workflow before you scale.