McKinsey consultants spent this year sitting down with executives at 15 companies widely regarded as the best at putting AI to work. The first sentence of the report that came out of those interviews reads like this: almost every company has AI tools, but not many know how to use them well. If you came looking for the secret to success, it's a deflating place to start.
What makes the report worth a look is that it doesn't lead with what these companies do right. It leads with the mistakes they have in common. Even the firms people call AI-native get stuck in the same spots. The patterns were boiled down to seven of them. And a good number of those seven had nothing to do with the power of the tools or the size of the budget. The gap between stacking up tools and actually digesting them turned out to be deeper than expected. That gap is exactly where most organizations are getting stuck right now.
The Stretch Where You Only Think You Got Faster
The first error the report found across the board is the illusion of speed. The expectation that work will simply move faster once AI drafts the first version plays out differently on the ground. Reviewing what AI produces, editing it, deciding which parts to keep, and then communicating all of that within the team actually eats up more time. There are stretches that genuinely got faster. But at the same time, new slow stretches appear. If the gap between how fast the tool generates and how fast a person can absorb it never narrows, overall productivity can land below expectations.
The second pattern is knowledge isolation. The first people to start using a new AI tool are usually one or two enthusiastic team members. The trouble is that their experience stays a personal asset. Six months later, the same people are still using the tool the same way. AI capability gets pinned to individuals rather than the team. As a result, the productivity gap between the people who use the tools and the people who don't only widens over time.
The third is verification avoidance. If you don't explicitly design, from the outset, a system for how AI output will be checked, people end up trusting that output by default. When an error surfaces, someone says "the AI got it wrong," but the deeper problem is that no one designed the verification process in the first place. Among the 15 companies, the difference between those that built this structure explicitly and those that didn't showed up less in their error rate than in how fast they caught errors.
The remaining patterns landed along similar lines. Data siloed by department, forcing every AI project to renegotiate access from scratch. Ambiguity over who actually has the authority to act on the suggestions AI generates. Reading AI adoption as a signal of headcount cuts, so the people whose time should go to operating the tools never get enough of it. And tool fatigue, where new tools keep piling on until just figuring out what to use where becomes a burden of its own.
One thing stands out the first time you see these seven. There's no technical problem in the list. Nothing about GPT-4 being slow, the API being unstable, or costs running too high. They're all problems that arise from how people and organizations take the tools in. Now that AI performance has risen high enough, the bottleneck has shifted away from the tools and toward the habits around handling them.
The People Who Don't Buy This Report
Among CPOs at growth-stage startups who have read the report, one reaction is that "the seven principles are already the operating language of big companies." Apply a verification framework or data-governance discussion premised on an organization of several hundred people to a three-person team or a solo operator, and you end up spending more time on design than on execution. The criticism isn't entirely wrong.
There's a more fundamental objection, too. The argument is that the very act of codifying principles as operating truths can suppress experimentation and iteration. There are actual cases on record of organizations that fail fast and adjust growing faster at AI adoption than organizations that design the structure first. The McKinsey report itself doesn't fully reject this point. Some of the 15 companies reported that an AI-usage culture took root spontaneously, without any formal guidelines. The principles didn't come first; the patterns emerged as experiments piled up.
Separately, the criticism that the seven truths were never measured carries weight as well. An interview-based report leans on executives' own self-perception. The report contains no data measuring, in a controlled way, how much knowledge isolation or verification avoidance actually affects performance. A sample of 15 companies is also limited. And what "ahead on AI adoption" even means is never clearly spelled out.
That said, this skepticism isn't an argument for throwing the report out. The seven patterns are more useful as an after-the-fact diagnosis than as an up-front design. Used as a reference point for checking "where is our team stuck right now," the limitation of being a big-company document shrinks considerably.
Is My AI Routine Actually Working Right Now?
It's true that the McKinsey report is written in big-company language. But inside it are points that connect directly to solo operators and small teams.
It's worth examining whether the AI tools you're using now are actually involved in your decisions. If they stop at generating drafts, that's closer to having set a tool down than to using one. Spelling out which judgments you delegate to AI and which you insist on making yourself reveals your real level of use. The starting point, in particular, is to track how much you revise a proposal draft after pulling it from AI, and what you use as your criteria during that revision.
Your verification routine needs a check, too. If you don't explicitly decide the criteria you use to review the content, plans, and summaries AI produces, at some point the review just disappears. For a solo operator, this process doesn't have to be an elaborate internal system. Deciding in advance on "the three things I have to verify myself in this draft" is already a verification routine. A single simple checklist can make a noticeable difference in the quality of a blog post or a proposal.
The number of tools you hold is a check item as well. Add a tool every time a new one ships, and at some point more of your energy goes into deciding what to use where. The tool fatigue the McKinsey report points to isn't a big-company phenomenon. Once a quarter, reviewing the list of tools you currently use and clearing out the ones that don't actually contribute to your decisions helps you maintain your level of AI use.
There's something the discussions about what capability remains to humans in the age of AI keep observing. The faster a tool's features expand, the wider the gap grows in the ability to judge in what context, and up to what point, to use that tool. That's why simply spending more time on the tools doesn't close the gap. The disposition to read context and know when to hold a judgment is what creates the real difference in capability, and that operates on a different level than how familiar you are with any particular tool.
What McKinsey observed across the 15 companies pointed the same way. Before which tool you use comes what you decide with the tool in front of you. The teams people call AI-native weren't the first to roll out every new feature the moment it shipped. They were the ones that decided in advance which tool to use for which purpose, and had a routine for adjusting quickly when that decision turned out to be wrong.
