How I think about AI at work

The most expensive mistake in AI adoption is assuming that AI will compensate for the process underneath it. It won't. If your lead qualification was inconsistent before AI, AI will produce inconsistent qualification faster. If your content review was unclear before AI, AI will produce unclear content in more volume. If your sales handoff was leaking before AI, AI will leak the same percentage of a larger pipeline.

This is the most validated finding across everything I've seen in the last two years. The 95% of AI investments that produce zero P&L impact aren't failing because the AI is bad. They're failing because the operation underneath the AI was unclear, and clarity doesn't automate. AI amplifies signal, but it also amplifies noise, and it amplifies noise more reliably than signal because noise is what most organizations have more of.

When I walk into a company that's frustrated with its AI investment, the first thing I look for is what was already broken. The AI didn't create the problem. It made the problem visible, louder, and faster. The fix isn't a different AI tool. The fix is the process the AI was supposed to help.

The 75% of knowledge workers who report individual productivity gains from AI are telling the truth. So are the 95% of companies that report zero measurable business impact from AI investment. Both numbers are real. The gap between them is the story nobody is telling.

At the individual level, AI works. A skilled person can write faster, analyze faster, code faster, summarize faster. The output is real, the time savings are real, the person feels more productive and often is.

At the organizational level, almost none of this compounds. The gains stay trapped at the individual workstation. They don't show up in quarterly numbers. They don't change the customer experience. They don't move the pipeline. They get absorbed into higher expectations, into rework cycles when someone else has to verify the AI output, or into time the person now spends learning the next tool.

The gap is not an AI problem. It's a systems problem. Organizations that close the gap aren't the ones with the best AI tools. They're the ones with the clearest operating design for deciding what to automate, what to review, and what to leave alone. Most companies never got to that design because they treated AI adoption as a tooling decision when it was actually an operating decision.

Most executives I talk to can describe their company's AI strategy. Fewer can tell me, specifically, which tools their teams are using every day. Almost none can tell me what those tools are being used for, how much the outputs are being verified, or where the work is being quietly redirected because the AI got it wrong.

This is the Shadow AI Stack. The tools people are actually using, the workflows those tools have actually embedded into, and the decisions that are now being made partly by AI outputs that nobody reviewed. MIT's 2025 State of AI in Business report found that only 40% of companies have official AI subscriptions, but employees at over 90% of surveyed organizations use personal AI tools for work tasks daily. The official picture and the operational picture aren't even close. Some of those tools are saving time. Some are quietly introducing errors nobody is catching. Some are replicating work the sanctioned tools were already doing. Leadership is making headcount and budget decisions against a picture of AI usage that doesn't match the real one.

This doesn't get fixed by policy memos or tool lockdown. It gets fixed by leadership actually seeing, in specific terms, what their operation has become. That seeing is most of what a Diagnostic produces.

Anyone with a Claude or ChatGPT subscription can now build a working AI tool in an afternoon. Within a year, the floor will drop further. The skill of building a tool is collapsing toward zero as a competitive advantage.

What doesn't collapse: the judgment about which tool is worth building in a specific business context, with specific constraints, serving specific workflows, for specific people. Knowing what not to build. Knowing which manual process is signaling a deeper structural problem that automation would hide. Knowing when a tool replaces judgment in a dangerous way versus when it amplifies judgment in a valuable way.

This is the same pattern that played out in graphic design when Canva arrived. Mid-skill designers got replaced by software. Senior designers, the ones making brand-level decisions about what should and shouldn't exist, got more valuable. The middle evaporated. The same shape is happening now in consulting, analytics, and software implementation. The work that survives commoditization is the judgment work at the operating layer, and the artifacts that prove that judgment was right.

I build custom tools inside client engagements because the tool is proof that the judgment was grounded in technical reality. But the tool is not the deliverable. The judgment is.

Every piece of work has decisions embedded in it. When you automate the work, you don't remove the decisions. You move them. They show up earlier in the workflow, in the form of prompts, parameters, guardrails, review policies, and governance rules. And they typically show up in the hands of people who were never hired to make those decisions.

The marketing manager who used to approve campaigns is now writing the prompt that will generate fifty variants of a campaign. The prompt is now the decision. Is the person who's writing the prompt equipped to make that decision at that scale? Most of the time, no. The company never designed for it. Training didn't cover it. The review processes assume the old workflow where a human made each decision explicitly.

This is the Busywork Trap. People are busier than ever because they are now making decisions they weren't equipped to make, at speeds they weren't trained for, with fewer checkpoints than the old process had. They feel exhausted, the output quality is uneven, and nobody can quite explain why. The tools are working. The process around the tools is failing.

The fix is not more tools. The fix is redesigning who makes which decision, at what point in the workflow, with what information, before the automation amplifies whatever that person decides.