Every AI Vendor Will Tell You Where To Put It. Almost No One Will Tell You Where Not To.
Fifty-four percent of workers bypassed their company’s AI tools in the last 30 days and did the work by hand instead.
That figure comes from WalkMe’s fifth annual State of Digital Adoption report, a survey of 3,750 executives and employees across 14 countries. Another 33 percent have not touched AI at all. Combined, roughly eight in ten people are routing around software their employer is actively paying for.
The standard read on that number is a people problem. Resistance. Change management. Somebody needs more training.
We would push back on that. When more than half your workforce independently decides the manual path is faster, they are usually right. They are not rejecting AI. They are reporting, through behavior, that it was pointed at the wrong work.
The Give And The Take Are Almost Identical
Two numbers sit uncomfortably close together in the 2026 data.
Goldman Sachs economists estimate that workers using AI well save 40 to 60 minutes a day. Meanwhile the WalkMe research found workers now lose the equivalent of 51 working days a year to technology friction, about 7.9 hours a week, up 42 percent from 2025.
The productivity AI hands to people who can make it work is roughly equal to the productivity it destroys for people who cannot.
That symmetry explains the most deflating statistic of the year. A National Bureau of Economic Research survey of about 6,000 executives across the US, UK, Germany, and Australia found 69 percent of firms using AI while 89 percent of managers reported no change in productivity measured as sales per employee over three years.
Enormous adoption. Flat results. The gap between those two facts is not model quality. It is placement.
The Verification Tax
Here is the test we run with clients before recommending any AI build, and it costs nothing to apply yourself.
For a given task, ask how long it takes a competent person to verify the AI’s output compared to how long it would take them to just do the work.
When verification is fast and the work is slow, AI is a strong fit. Drafting a first-pass summary of a long document qualifies, because a human skims it in a minute against thirty minutes of reading. Categorizing a thousand inbound tickets qualifies, because spot-checking a sample tells you whether the whole batch is sound.
When verification costs as much as doing the task, the economics collapse. That is most judgment work. Reviewing a proposal for an unusual client situation. Deciding which of three vendors to advance. Diagnosing why a production line is running slow this week and not last week. To confirm the AI’s answer, a human has to reconstruct the entire reasoning path, at which point they have done the job twice.
That is the hidden cost inside those 51 days. Not the tool failing loudly. The tool producing something plausible that somebody then has to fully re-derive.
Where AI Genuinely Belongs
The pattern across the projects that work is consistent, and it is narrower than the market implies.
High volume, low variance tasks. The same shape of work repeating hundreds of times a week. Classification, extraction, routing, summarization, first drafts.
Work with an obvious right answer. Somebody can look at the output and know in seconds whether it is correct. Pulling PO numbers off scanned invoices has a right answer. Deciding your Q4 strategy does not.
Work that tolerates a draft. The output is a starting point a human finishes rather than something that ships untouched to a customer.
Work where the input is already written down. AI operates on what exists. When the real expertise lives in somebody’s head and was never documented, no model recovers it.
Notice what those four have in common. They describe the boring middle of a workflow, not the decision at the end of it. The most durable AI implementations we see in Wisconsin manufacturing, professional services, and construction are unglamorous on purpose.
Where It Does Not Belong
The honest list matters more, because nobody selling AI will hand it to you.
Low volume, high stakes decisions. Something that happens four times a year and carries real consequence does not justify a system, and it will not have enough examples to build one properly.
Anything where the reasoning is the deliverable. Clients are not buying the recommendation. They are buying a person who can defend it under questioning.
Work built on undocumented judgment. The thirty year estimator who prices a job by looking at a site has knowledge that was never written down anywhere a model can reach.
Relationship moments. The escalation call, the bad news, the negotiation. Companies keep automating these because they are expensive in labor hours, and keep discovering the labor hours were the product.
Anything requiring a guarantee of correctness. Probabilistic systems produce probabilistic output. Contracts, compliance filings, and safety calculations need a different answer than “usually right.”
Stacking Tools Makes It Worse
One more finding worth putting in front of any leadership team mid-rollout: research this year found that piling on four or more AI tools causes self-reported productivity to fall rather than rise.
Every tool carries its own login, interface, quirks, and failure modes. Past a certain count, employees spend more time deciding which tool to open than doing the work. That is a meaningful share of those 7.9 hours a week.
The organizations getting real returns did fewer things. Gallup’s July 2026 data put US workers at a record 47 percent saying their organization has integrated AI, up from 41 percent the previous quarter, and found employees given an actual integration plan were 15 percent more engaged than those without one.
A plan, in practice, means naming which work AI touches and which work it does not. The second half of that sentence is the half that gets skipped.
What To Do With What You Already Built
Most companies reading this are not at the starting line. There is already a tool deployed, a budget spent, and an adoption number somebody reports upward every month.
The instinct in that situation is to push harder. Mandate usage. Track logins. Put it in a review.
The data argues against that firmly. GoTo’s 2026 Pulse of Work survey of 2,500 employees and IT leaders found 60 percent of employees already feel pressured to use these tools, and nearly one in four IT leaders said AI mistakes have already affected customers, clients, or the bottom line. Pushing adoption on a poorly placed tool does not produce value. It produces output nobody checked.
The better move is to shrink the surface. Take the tool off the tasks where verification costs as much as the work, and point it at the two or three places that clear the test. Adoption on a narrow, well-chosen application beats adoption on a broad, badly aimed one every time, and it is far easier to measure.
Nothing about that requires writing off the investment. It requires admitting the scope was set by a vendor rather than by the work.
Ask Your Team, They Already Know
The cheapest AI audit available starts with a question nobody asks.
Go to the people who have already bypassed the tools and ask what they did instead and why. They are not being difficult. They ran the verification math intuitively and made the correct call, and the answer costs you a fifteen minute conversation rather than a consulting engagement.
Then take the tasks they rejected and pull them off the roadmap. Take the ones they described as mind-numbing and repetitive and put those at the top. That list, built from behavior instead of a vendor’s slide, is worth more than most AI strategies we get handed.
Fifty-four percent bypass is not a failure of adoption. It is free diagnostic data about where the work actually is.
At Earthling Interactive, we help organizations across Madison, Milwaukee, and the continental U.S. decide where AI earns its place and where it does not. AI consulting and AI development built around the work your team actually does, including the honest recommendation to leave some of it alone.
The best AI advice we give clients is frequently about what not to build.
