The 37% Problem: Why AI Adoption Keeps Outpacing AI ROI

Eighty-eight percent of companies use AI now. Six percent can point to real money from it.

That's McKinsey's 2026 State of AI data, not a guess: 1,719 respondents, 97 countries, fielded this past May and June (McKinsey, Aug 2026). If you're an SME founder who already bought the tools and is still waiting for the number to move, this one's for you. It's not another "AI is coming" post. It's what separates the 6% from the other 82%, and it isn't a bigger budget.

What does McKinsey's 2026 data actually show about AI adoption vs ROI?

The headline numbers from the survey:

• 88% of organizations report using AI in at least one business function
• 37% attribute any EBIT impact to that AI use, and most of that 37% put the impact under 5%
• 6% are "high performers," meaning they attribute 5% or more of EBIT to AI. Unchanged from 2025

Sit with that last one for a second. A full year of extra investment, and the high-performer share didn't budge.

Individual productivity did improve: 80% of AI users report getting more done personally (McKinsey, Aug 2026). That gain isn't reaching the P&L at anywhere close to the same rate. Adoption climbed the ladder. Profit didn't follow it up.

Why does redesigning work matter more than adding more AI tools?

Across 2026 coverage of this gap, one pattern repeats: companies that redesign a workflow before automating it land in the high-performer group roughly 2.8 times more often than companies that bolt AI onto the process they already had (CIO.com, 2026 and Resultsense, Aug 2026).

Workflow redesign means rebuilding how the process works, not inserting an AI step into the old one and calling it done. That distinction is most of the gap between the 6% and everyone else.

Most companies ask the wrong question first: "where can we plug in a copilot?" instead of "if AI agents existed on day one, how would we have built this?" The first question keeps every inefficiency the old process had and stacks a tool on top. The second starts from the outcome and works backward.

Once we were building an agent platform for a company listed on the stock market. Their business process is quite simple. They need to issue a monthly report on stock transactions of major shareholders (which is actually a compliance duty for listed companies). Every month they have to gather information from different sources. Each source has its own data format and its own naming convention, which makes the finance team waste a lot of time on cleaning, merging and formatting the data. They said they were hiring 4 people only to do the clean-up tasks. Now they need a way to automate this repetitive process, which is totally feasible with today's AI capabilities. It sounded simple, but not really. A simple and straightforward process is actually simple because the person in charge knows the work and process, all the know-how stays in her head. The first thing we had to do when stepping into this project is not coding the agents, but asking the insightful and valid questions to retrieve her know-how, structure it in a markdown file in a way that agents can understand (not guess) and clean, restructure the data of that finance department in a way that agents can read, retrieve, process and analyze. This human work is the most time-consuming task in every AI transformation project. In other words, we were in fact rebuilding their working process and turning it into a new agent-friendly one. So for the department lead, the expected ROI of this AI transformation project was crystal clear: they wanted to save time, human effort, and own the whole core data and knowledge of this process.

The case against this take, stated plainly: sometimes speed beats perfect process, and a fast tool bolted onto an ugly workflow still beats no tool at all while the redesign gets scoped. That's fair for a stopgap. It stops being fair the moment the "temporary" bolt-on becomes the permanent process, which is what happened to most of the 82% stuck between adoption and payoff.

What do companies that redesign workflows do differently before they automate?

The 2026 research on this converges on a short list:

1. They simplify the end-to-end process first. Automating a broken process scales the breakage faster; the team still cleans up the fallout, just at machine speed now.
2. They assign an owner and a baseline metric to every pilot before it starts, not after it shows promise.
3. They decide up front which parts of the workflow stay human and which change, instead of asking AI to absorb the whole thing unchanged.
4. They run two tracks together: specific use cases for early wins, broader enablement that compounds over quarters.

None of this needs a bigger AI budget. It needs a decision, made before anyone buys a tool, about what the redesigned process should look like.

Why is the skills gap not really a skills gap?

Skills take the blame a lot. Over 60% of UK SMEs name it as their top AI barrier, and 76% of SMEs using AI self-rate as "novices" (Startup Edition, 2026). But skills gaps and process gaps look identical from the outside: AI gets used, nothing structural changes, ROI stays flat.

A skills gap means people don't know how to run the tool well. A process gap means the process was never going to pay off, no matter who's running it. Training fixes the first one. It does nothing for the second.

Here's the uncomfortable test: hand your best-trained person the AI tool and the unredesigned process. Does EBIT move? For most companies I've looked at, honestly, no. The constraint was never the operator.

How should a company measure whether AI redesign is working?

Five things, tracked together, not one at a time:

• Time: did the redesigned process take less human time per unit of output?
• Cost: did the fully-loaded cost per unit drop, not just the software line item?
• Quality: did error rate or rework move, and in which direction?
• Revenue: did the redesign touch anything customer-facing enough to move revenue?
• Capacity: can the team handle more volume without adding headcount?

A pilot that saves time but hurts quality isn't finished. And a pilot that improves quality but adds review overhead that eats the savings isn't finished either. High performers check all five before they call a redesign done. That habit, more than any specific tool, is probably what accounts for the 2.8x.

I'd rather be wrong about the exact multiplier than wrong about the direction: tool-first bets keep missing, redesign-first bets keep compounding. If your experience says otherwise, I want to hear the counterexample, not just the tool name.

Frequently asked questions

What percentage of companies see real ROI from AI in 2026?

37% report any EBIT impact; only 6% report an impact of 5% or more, McKinsey's bar for "high performer" status.

Why do companies with high AI adoption still see low ROI?

Most apply AI to workflows they never redesigned. Companies that redesign first are about 2.8x more likely to reach high-performer status.

Is the AI ROI gap mainly a skills problem?

Not primarily. An unredesigned process tends to fail about as often with a well-trained operator as with a novice.

Key takeaways

• 88% of companies use AI; only 37% see any EBIT impact, and just 6% qualify as high performers, flat versus 2025 despite a year of extra spend.
• Redesigning a workflow before automating it makes a company roughly 2.8x more likely to become a high performer.
• "Where can we add AI?" and "how would we build this if AI existed on day one?" produce different companies. Only the second question closes the gap.
• Training fixes a skills gap. It doesn't fix a process that was never going to pay off.
• Measure redesign success across time, cost, quality, revenue, and capacity, not just the one metric that looks good first.