The 37% Problem: Why AI Is Making People Faster, but Not Making Companies Better

AI has already won the productivity argument.The harder question is why so little of it reaches the P&L..

At TPG, one idea sits at the centre of how wework with our Principals: owning outcomes, not just delivering advice. It isn'ta rule anyone hands down. It's an ambition we share and keep sharpening together, drawing on decades of combined experience of what actually moves abusiness.

Part of my own contribution to that comes fromyears of working in a model where you only got paid for results. No time-basedfees. A share of the savings delivered, or nothing at all. That model teaches you one lesson very quickly, and sometimes expensively. There is a big difference between a saving that is identified and a saving that is banked.

Identifying the saving was rarely the hard part. A sharp analyst with good data could usually find it. The hard part was everything that came after. Someone senior enough had to hold the line in a supplier negotiation. An internal process had to actually change. And a budget owner had to accept that the money was gone from their line and wasn't coming back.

That second part is where the value lived. It was also where most of it was lost.

I kept thinking about that gap while readingMcKinsey's State of AI 2026 report last month. It describes the samegap, now playing out across the whole economy.

The numbers that don't add up

Nearly nine in ten organisations now use AI regularly in at least one business function. Four in five respondents say ithas made them personally more productive. And yet only 37% say AI is contributing positively to their organisation's EBIT. That is essentially the same figure as a year ago, despite another year of heavy investment.

So AI is showing up on people's screens, butit isn't yet showing up in the numbers.

This is not an argument that AI doesn't work.It clearly does. A document drafted in ten minutes instead of two hours is areal gain, and so is analysis that once needed a team of three. But a task thatruns 30% faster does not make a company 30% better. Between the two sits a setof questions no tool answers on its own

• Should this task still exist at all?
• If the process around it were designed from scratch today, what     would it look like?
• Where does human expertise genuinely need to stay?
• And what happens to the capacity AI frees up? Does it get     redeployed, or does it simply evaporate into more meetings?

Those are management questions. Not technology questions.

What the 6% are doing differently

McKinsey identifies a small group of "AIhigh performers." These are organisations that attribute at least 5% oftheir EBIT to AI and report significant value from it. They make up about 6% ofrespondents.

What separates them is not access to bettermodels. Everyone has access to broadly the same technology now. Two thingsstand out instead:

First, nearly three-quarters of them havefundamentally redesigned workflows because of AI, compared with roughly aquarter of everyone else. Second, they are 3.3 times more likely to say theyintend to use AI to transform the business, not just make it more efficient.

A fair caveat is that this is self-reportedsurvey data, and it shows correlation, not proof. But the pattern matcheseverything I've seen in twenty years of watching organisations try to change.Most companies are asking: "Where can we add AI to what we alreadydo?"

The few getting real returns are askingsomething much harder: "If we were building this business today, withAI available from day one, how would we organise the work?"

The first question gets you a faster versionof the organisation you already have. The second can get you a differentorganisation altogether.

When analysis becomes cheap, judgement becomes expensive

There's an easy assumption that as AI getsmore capable, experience matters less. I think the opposite is happening.

For most of the last century, organisationspaid people to find information and process it. That is exactly the work AI isnow making abundant and cheap. Research, synthesis, scenarios and first draftscan now be produced in hours rather than weeks.

What AI doesn't remove is the need to decide.When you have ten well-argued options instead of two, choosing between themdoesn't get easier. It gets harder. The questions that matter move up a level: Whatactually matters here? What do we trust? What is this analysis missing? Whathappens second and third, after the obvious first effect? And, ultimately, whatdo we do?

PwC's 2026 AI Jobs Barometer points the sameway. Even junior roles exposed to AI are increasingly being asked forcapabilities traditionally associated with seniority, including leadership andjudgement. When the production of analysis gets automated, the job that remainsis the judgement call.

Human in the loop, at the right point in theloop

A lot of the AI conversation treats"keeping a human in the loop" as a temporary limitation, a safety netuntil the technology improves. At TPG, we see it differently: the questionisn't whether humans stay involved. It's where they add the mosteconomic value. Increasingly, AI can do the research, prepare the analysis andrun the defined process. The human contribution moves to direction, challenge,context and accountability.

That isn't less human involvement. It is humaninvolvement placed where it is worth the most.

The constraint is the organisation, not thetechnology

McKinsey's own conclusion is telling. Thelimiting factor, its researchers argue, is increasingly the organisation'sability to absorb change.

That rings true. The technology is movingfaster than most companies can redesign themselves around it. And redesigninghow work gets done is not a software rollout. It takes people who understandhow organisations actually behave. They recognise the pattern from a previoustransformation, and they know when the data is telling only half the story.They are prepared to make a call when there is no perfect answer.

In other words, it takes the same thing thatseparated an identified saving from a banked one.

AI × experience

This is why we built The Principals Group theway we did.

The future of senior expertise isn't a choicebetween AI and experienced people. It's the combination. A senior operator withtwenty or thirty years of real decisions behind them, working with AI that canresearch, model and challenge at extraordinary speed, is a differentproposition from either one alone.

Work that once needed a large team and a longengagement can increasingly be done by a small, very senior team. Moreimportantly, it can be done by people who are prepared to stand behind theoutcome, not just the recommendation.

AI multiplies the intelligence available to adecision-maker. Experience decides what to do with it.

If your organisation is sitting on a growingpile of AI pilots and wondering why none of them show up in the numbers, that'sprecisely the conversation we're built for.

Senior Judgement. Delivered.

Sources: McKinsey & Company /QuantumBlack, "The State of AI in 2026: On the Road to ROI" (August2026); PwC, "2026 Global AI Jobs Barometer."

‍