TL;DR
Too Long, Don't Read? No problem: here's the Promptcast 😉
Normally, a convincing management study requires a global consulting firm, 1,719 respondents from 97 countries and probably more PowerPoint slides than the average printer can survive in its entire lifetime.
Or you read it's promp.td.
Not entirely serious, of course. McKinsey has an empirical reach no independent blog can match. The new study “The State of AI in 2026: On the Road to ROI” is based on a global survey conducted in May and June 2026 and shows how far companies have actually come in using, scaling and monetising AI.
Still, it's a little bit serious. Because McKinsey's numbers describe, with surprising accuracy, several problems whose mechanisms were analysed on it's promp.td earlier. And some questions that have long been asked here don't even appear in the study's measurement model yet.
89 percent use AI. So where's the ROI?
89 percent of the organisations surveyed use AI regularly in at least one business function. 44 percent are already scaling it across the enterprise. Yet only 37 percent attribute any positive EBIT contribution to AI at all, practically the same as in 2025. A mere 6 percent (!) qualify as “AI High Performers” with at least 5 percent EBIT impact and significant value contribution.
At the individual level, the picture looks completely different: 80 percent report higher personal productivity, around half report better decisions and new skills.
That's no contradiction. It's the difference between an accelerated task and an improved company. If you build a presentation in an hour instead of a day, but then still have to push it through seven approval loops and four meetings, you've optimised your task. Not the process.
Or, as it was put here back in 2025:
It's always the Process, Stupid!
AI doesn't fix a bad workflow. It just makes it produce nonsense faster.
Behind this lies the POC trap: someone discovers a tool, a team builds an impressive proof of concept, management applauds, and then remarkably little happens. Because a POC only answers whether a technology can do something under lab conditions. Not whether a company can change its real end-to-end process, its roles, data and incentive systems in a way that creates lasting value.
McKinsey's high performers confirm exactly that: almost three quarters of them have fundamentally redesigned their workflows with AI, compared with only about a quarter of the remaining companies. They are also 3.3 times more likely to aim for a fundamental transformation of the business rather than mere efficiency.
So the successful ones don't install AI into the existing organisation. They change the organisation.
The thing that actually needs to scale isn't the model, the prompt or the copilot. It's the redesigned end-to-end process.
McKinsey has now supplied the bar charts for that. Thanks, they weren't in the budget here. 😉
Coding is changing “build or buy”
The thesis that coding will be the killer app of generative AI also gets remarkable backing.
Around a fifth of companies are already scaling coding agents; among companies with more than one billion US dollars in revenue, it's 31 percent. More importantly: 32 percent have skipped at least one software purchase because they could build the required functionality in-house with agentic coding tools. Among high performers, it's almost half.
That means more than “developers code faster”. It changes the economics of software. When custom software gets cheaper to produce, companies no longer have to squeeze every process problem into an off-the-shelf solution. The question shifts from “Which software can we buy?” to “Which process do we need, and which part of it can we generate ourselves?”
Budgets may therefore migrate from licences to internal process logic, integration, architecture and quality assurance. Exactly the shift it's promp.td had predicted.
Agents are organisational design
Among large enterprises, the share scaling AI agents in at least one function rose from 27 to 40 percent within a year. Among smaller companies, it stayed almost flat at 22 percent.
But the interesting question isn't just whether agents are coming. It's this: what happens when software no longer just answers, but becomes an acting participant in the organisation?
An agent needs context, system access, roles, permissions, escalation rules, verifiable intermediate steps and a clear line between autonomous action and human decision.
That is precisely the organisational maturity McKinsey finds more often among high performers: human-in-the-loop rules, risk management, impact measurement, strategic workforce planning, workflow redesign and semantic data structures. In most of the functions examined, they scale agents more than three times as often as other companies.
That's why agentic AI is not an ordinary software roll-out. It's organisational design.
After all, an agent without context, rights and controls isn't a digital employee. It's a very eloquent intern someone accidentally handed the master key.
Now the prompt tax becomes visible
McKinsey also finds that AI-related operating costs, tokens included, are already limiting usage at 20 percent of organisations, among high performers especially with heavy use of coding agents. At the same time, 60 percent plan to increase their AI investments further.
This exposes a problem described here as the prompt tax: if a company has to send the same rules, examples, documents and experience to an external model with every call, it keeps paying to explain its own business to the AI.
Falling token prices don't solve this automatically. Context volumes, reasoning steps, agent loops and usage can grow faster than unit costs fall. AI is not “too cheap to meter”.
So the metric that matters isn't the price per million tokens. It's the cost per correctly completed process instance, including retries, oversight, integration and human rework.
McKinsey is now measuring the symptom. The structural question behind it: How long does a company want to keep renting back its own knowledge on demand?
What McKinsey isn't measuring yet
Precision matters here. When a study doesn't confirm a thesis, there are three possible reasons: the thesis is wrong. The phenomenon isn't visible enough yet. Or the study never collected the categories that would reveal it.
For several of the theses on it's promp.td, it's mainly the third:
- McKinsey measures whether companies use specialised AI tools, but not whether small specialist models handle routine processes more economically than frontier models.
- The study measures AI operating costs, but doesn't break them down by repeated context, RAG, fine-tuning, local deployment or ownership of models and feedback data.
- It measures total headcount, but not whether entry-level tasks are disappearing and the talent pipeline is eroding.
The latter is what was described here as the “seniorisation of juniors”: in 2025, 32 percent expected an AI-driven reduction in headcount; in 2026, it actually happened at only 14 percent. But that doesn't refute the thesis. It isn't about mass layoffs, it's about the career ladder. When AI takes over research, boilerplate code and first-pass analysis, headcount initially stays stable while exactly the tasks disappear on which newcomers used to learn their trade. McKinsey counts heads. The more exciting question is:
Where will tomorrow's senior heads come from if the junior tasks disappear?
McKinsey measures symptoms. it's promp.td described mechanisms
What McKinsey measures in 2026 | The mechanism it's promp.td described |
|---|---|
High personal productivity, but little EBIT impact across companies. | Task acceleration without process redesign doesn't produce company-level ROI. |
High performers fundamentally redesign workflows. | AI transformation is business process optimisation, not tool roll-out. |
32 percent replace at least one software purchase with an in-house build. | Generated code shifts the line between build and buy. |
AI costs are already limiting usage. | Repeated context creates a prompt tax. |
High performers use semantic layers or knowledge graphs more often. | Institutional knowledge becomes a strategic AI asset. |
Agents scale mainly at organisationally mature large enterprises. | Agentic AI requires rights, context, governance and process architecture. |
Headcount reductions fall short of expectations. | Early disruption shows up in tasks and career paths, not immediately in headcount. |
McKinsey shows very well what is happening. In several cases, it's promp.td had described earlier why it would happen.
A brief package insert: McKinsey's study relies mostly on self-reporting, the high-performer group comprises just 92 respondents, and correlation isn't causation, not even in a pretty chart. Conversely, a blog thesis doesn't become empirical research just because later data happens to fit. The comparison is only sound if the earlier claim, its publication date and the later finding sit side by side in a traceable way, including the predictions that are still open or turned out wrong.
So the defensible claim is not that McKinsey is copying from here. That would be nonsense, if extremely flattering. 😉 It is this:
McKinsey's 2026 data support several explanations of AI implementation previously put forward on it's promp.td. Other theses go beyond the study's measurement model and remain testable predictions.
Why can a blog be faster here? Not because of better data. Quite the opposite. A large survey needs stable terms and sufficiently widespread phenomena; it works a posteriori. An analytical blog can ask ex ante: if technology, costs and organisational logic are heading in this direction, what problem will that create tomorrow? That's faster, but also riskier. Hence the rule: document predictions, don't mistake hits for certainty, and make mistakes just as visible as confirmations.
What's probably coming next
Particularly interesting is McKinsey's observation that high performers more often use a semantic data layer or a knowledge graph. A knowledge graph answers the question: How do we make our company knowledge accessible to AI? Because frontier models know the internet. Not your company, not its unwritten rules, past decisions and customer-specific exceptions.
The piece “Intelligence Ownership Becomes the Business Case” takes this further. The next strategic question isn't just how knowledge gets to the AI, but where the resulting intelligence lives and who owns it. As long as context is shipped off to someone else's model with every transaction, the company remains a tenant of its own intelligence.
The next level of maturity is likely to be a hybrid architecture:
- Frontier models handle rare, open-ended and creative tasks.
- Knowledge graphs and RAG supply current, verifiable facts.
- Small specialist models learn frequent, clearly measurable decisions.
- Agents orchestrate the components along the process.
- Humans control exceptions and consequential decisions.
The lasting advantage then lies less in the name of the base model. Models are becoming more interchangeable. What isn't interchangeable: your own experts' labelled judgements, taxonomies, exception rules and feedback loops. In short, the institutional memory.
So the question after “build or buy?” may well be:
Which intelligence can we rent, and which should we own?
Now what?
Above all, the study makes one thing clear: the first phase of enterprise AI is winding down. Access to tools is no longer the problem; 89 percent already use AI somewhere. The second phase starts with less comfortable questions:
- Which complete business process is supposed to change?
- Which steps get eliminated, not just accelerated?
- How is individual time saved translated into revenue, margin or better service?
- Which institutional knowledge does the AI need on a permanent basis?
- Which decisions may an agent make, and who checks the result?
- What does a correctly completed outcome cost, not just a token?
- Where will tomorrow's experts come from if AI takes over the beginners' work today?
Anyone who answers these questions with nothing but the name of their favourite model doesn't have an AI strategy. They have a subscription.
So perhaps a future McKinsey study won't just ask about knowledge graphs, agents and token costs, but about proprietary specialist models, training data, expert labels and a strategy for Intelligence Ownership. When that day comes, readers of it's promp.td will already have internalised the topic. 😉
Because in 2025, I launched the leitmotif:
It's always the Process, Stupid!
In 2026 came the addendum:
But it's also the Data. And in the end, it's about who owns your Intelligence.
Same direction. Fewer slides. Published a little earlier.
And, well … for free. 😉
