Your
Trusted
Get Fully AWS Funded
Cloud Migration
Cloud Migration
AI use can mean that an employee or team has access to a tool and uses it for a task. Meaningful adoption involves a workflow that has been deliberately changed, an outcome that can be measured and an owner who is accountable for how the capability performs.
Surveys use different units of analysis. Some ask leaders whether AI is used anywhere in the organisation. Others measure firms that have deployed AI in a business function, or examine whether AI is integrated into a product or operating process. The numbers can all be accurate while describing different stages of maturity.
Enterprise-wide scale is a decision about operating fit. A company may achieve more by improving one high-friction workflow than by licensing a general tool for every employee. Scope should follow a real operating problem, the quality of available data and the organisation’s capacity to own the result.
Start with the workflow: throughput, quality, rework, customer or employee experience, exceptions, operating cost and risk. Active seats and prompt counts provide context; they rarely show that the business is performing differently.
Once AI connects to business data, tools or production systems, infrastructure decisions affect the outcome directly. Identity, access, observability, reliability and the ability to stop or roll back an action are part of making a capability usable in real operations. More Das Meta cloud infrastructure articles explore these operating foundations.
“AI is receiving more capital, attention and technical momentum than any business technology before it. Yet inside many companies, the reality still looks surprisingly familiar: a few power users move faster, most people experiment occasionally, and the work itself stays largely unchanged.”
Global private investment in AI reached $344.7 billion in 2025, up 127.5% year on year. Generative AI captured $170.9 billion of that total, after growing by more than 200% from the previous year. Frontier models gained 30 percentage points on a benchmark designed to stay difficult for both machines and human experts. On structured computer-use tasks, agent accuracy rose to 66.3%.
The acceleration is visible in funding rounds, product launches and benchmark charts. It is visible in the volume of AI features that now arrive inside tools teams already use. There is a real technology story here, and it is moving quickly.
Then a company buys licences, runs a training session and calls the rollout a success. A usage dashboard goes up. A handful of people become exceptionally good at the tool. A much larger group occasionally asks it to rewrite an email, or never returns after the first awkward output. The workflow keeps moving at roughly the same pace.
“AI adoption has become a catch-all term. It can describe a licence, a prompt, a pilot, a workflow, or a financial result. Treating those as the same thing makes almost every dashboard look healthier than the work underneath it.”
A recent essay makes this point through a deliberately sharp claim: you may already be among the top 1% of AI users in your company. It’s a useful provocation. The 1% comes from practitioner observation, and the underlying question is stronger than any market-wide estimate: access to AI and the ability to get durable value from it are very different things.
The headline data can sound almost conclusive. 88% of organisations report regular AI use in at least one business function. Yet only about one-third say they have started to scale their AI programmes across the enterprise; 39% attribute any level of enterprise-wide EBIT impact to AI, and the group reporting both significant value and at least 5% EBIT impact accounts for around 6% of respondents.
Those figures describe reported use, which makes the definition central to their interpretation. A survey respondent can truthfully report regular AI use when a team uses a tool in marketing, support or IT. That answer tells us little about whether the underlying workflow was redesigned, whether the result is reliable, or whether someone can trace the commercial effect.
Official business data creates a useful second view. 20.0% of EU enterprises with ten or more employees used AI technologies in 2025. A nationally representative US survey found 18% of firms using AI in a business function in late 2025 and early 2026. Broader executive surveys, official business data and disclosure-led research answer different questions, apply different thresholds and sample different populations.
The word “adoption” is doing too much work. It can cover an employee with a chatbot tab open, a team that uses AI for one task, a product feature in production, or a company whose operating model now depends on an AI-supported process. These are different stages, with different costs, risks and evidence of value.
The strongest clue sits in the middle of the adoption journey. Among US firms that use AI, 57% use it in three or fewer business functions, while 65% limit it to three or fewer worker tasks. Most use is still augmentation: helping someone write, search, summarise or analyse faster inside the work they already do.
That can be valuable. It can also be where an AI programme stops.
A recent preprint that reads S&P 500 disclosures for evidence of deeper integration estimates that 11% of firms had AI deeply integrated into business processes in 2025. A further 10% used it in producing goods or delivering services. It introduces a new measurement approach to AI integration through public disclosures, within a narrow S&P 500 sample. Its useful contribution is the distinction between personal or team usage and systems that carry part of the business process.
This is where AI readiness becomes practical. Does the team know which workflow should change first? Is there a trusted source of context? Can a human correct the outcome before the error travels? Is there an owner who can decide what happens when quality drops, a data source changes or a model fails?
These questions can look operationally dull next to a model demo. They are often where the difference between an impressive pilot and a repeatable capability is decided.
The adoption gap needs a counterweight: AI can create measurable value.
In a field study involving more than 5,000 customer-support agents, access to an AI assistant increased issues resolved per hour by 14% on average. The gain reached 34% for novice and lower-skilled workers, while experienced workers saw little average improvement. The result is striking precisely because the setting was specific: a known task, a defined outcome, a steady stream of comparable work and people who could accept, adapt or reject the model’s suggestion.
That is a better way to read the investment boom. Better models expand what could be useful, while workflow redesign, automation boundaries, exception ownership and commercial measurement remain organisation-level decisions.
Prompt training has a place, especially when it reveals people who already understand how to experiment responsibly. Workflow design, data boundaries, review points and a measure beyond active seats remain essential.
A support team might measure resolution time, repeat contacts and quality assurance. A finance team might measure exception rate, time to close and the share of work that still requires rework. A product team might measure task completion, escalation rate and whether users return to the feature. The appropriate metric changes with the workflow. Every workflow still needs a measure that reflects its own outcome.
The companies that get beyond shallow usage tend to make the work visible before they make the AI bigger. They choose a high-friction workflow, define what a good outcome looks like, connect only the relevant context and decide where human judgment remains part of the path.
For an AI agent, that also means deciding what it can read, which tools it can call, where its permissions end and what evidence remains after an action. An earlier discussion about agent identity explores why those controls become urgent once an agent moves from answering questions to reaching live systems.
This is the point where an AI capability layer earns its place. The model sits within a wider operating system. The layer around it connects the right data and tools, keeps authority bounded, makes behaviour observable and gives a team a way to intervene when real-world conditions diverge from the demo.
A useful board conversation asks: which named workflows now perform differently; what has improved; what new risk or operating cost appeared; and who is accountable for the result?
That conversation is harder to conduct. It is also much closer to the truth.