Imagine two companies answering the same survey. One has a team trying a chatbot on internal documents. The other uses machine learning to screen transactions throughout the working day. Both can answer yes to "Does your organization use AI?" You would want considerably more detail before giving either company your technology budget.

That is the difficulty with global AI adoption numbers. They tell us something about reach, but much less about how deeply the technology has entered everyday work. To understand who is adopting AI, we need to distinguish trying a tool, deploying it, and relying on it. Industry and geography still matter. So does knowing exactly what the survey counted.

What the headline measures

In McKinsey's early-2024 survey, 72% of respondents reported AI adoption in at least one business function. The survey ran from February 22 to March 5, 2024, with 1,363 participants across regions, industries, roles, and company sizes. Results were weighted by each respondent's country's contribution to global GDP. The report also put regular generative AI use in at least one function at 65%.

Those measures describe breadth of use. They do not mean 72% of all businesses worldwide had mature AI operations, or that every department in a respondent's company used AI. A single function can qualify an organization for the headline.

The financial evidence measures something else again. Stanford's 2024 AI Index, chapter 4, reported $189.2 billion in global corporate AI investment during 2023. That included mergers and acquisitions, minority stakes, private investment, and public offerings. It was down about 20% from 2022. Buying an AI company and putting a model into daily service are different events; the investment total cannot confirm the adoption percentage.

I would treat these as separate entries on a dashboard: how much money is moving, how many respondents report use, and how much work actually depends on the technology. Combining them into one story of universal progress loses information that a decision-maker needs.

The distance between exploration and deployment

A useful comparison comes from a single survey with explicit stages. IBM's Global AI Adoption Index 2023 release reported that 42% of surveyed enterprises with more than 1,000 employees had actively deployed AI; another 40% were exploring or experimenting without deployment. The full November 2023 survey covered 8,584 IT professionals across multiple countries, including smaller employers; the headline figures concern the large-enterprise subset.

That is evidence of substantial activity at different stages. "Actively deployed" still doesn't establish company-wide coverage, dependable performance, or a positive return. Nor can we subtract IBM's deployment rate from McKinsey's usage rate to calculate how many organizations are stuck in pilots. The samples, dates, and questions differ.

The saved chart below shows why this matters. Its 35% production, 37% pilot, and 28% non-use slices have no verified common survey basis. In particular, the 37% slice comes from subtracting an unverified 35% production figure from the 72% usage headline. Read this preserved graphic as an example of an unsupported partition, not as a measurement of organizations or a McKinsey result.

Unverified partition, not measured adoption rates. The saved 35% production / 37% pilot / 28% non-use chart lacks a common survey basis. Subtracting an unverified production rate from McKinsey's 72% usage figure does not establish a pilot population. The percentages and category labels must not be used as market evidence.

A pie chart makes categories look settled. Before using one to set strategy, ask whether every slice describes the same population at the same time, whether the categories exclude one another, and whether they cover everyone. An organization can have a production fraud model and an experimental chatbot simultaneously. A survey needs a rule for classifying that organization before its answers become slices.

Different industries have different finishing lines

Finance offers tasks with visible consequences: detecting fraud, evaluating credit risk, extracting information from loan documents, or generating trading signals. A useful way to assess adoption is to follow one decision. For a fraud model, a score must arrive while the transaction can still be stopped. Someone must handle disputed decisions. Missed fraud and legitimate transactions blocked by mistake both have costs.

IBM's 2023 survey reported active deployment at about half of surveyed financial-services respondents' organizations. That supports substantial use in that sample. It does not tell us how much of each firm's work uses AI, or whether finance's lead is widening. The same IBM release identifies fraud detection and document processing among enterprise AI applications. A list of applications tells us where to look; operational evidence tells us how well they work.

Healthcare has several different adoption stories inside it. The FDA's AI-enabled device list records marketing authorizations, including imaging applications. Authorization is evidence that a device met applicable premarket requirements. It is not a count of hospitals using it. Likewise, AI in drug discovery can help with particular stages, but a faster molecular-screening task does not establish that the whole development process has shrunk from years to months. The FDA's discussion of AI in drug development treats uses across that lifecycle separately.

Clinical documentation offers a more bounded example. A Stanford Health Care pilot study of Nuance's DAX Copilot reported a median reduction of 0.57 minutes per note compared with the period before the pilot. That is a specific workflow result, with a before-and-after design rather than proof that every deployment will save the same time. Scheduling, claims processing, drafting a note, and supporting a diagnosis also deserve separate assessments. Their errors have different consequences.

Software development illustrates another distinction: individual tool use versus organizational maturity. In Stack Overflow's 2024 survey, 62% of respondents said they were using AI tools in their development process, while 76% were using or planning to use them. That supports widespread use among respondents. Support routing, sales assistance, and product recommendations pose further workflow questions that a coding-tool survey cannot answer.

Manufacturing makes the transition especially tangible. Predictive maintenance uses equipment signals to anticipate problems; visual inspection helps identify defects; supply-chain applications help plan material flows. NIST's manufacturing work addresses how to measure and manage industrial AI. Consider a hypothetical maintenance model that predicts bearing failure. Its value depends on the warning arriving early enough, a technician being available, and the replacement part being in stock. An accurate prediction can still fail to prevent downtime if nobody can act on it.

These examples give "production" a useful meaning: a model participates in a real workflow with a responsible owner and a way to handle mistakes. Even then, the next question is whether the workflow improved.

Geography needs the same care

The US investment lead is clear in Stanford's 2024 AI Index overview: US private AI investment reached $67.2 billion in 2023, roughly 8.7 times China's total and 17.8 times the UK's. Those are private-investment comparisons. They don't rank how extensively ordinary businesses in each country use AI, and the UK figure is not an EU total.

For a different view, Eurostat reported AI use by 13.5% of EU enterprises with at least ten employees in its covered sectors in 2024, up from 8% in 2023. Placing that beside McKinsey's 72% and calling the difference a European adoption deficit would repeat the same mistake: different populations and survey methods can produce very different numbers.

Within IBM's survey, China had a reported active-use rate of 50%, compared with 26% in France. Those dated survey results describe respondents' organizations. They cannot establish that state investment, data access, or regulation caused the difference, and they say nothing directly about which country builds the best frontier model.

Policy affects deployment requirements. The EU AI Act uses a risk-based approach, so the intended use matters. But a claim that regulation explains a country's adoption rate needs evidence connecting the two. For a business choosing where to deploy, local data access, available expertise, and the rules for its particular use case are more useful planning inputs than a single national ranking.

What has to change after the demo

The production problem becomes easier to understand if we follow a hypothetical loan-document extractor. In a demonstration, it reads a clean application and returns the right fields. In daily use, it encounters a scanned copy, a revised form, and an applicant whose name differs between documents. The database may already contain a record under another identifier. Extracting text correctly is only part of completing the job.

That is why data quality and integration deserve attention early. The team needs to know which record is authoritative, how to detect duplicates, and where uncertain results go for review. It also needs response-time limits, monitoring, fallback behavior, and permission to access the right records. These are concrete requirements behind the phrase "production readiness."

Skills are one constraint among several. IBM respondents cited limited AI expertise, data complexity, ethics, integration, and cost as barriers; the ordering is specific to that survey and question. Hiring another model specialist will not resolve an unclear business objective or contradictory customer records.

For this hypothetical document workflow, I would define success before expanding the pilot: correctly processed applications, review time, exceptions missed, and total operating cost compared with the existing process. The people handling applications should help decide what counts as an acceptable mistake. Model monitoring and retraining, often grouped under MLOps, then support a defined service instead of becoming infrastructure in search of a purpose.

A narrow internal assistant and a system involved in credit decisions need different delivery plans. Scope, dependencies, validation, and the consequences of error determine the work.

Measure the work that changed

The evidence supports broad AI use and a meaningful distinction between exploration and deployment. It also shows why an adoption percentage alone is a poor target. A company can raise its count of AI users without improving the work those users do.

I would track the share of relevant work handled, the quality of the result, the human effort still required, and the cost of keeping the service running. For document processing, that means following applications through completion, including the ones sent back for correction. For maintenance, it means checking whether warnings led to useful interventions. Those measures let a team decide whether to expand, repair, or stop a deployment.

Before setting the next adoption target, ask: which task has become dependable enough that people choose to rely on it? Answer that with operational evidence, and "we use AI" starts to tell you something useful about the organization.