Everyone has AI. Few can turn it into work.

Measurement discipline

AI adoption numbers hide the real problem

AI adoption is often discussed as a single curve. That framing hides the real problem. Different sources measure different units: organizations, firms, workers, employment-weighted firms, product usage, and tasks.

McKinsey reports that 88% of survey respondents say their organizations use AI regularly in at least one business function. The Census Bureau reports a much lower past-two-week current-use range for BTOS businesses. The Federal Reserve shows why both can be true: adoption looks different when the unit is the firm, the worker, or the worker inside an AI-adopting firm.

Operating absorption

Using AI is not the same as operationalizing it

The most useful McKinsey finding is not just that 88% of respondents report regular AI use. It is the gap between that figure and the smaller set of organizations reporting scaled programs, EBIT impact, and high-performer status. McKinsey

Value capture

Where AI value dies

Value rarely dies at the moment of generation. It usually dies afterward. A pilot can produce a strong answer and still fail if the company cannot place that answer inside the way the business actually runs.

The sources point to five recurring failure modes: value is not measured, governance trails deployment, data context is fragmented, agents are given more ambition than controls, and employees use AI faster than the organization redesigns the work around them. BCG IBM

This is the downside of the current cycle. Companies can spend on licenses, demos, and pilots while still leaving managers with unowned outputs, unclear review standards, uncertain data quality, and no durable way to learn from completed work. The organization may look active without becoming more capable.

IBM's 2025 CEO study points to the same issue from the executive side: surveyed CEOs reported that only 25% of AI initiatives had delivered expected ROI over the previous few years, only 16% had scaled enterprise-wide, and 50% said rapid investment left disconnected technology. Deloitte's 2025 ROI survey found rising investment, but only around one in five surveyed organizations qualified as AI ROI Leaders. These are different measures. They point to the same operating problem: spend is easier to approve than value is to measure and absorb. IBM Deloitte

Task diffusion

AI changes tasks before it changes jobs

Anthropic's Economic Index is valuable because it looks below the org chart. The first report found that roughly 36% of occupations had AI use in at least a quarter of associated tasks, with observed usage leaning toward augmentation at 57% versus automation at 43%. Anthropic

The March 2026 update showed usage becoming less concentrated inside Anthropic's Claude data. The top 10 tasks fell from 24% of Claude.ai traffic in November 2025 to 19% in February 2026, and about 49% of jobs had at least a quarter of tasks observed in Claude usage. Anthropic Economic Index

This is how enterprise AI should be evaluated. The better question is not whether a job is "automated." It is which tasks have enough context, repetition, evidence, and review discipline to become repeatable AI-supported work.

Productivity evidence

AI works when the workflow is ready

Productivity is a workflow property, not a model property. The evidence is not a single number. It is a set of boundary conditions. A QJE study of 5,172 customer-support agents found that access to a generative AI assistant increased issues resolved per hour by 15% on average, with larger gains for less experienced and lower-skilled workers. QJE

METR found the opposite in a different setting. In a randomized controlled trial with 16 experienced open-source developers working on familiar repositories, access to early 2025 AI tools made completion time 19% longer. Developers had expected a 24% speedup. METR now labels the result as historical evidence because model capability has moved since the study window. METR

The variance is the lesson. AI helps when the task is bounded, the source material is available, the output can be reviewed, and the feedback loop is short. It can slow work down when the user has to reconstruct context, verify too much, or absorb errors that the workflow did not anticipate. The model matters, but the surrounding work system determines whether capability turns into productivity.

Agent reality

Agents expose the operating-model gap

Agents matter because they expose what chat can hide. A person can use a model informally and absorb the risk alone. A system that acts on business work needs authorization, source discipline, policy boundaries, exception handling, and a reviewer who can trust the path taken.

The data does not support a claim that agents are already running the enterprise. It supports a more precise claim: agent activity is real, production deployment is early, and weak operating controls are a likely failure mode.

This is why agents are not mainly a feature question. They are an operating-model question. If software can act, the organization must define the allowed action, the evidence threshold, the reviewer, the exception path, and the feedback loop.

Operating-model effect

The bottleneck moves from output to review

Faster work is the first-order effect. More work entering the system is the second-order effect. The third-order effect is that review, approval, governance, and quality control become the bottleneck. The fourth-order effect is a new operating model.