Access to a capable AI system and the ability to produce capable work are separate conditions. The interface can make them look identical. Their operating requirements are different.
Four layers shape the distance between a model request and a result that can survive contact with the world: model capability, context readiness, workflow authority, and outcome quality. Model access is the easiest layer to purchase. The other three depend on decisions an organization has to make for itself.
Layer 1: Model capability
Frontier models can reason across documents, write code, synthesize research, generate media, and operate tools. Model choice still matters. Task fit, privacy, cost, latency, reliability, and provider policy can all change the result.
Capability sets a ceiling. It does not determine what source material enters the workflow, which policy version counts, who may delegate an action, or who accepts responsibility for the final output. Those conditions live outside the model.
Layer 2: Context readiness
Documents, records, policies, notes, and source files become operational infrastructure when an AI system uses them. Conflicting versions and stale instructions survive the transition into an AI workflow. They may become harder to notice because the model can express the conflict in smooth prose.
OpenAI’s May 2026 B2B Signals analysis offers a useful adoption signal. Firms at the 95th percentile of AI use generated 3.5 times as many tokens per worker as the median firm, up from 2 times in April 2025. Message volume explained 36% of the gap. OpenAI associated the remainder with deeper use, richer context, and more substantive outputs. Frontier firms also sent 16 times as many Codex messages per worker as typical firms.
The evidence has boundaries. OpenAI describes tokens as an imperfect proxy for business value. The analysis is based on de-identified, aggregated enterprise usage data and does not establish causation. A vendor usage study also reflects its own products and customers.
The safe inference is narrower: advanced AI use increasingly involves richer context and more delegated work. That makes the condition of the surrounding material more consequential. Calling content readiness a bottleneck is our synthesis from that shift.
Layer 3: Workflow authority
Anthropic’s June 2026 Economic Index report found that 93% of sampled chat and Cowork conversations produced an identifiable artifact. Explanations accounted for 17%, documents and reports for 15%, and guidance for 11%. In work conversations, documents and reports led at 20%.
The report also compared product surfaces. A median chat or Cowork session producing a blog or article involved 13 rounds of human-AI exchange. A median blog-producing Claude Code session contained one human prompt. Anthropic interprets this as a difference in autonomy across product surfaces. The comparison says nothing by itself about output quality.
Fewer handoffs make authority design more important. An organization needs to know which outputs are suggestions, which may enter a draft workflow, which require specialist review, and which actions remain blocked. The accountable reviewer should be named before an error appears.
Authority is a workflow property. The model can generate a polished artifact without knowing whether it has permission to settle a factual dispute, interpret a policy, expose sensitive material, or send the result.
Layer 4: Outcome quality
Outcome quality records the interaction among the first three layers. A strong model working from current sources inside a bounded review process has better starting conditions. A strong model working from stale sources with unclear authority can produce confident work that moves quickly in the wrong direction.
Anthropic’s April 2026 study of 80,508 Claude users adds a human consequence. Among respondents who explicitly discussed productivity effects, 48% emphasized expanded scope and 40% emphasized speed. People reporting the largest speedups also expressed greater concern about job displacement.
The survey drew from personal Claude accounts and voluntary responses. Many attributes were inferred from open-ended answers with Claude-powered classifiers. It is not representative of all workers. Even with those limits, “expanded scope” matters: people are attempting work they previously could not do. That increases the importance of knowing where source material ends, where inference begins, and when outside expertise is required.
Five questions before deeper delegation
- Source: Which file, policy, or record is authoritative today, and who last verified it?
- Status: Can the system distinguish current material from drafts, archives, superseded versions, and unverified notes?
- Permission: Which decisions may the AI support, and which actions require a named human approval?
- Trace: Can a reviewer recover the sources, prompts, edits, and approval state behind the result?
- Repair: Who owns correction, rollback, notification, and learning when the output fails?
These questions turn “AI readiness” into something inspectable. The model will keep improving. The surrounding operating system still belongs to the people using it.
Public sources and posture
- OpenAI, “B2B Signals,” May 6, 2026. Vendor-authored, aggregated usage research; observational and product-specific.
- Anthropic, “Economic Index report: Cadences,” June 26, 2026. Anthropic product telemetry and survey work; platform and sampling limits apply.
- Anthropic, “What 81,000 people told us about the economics of AI,” April 22, 2026. Voluntary personal-account survey with classifier-based inference and disclosed representativeness limits.