# AI productivity is becoming a queue at the human brain

Anthropic’s agents are working six times as many hours as their human handlers, yet spend roughly 80% of interactive-session time waiting for people. Output is multiplying faster than organisations can absorb, judge and integrate it. CHROs and CIOs need to budget human attention alongside tokens.

**AyEye — Workforce Management · 2026-10-17**

AI can produce more work than an organisation can use.

That sounds like abundance. Inside a company, it can become congestion. Someone still has to understand the output, decide what matters, connect it to other work, accept responsibility and act.

New evidence from Anthropic makes the imbalance unusually visible. By August, employee-directed agents were working six times as many hours as their human handlers, yet agents spent roughly four-fifths of their interactive-session time waiting for human input. Output expanded; so did the human work around it.

The next productivity constraint may not be model intelligence. It may be the finite rate at which an institution can turn machine activity into accountable judgement.

## The executive brief

- Anthropic has published an unusually detailed internal study of work at a frontier-AI company. It combines usage and output measures with employment data, surveys, communications, calendars and 17 interviews. By August, employee-directed agents recorded six times the hours of their human handlers.

- The same agents spent about 80% of interactive-session time waiting for people. Anthropic says even non-technical employees would need more than half a working week merely to read all the output Claude produced for them.

- Coordination did not disappear. “Meta-work” rose from 10% of interactive sessions in autumn 2025 to 22% in spring 2026, while median meeting time increased 20% year on year.

- Junior technical work is thinning inside one fast-growing firm. Junior roles fell from 21% to 14% of technical staff since early 2025, and from 27% to 9% of technical hires. This is one exceptional company, not a labour-market estimate.

- Enterprise platforms are making machine consumption visible before human absorption. ServiceNow has announced a signed “passport” for agent scope and expanded cost-to-value controls. The missing companion is an attention budget: the review, coordination and learning capacity required to convert agent output into reliable value.

## AI productivity is becoming a queue at the human brain

ORIGINAL SYNTHESIS · Confidence: medium-high on the mechanism, low on prevalence · Horizon: now to 24 months.

### Signal one: machine work is already outrunning human handling in one frontier firm

REPORTED FACT. Anthropic published an internal employment study on 16 October using company employment records, AI-use measures, output indicators, surveys, internal communications, calendars, hiring data and 17 interviews.

It reports that by August, employee-directed agents accumulated six times as many hours as their human handlers. That does not mean six times as much value or six autonomous digital employees for every person. “Agent hours” count the duration of agent activity initiated by employees, and much of that activity overlaps in time.

The company also reports that agents spent about 80% of interactive-session wall time waiting for human input. Even a non-technical employee would need more than half a working week to read all the output Claude generated for them. [Anthropic, 16 October 2026](https://www.anthropic.com/research/employment-at-anthropic)

ANALYSIS. The important ratio is not agent hours to employee hours. It is produced output to usable human attention.

An agent can draft, search, code and analyse concurrently. A person cannot read six streams with the same depth, connect each to organisational context, test the evidence and remain accountable for the consequences. The machine's capacity is parallel. Human judgement is selective and path-dependent.

Waiting is therefore not necessarily technical inefficiency. It may be the visible edge of an institutional bottleneck: the moment machine abundance meets a finite human capacity to absorb, decide and coordinate.

### Signal two: the work around the work is growing

REPORTED FACT. Anthropic defines “meta-work” as activity that manages, reviews, communicates or coordinates other work. It says meta-work rose from 10% of interactive Claude sessions in autumn 2025 to 22% in spring 2026.

Median employee meeting time increased 20% year on year. Among technical individual contributors it rose 16%. Anthropic also reports a fivefold increase in code commits per engineer over six months and higher output measures in several functions, while warning that adoption is not random and the company was growing explosively.

ANALYSIS. Productivity and coordination can rise together. That is not a contradiction.

When the cost of producing an option falls, the number of options can increase. When execution accelerates, more work reaches a boundary where priorities conflict, dependencies surface or someone must accept a consequence. When drafts become cheap, comparison, selection and integration become expensive.

The familiar productivity question—“How much faster did the task become?”—misses the new load. The relevant question is: “What human work did the faster task create elsewhere?”

### Signal three: early-career work is narrowing while expectations rise

REPORTED FACT. Anthropic says junior employees declined from 21% to 14% of its technical workforce since early 2025, while the junior share of technical hiring fell from 27% to 9%. Technical employment still grew in absolute terms as the company tripled headcount.

Managers interviewed for the study described placing more value on autonomy, judgement and the ability to work at higher abstraction. In a separate survey published the same day, Anthropic staff expected AI to be able to perform a median 75% of their current work hours by 2030, compared with 28% among US workers. Yet 95% of Anthropic respondents also expected company headcount to grow over the next year, with a median forecast of 66% growth. These are expectations, not a company plan or a forecast with established accuracy. [Anthropic, 16 October 2026](https://www.anthropic.com/research/ai-expectations-gap)

ANALYSIS. The evidence does not support “AI eliminates jobs”. It supports a more difficult possibility: organisations can add people while removing the work through which inexperienced people learn to become dependable.

If employees enter at a higher abstraction level, somebody or something must still supply grounded knowledge of customers, code, policy, operations and failure. The attention bottleneck and the capability-formation problem are therefore connected. Experienced people may spend more time judging machine output just as fewer novices receive the object-level practice that creates future judges.

### Signal four: machine identity and spend are becoming governable objects

REPORTED FACT. ServiceNow announced on 15 October an “AI Passport” intended to cryptographically verify an agent's identity and permitted scope at runtime. It also described additions to its control-tower product for discovering agents and measuring their cost and value. The passport is expected in the fourth quarter of 2026; the claims describe a supplier roadmap, not independently verified production outcomes. [ServiceNow, 15 October 2026](https://newsroom.servicenow.com/press-releases/details/2026/ServiceNow-advances-its-operational-harness-for-enterprise-AI-with-AI-Passport-and-expanded-governance-capabilities/default.aspx)

ANALYSIS. Enterprises are learning to meter tokens, runtime scope, identity and machine cost. Human absorption remains largely hidden inside salaries, meetings, interruptions, cognitive switching and rework.

An agent can have a passport and a budget while its human counterpart carries an invisible queue. That is a measurement asymmetry. It makes machine output look scalable and human judgement look like free residual capacity.

### The attention-conversion model

StageManagement questionFailure if ignored

ProductionHow much work can agents generate concurrently?Output volume is mistaken for value.
CompressionWhat can be summarised, sampled or suppressed without hiding material uncertainty?People drown in detail or trust lossy summaries.
ComprehensionWho can understand the evidence and its organisational context?Review becomes a ceremonial click.
DecisionWho has authority to choose, pause or redirect?The queue waits for a person who cannot act.
IntegrationWhat other work, people and commitments does the output change?Local speed creates system-wide coordination debt.
LearningWhich human capability is strengthened or displaced by the hand-off?Today's throughput consumes tomorrow's judgement.

What to notice: the bottleneck can move. Faster generation may shift the constraint into comprehension, decision rights, integration or learning. Optimising the agent alone does not optimise the institution.

ORIGINAL SYNTHESIS. Organisations need to manage attention conversion: the capacity to turn machine activity into understood, authorised, integrated and learnable work.

This is not an argument for humans to read everything. They cannot. Nor is it a call to automate review until no person remains. Recursive machine checking can reduce noise, but it cannot decide which uncertainty is material to the institution or who should carry the consequence.

The design problem is to determine:

- which outputs may flow without review because they are low consequence and reversible;

- which can be sampled against known failure patterns;

- which need source evidence and human judgement before they change a person, system or commitment;

- which should never be generated because the likely value is lower than the absorption cost;

- where protected practice must remain so future human judgement continues to form.

### The unexpected connection

Agent hours + rising meta-work + fewer junior technical roles + machine passports

These developments belong to productivity measurement, organisation design, workforce development and enterprise governance. Together they suggest that the scarce resource in an agentic organisation is not merely compute, data or access. It is the institution's ability to convert abundant machine effort into accountable action without exhausting the people or degrading the capability on which that judgement depends.

PROVOCATION: Stop measuring how much work AI produces. Measure how much judgement the organisation can absorb.

An organisation can report more code, drafts, analyses and completed tasks while becoming slower at choosing, integrating and learning. The output is real. So is the queue it creates.

If leaders reward generation without pricing absorption, people will either review superficially, create informal filters, extend their working day or allow machine-produced work to accumulate as unowned organisational debt.

### What if we are right?

Opportunity. AI could remove not only production work but low-value information handling. Agents could prepare evidence, identify uncertainty, compare alternatives and route only the decisions that genuinely need human attention. Protected human time could move towards judgement, coaching, exception learning and relationship.

Organisational consequence. AI business cases would include an attention budget alongside token, licence and implementation costs. Workforce plans would estimate review demand, decision latency, coordination load and capability formation. Managers would have explicit limits on the number and consequence of machine workstreams they can responsibly absorb.

Likely horizon. A workflow-level measure is possible now. Portfolio-level attention accounting is plausible within 12–24 months, first where agent telemetry can be joined to workflow, decision and calendar data.

### What would prove us wrong?

The thesis weakens if the Anthropic pattern is mainly an artefact of one rapidly growing frontier laboratory; if agent output becomes sufficiently accurate, compressed and self-integrating that human review demand falls faster than generation grows; or if rising meetings and meta-work are explained by organisational growth rather than AI adoption.

It also fails if attention measures become surveillance of individual employees or an excuse to slow harmless automation. Time spent reviewing is not automatically valuable; a poor approval process should be removed, not celebrated as human judgement.

The operational disconfirmation is observable: if agent volume rises while decision latency, review effort, meeting load, rework, unresolved exceptions and capability gaps all fall, human absorption is not the binding constraint in that workflow.

### Optimistic possibility: machines learn to respect attention

The constructive outcome is not a larger human checkpoint at the end of every machine process. It is an architecture that treats attention as a protected resource.

Low-consequence work can move quickly. Material uncertainty can travel with the output rather than being buried beneath it. An authorised person can see why a case needs them, what changed and which decision they alone must make. The system can retain enough object-level practice for people to learn, while agents remove the clerical work around that practice.

Productivity then becomes less about producing everything that can be produced and more about making the right human intervention both rare and decisive.

## Revisited: every autonomous agent creates management work

On [23 September, Workforce argued that every autonomous agent creates management work](https://www.lecxie.com/publications/ayeye-workforce/2026-09-23.html): briefing, routing, monitoring, review and intervention move into ordinary roles even when formal layers disappear.

Anthropic's internal evidence materially strengthens that thesis in one unusually advanced environment. The increase in meta-work and meetings is not proof of causation or prevalence, but it makes the hidden load measurable. The update is sharper: agentic management work accumulates as a queue against finite human attention.

On [13 October, Workforce warned that autonomous HCM can consume capability faster than it creates it](https://www.lecxie.com/publications/ayeye-workforce/2026-10-13.html). Anthropic's junior-role shift is one-company evidence consistent with that risk, while its absolute hiring growth complicates any displacement story. Capability extraction can occur inside a growing workforce.

On [15 October, Workforce proposed a decision spine](https://www.lecxie.com/publications/ayeye-workforce/2026-10-15.html). Today's addition is that the spine needs capacity. Preserving evidence and authority is insufficient if the responsible person has no time to understand either.

## The review agent could become another producer

TENUOUS BUT PLAUSIBLE · Confidence: medium-low · Horizon: 12–36 months.

As machine output grows, organisations will deploy agents to filter, rank and review other agents. That is sensible. It also risks a recursion trap.

SPECULATION. A review agent may reduce the number of items a person sees while increasing the distance between the person and the underlying work. If it generates explanations, risk scores and exception summaries, it creates a new output stream that itself requires calibration and challenge.

The causal chain is plausible:

Agent output expands → review becomes the bottleneck → review is automated → people see fewer primary artefacts → calibration and context weaken → exceptions become harder to recognise.

What to watch: whether reviewers retain access to primary evidence; whether sampling includes cases the review agent considered safe; whether people can detect drift without doing the object-level work; and whether false reassurance falls when a second model agrees with the first.

## Operating-model implication

### Give consequential automation an attention budget

DecisionEvidence requiredPrimary steward

What may flow without review?Consequence, reversibility and observed failure patternsProcess owner and risk
What deserves scarce human attention?Material uncertainty, affected people and decision rightsBusiness owner and qualified reviewer
How much can one role absorb?Queue length, review time, interruption load and decision latencyManager and workforce planning
What must remain visible?Sources, model/action trace, omissions and unresolved dissentTechnology and assurance
Which practice must be protected?Judgement pathway, novice exposure and coaching evidenceCapability owner and HR
When should generation stop?Marginal value versus absorption and remedy costExecutive sponsor

An attention budget is not a cap on useful automation. It is a commitment not to create consequential machine work without a credible path through comprehension, authority and learning.

## Human control watch

Assistance: AI creates options or prepares evidence for a person. Control depends on whether the person can inspect sources and has enough time to exercise judgement rather than merely acknowledge output.

Delegated execution: agents perform bounded work between reviews. Control depends on sampling, exception compression, queue visibility and a reviewer whose mandate matches the consequence.

Autonomous control: systems act continuously across workflows. Control depends on limits that prevent machine production from outrunning decision capacity, plus interruption and remedy that do not rely on an overloaded individual.

Today's shift: human control is not the presence of a person somewhere in the loop. It is the availability of informed human attention at the moment it can still change the outcome.

## Capability-model update

Gaining valueUnder pressure

Attention-conversion designerAgent-output volume used as productivity
Exception-compression stewardEverything routed to a nominal reviewer
Machine-workload plannerHuman review treated as free capacity
Judgement-pathway designerJunior work removed without replacement practice
Primary-evidence samplerReview agents validating review agents
Decision-latency analystTask speed measured outside the whole workflow

## IF I WERE TO DO ONE THING NOW

### Measure one attention path

OUTPUT → READ → CHECK → DECIDE → INTEGRATE

Return to the workflow whose whole decision you named in the previous exercise—or choose one agent-assisted workflow now. For five completed cases this week, time the human work that happened after the agent produced its output: reading, checking, clarifying, coordinating and remedying. Mark the one human intervention that actually changed the decision. Do not create an enterprise dashboard or introduce a new approval. Use the five cases to remove one avoidable review step, improve one piece of evidence or set one boundary on generation. This converts the decision spine into a realistic measure of the attention required to keep it alive.

## Mental-model update

The regenerative enterprise has learned to form capability, preserve whole decisions and connect machine work to responsibility.

Today it recognises attention as conversion capacity.

Machine effort is abundant only when the organisation can turn it into understood, authorised and integrated action. Otherwise abundance becomes a queue, and the person named as “in the loop” becomes its hidden infrastructure.

The emerging North Star is an attention-literate enterprise: fast enough to use machine abundance, selective enough to protect judgement and deliberate enough to keep producing the people who can exercise it.

## Questions for the executive table

- Where is agent output growing faster than the organisation's capacity to understand and use it?

- Which AI business case counts production time saved but omits reading, checking, coordination and remedy?

- Which person is nominally “in the loop” but lacks either the time or authority to change the outcome?

- Where are junior people losing the object-level practice needed to become tomorrow's reviewers?

- Which machine-generated work should not be produced at all because its absorption cost exceeds its value?

Evidence note. Anthropic's employment study concerns one rapidly growing frontier-AI company and combines direct measures with surveys and 17 interviews. It cannot establish causation or prevalence across employers. Its separate expectations survey records beliefs, not company plans or realised outcomes. ServiceNow describes forthcoming and existing products; capability and value claims require independent field evidence. Increased meetings and meta-work may partly reflect organisational growth, measurement choices or selection into AI use.

The concepts attention conversion, attention budget and machine abundance becoming a queue are original AyEye analysis, not claims made by the cited sources.

Canonical: https://www.lecxie.com/publications/ayeye-workforce/2026-10-17.html