AI capability changes gradually
Management decisions do not

A calm curve can still end in a sharp decision

Imagine a support function that has tested AI for a year. Each model release makes drafting, classification and document review slightly better. None of the improvements is decisive. The pilot remains a pilot, the team remains the same size and the workflow retains its old approvals.

Then several things meet at once. The software contract comes up for renewal. Two vacancies are open. The latest system passes the organisation’s quality tests. A connector removes the manual hand-off that made the earlier pilot uneconomic. A new executive wants the process redesigned rather than merely assisted.

The capability did not suddenly appear that morning. The management decision did.

This is the timing problem hidden inside many AI forecasts. Technical progress, organisational adoption and employment consequences are often discussed as if they run on one curve. They do not. Capability can improve continuously. Organisations act through budgets, contracts, controls, vacancies, restructurings and accountable decisions. Employment consequences then arrive unevenly across particular teams, locations and career stages.

Capability curves can be continuous. Management decisions are thresholds.

The technical signal is a rising tide

MIT FutureTech’s preprint Crashing Waves vs. Rising Tides offers the technical trigger for this distinction. The researchers tested a pool of 41 large language models across 6,648 text-based or partly text-based labour-market tasks derived from the United States Occupational Information Network. Their analysis covered 12,078 task instances and 60,845 model-response evaluations from people with relevant work experience. ¹

The workers judged whether model outputs would be acceptable without edits. Across all models and tasks in the study, 60 per cent of responses met at least the minimum-sufficient threshold, 49.7 per cent met or exceeded average-worker quality and 27 per cent were rated superior. For frontier models, success on tasks that people reported would take about one and a half hours rose from roughly 60 per cent in the second quarter of 2024 to above 70 per cent by the third quarter of 2025. ¹

The shape matters more here than the headline level. Performance improved broadly across task durations rather than moving from near-total failure to near-total success across a narrow band. That is the rising-tide finding.

It is also easy to over-read. Tasks were selected after GPT-4 classified them as offering at least 10 per cent potential time savings; that was a screening judgement, not observed time saved in workplaces. The self-contained prompts did not capture work requiring external files or software navigation, or repeated multi-turn interaction with other people. The acceptance measure was a worker’s judgement of what a hypothetical manager would accept, not observed deployment inside a live organisation.

The authors explicitly separate capability from adoption. Real implementation can be costly, information may be difficult to assemble and regulatory or operational constraints may make an apparently feasible task unattractive to automate. Their 2030 extrapolations are presented as a fast, upper-bound scenario if recent progress persists, not as a forecast. The paper is a preprint, and it discloses funding from Open Philanthropy and an unnamed technology company. ¹

Used with those limits, the research supports one claim: capability across a large set of relevant text tasks can improve broadly and relatively smoothly. It does not tell a board when to change a process, what a safe deployment costs or how many people a firm will employ.

Capability, adoption and employment run on different clocks

Technical capability is what a system can do under defined conditions. It can be measured through task evaluations, tests and observed improvements between model versions. Capability may rise even while an organisation does nothing.

Organisational adoption is the decision to put that capability into an actual workflow, with data access, integration, ownership, controls, review and funding. It is usually lumpy because those complementary conditions are lumpy.

Employment consequences follow the way work, demand and capacity are reorganised. A task improvement can complement a worker, remove contractor spend, reduce hiring, support growth, change a role or contribute to redundancies. Those outcomes are not interchangeable.

The distinction prevents two common mistakes. The optimistic mistake is to treat stable aggregate employment as proof that capability is commercially unimportant. The alarmist mistake is to treat a capability score as if the corresponding jobs have already disappeared. Both skip the organisation in the middle.

Firm adoption is uneven, narrow and two-directional

A revised United States Census Bureau working paper provides a useful view of that middle layer. Using a nationally representative supplement to the Business Trends and Outlook Survey, fielded from 17 November 2025 to 8 February 2026, it estimates that 17.9 per cent of firms used AI in at least one business function. The employment-weighted share was 31.2 per cent, reflecting higher use among large firms. ²

Even among users, breadth was limited. Among functional adopters, 57 per cent used AI in one to three business functions. Among firms reporting worker generative-AI use, nearly 65 per cent limited it to one to three of the nine task types surveyed. Among AI-using firms reporting any task effect, 66 per cent reported augmentation alone. Over the previous six months, 2.0 per cent of AI-using firms reported an AI-attributed net employment decrease. The survey did not measure the magnitude of that change or gross hiring and separation flows. ²

The pattern was not a single top-down rollout. Some workers used AI for tasks without formal firm-level adoption, while some firms reported adoption without worker task use. That two-directional diffusion helps explain why an organisation can contain advanced pockets, prohibited pockets and untouched processes at the same time.

The paper also found that broader functional integration and operational investment were positively associated with reported employment decreases, while worker task use alone was not significantly linked to headcount reduction after those factors were considered. This is correlation in survey data, not proof that broader integration caused the reductions. AI-attributed outcomes also remain respondents’ assessments. The result nevertheless supports a practical distinction: occasional worker use and an operating-model decision are different events. ²

The threshold sits in the organisation

Why does one firm act while another waits? Because the relevant boundary is not the model’s benchmark score. It is the point at which capability becomes worthwhile and governable inside a particular workflow.

A quality threshold may be crossed when representative local tests meet an agreed error rate and reviewers no longer have to reconstruct the answer. An economic threshold may be crossed when the total cost of integration, monitoring and exceptions falls below the cost of the current process. An integration threshold may be crossed when the system can reach the right data and hand work to the next stage without manual repair.

A governance threshold may be crossed when accountable owners, audit evidence, escalation paths and fallbacks are in place. A timing threshold may be crossed when a contract expires, demand falls, a vacancy appears or a new leader gains authority to redesign the work. None of these boundaries is universal. Two firms using the same model can rationally make different choices because their costs, risks, systems and strategic options differ.

This is why gradual capability can still produce a crashing organisational wave. The wave is not necessarily a technical discontinuity. It can be a delayed decision that releases several accumulated improvements at once.

An illustrative vignette

Consider a hypothetical finance-operations team that reconciles supplier records and prepares explanations for exceptions. For twelve months, AI assistance improves the first draft but still requires heavy checking. The organisation learns from the pilot but does not change roles.

At the next contract renewal, three conditions align. The model now passes a locally designed quality test. A new connector provides controlled access to the relevant records. Two experienced employees are due to leave and the finance director is asked to hold costs flat.

Management decides not to refill both vacancies. It redesigns the workflow so the system prepares matches and explanations, an analyst reviews exceptions, and a senior owner signs off higher-risk cases. The visible workforce change happens in one budget cycle, even though the enabling capability accumulated across several model releases.

This vignette is an editorial illustration, not a documented case or a claim that vacancies should be removed. A different organisation might cross the same technical threshold and use the capacity for growth, service quality or backlog reduction. The point is the timing: the organisational consequence is determined by the decision made at the boundary.

The labour evidence is uneven, not absent

Stanford Digital Economy Lab’s August 2026 revision of Canaries in the Coal Mine uses a balanced ADP-firm panel covering roughly 3.5 to 5 million full-time United States employees per month through June 2026. The authors find no evidence of widespread, economy-wide displacement in that analysis sample. They also report that employment among people aged 22 to 25 in the two most AI-exposed occupational quintiles was 19 per cent below where it would have been had it kept pace with the three least-exposed quintiles, with the divergence appearing mainly through reduced hiring rather than separations. ³

The authors do not present that as a causal estimate. The sample is large but not nationally representative; larger firms and some sectors are overrepresented. The pattern weakens when education is controlled for, some divergent trends predate generative AI, and the result is stronger in the payroll sample than in national survey benchmarks. They call it an early descriptive indicator. It is therefore evidence to monitor, not permission to attribute every missing vacancy to AI. ³

The International Labour Organization adds another boundary. Its 2025 task-level exposure index estimates that 24 per cent—roughly one in four—of workers globally are in an occupation with some generative-AI exposure, while 3.3 per cent of employment sits in the highest exposure category. The index is a static exposure snapshot at the start of 2025 and an upper-bound affected share under full implementation, not a realised employment effect or forecast. Because occupations combine tasks that still require human input, the ILO judges job transformation more likely than complete automation in the broad case. ⁴

Together these sources suggest an uneven sequence. Capability can advance across many tasks. Firms integrate it at different speeds and depths. Workforce effects may show first in hiring, contractors, workload or role design before they appear as an economy-wide displacement event.

The AI Decision Threshold Map

Use this map for one workflow at a time. Its purpose is not to manufacture a universal readiness score. It is to expose the conditions that would change a real management decision and the evidence required before that decision is taken.

On a narrow screen, each row is shown as a separate card.

For each boundary, keep a short threshold register: current evidence, accountable owner, next review date, action if crossed and the veto or pause condition. If error patterns, costs or control failures move in the wrong direction, the organisation should know when to stop or reverse.

What to ask before the boundary is crossed

• Which capability has actually improved, on whose evidence and under what conditions?
• What specific threshold would cause us to change this workflow, contract or role design?
• Have we measured the whole system, including integration, checking, failures and fallback?
• Who has authority to say that the quality and governance thresholds have been met?
• Which calendar event could turn a technical improvement into a rapid organisational decision?
• If the threshold is crossed, are we using the capacity for growth, quality, resilience, cost reduction or headcount change?
• What knowledge, learning path and human control must remain after the redesign?

Watch the boundaries, not the averages

Leaders will continue to see two apparently conflicting stories. One will show broad, fast improvement in AI capability. The other will show organisations and labour markets changing more slowly, partially and unevenly.

There is no contradiction. Capability is an input. Organisational change requires a decision. Employment effects depend on what the organisation does with the capability, what demand does next and which form of capacity is adjusted.

That makes the waiting period dangerous if it is mistaken for stability. A firm may look unchanged while several thresholds move closer: quality rises, integration improves, prices fall, contracts mature and leadership attention shifts. When enough conditions align, the decision can be rapid.

The practical response is neither panic nor passivity. Identify the boundaries now. Decide what evidence counts. Model the action on both sides. Protect the knowledge and control the future workflow will still require.

A rising technical tide does not dictate the decision. It changes when the decision becomes available.

Evidence note

This article separates technical capability, organisational adoption and employment consequences. The MIT source is a July 2026 preprint and is used for the shape of capability improvement, not as proof of real-world adoption or job loss. Census and Stanford findings are observational and are not treated as causal estimates. The finance-operations vignette and AI Decision Threshold Map are Andrew’s Intel editorial devices.