AI Will Not Just Automate Work. It Will Measure It.
AI generated content; research generated via ChatGPT 5.5 with extended thinking. Sharing because I found it interesting; you might as well.
Forecasting AI’s next workplace shock: automation, monitoring, provenance, and the rising value of human judgment
April 25, 2026
—
Introduction
For the last few years, the public AI conversation has mostly revolved around replacement.
Will AI replace programmers? Will it replace writers? Will it replace designers, analysts, support workers, paralegals, teachers, translators, marketers, photographers, or managers?
Those are real questions, but they are incomplete. The more immediate change may be stranger and more invasive than simple replacement.
AI will not only automate work. It will also measure work.
It will classify, summarize, score, monitor, compare, route, escalate, and evaluate. It will sit inside ticket queues, inboxes, chat systems, document repositories, customer-service platforms, code repositories, CRMs, HR systems, security tools, calendars, and productivity suites. Some of that will be useful. Some of it will reduce drudgery. Some of it will make organizations safer and more responsive.
But some of it may also create a form of workplace micromanagement at a scale that was not previously practical.
The near future of AI is therefore not just a story about machines doing human tasks. It is also a story about humans being reorganized around machine-readable work.
—
The “slowly, then all at once” phase
AI adoption is already broad, but its deeper organizational impact is still uneven. McKinsey’s 2025 global AI survey found that 88% of respondents said their organizations were using AI regularly in at least one business function, up from 78% the year before. At the same time, only about one-third said their organizations had begun scaling AI programs across the enterprise. Agentic AI is even earlier: 23% said their organizations were scaling an agentic AI system somewhere, while another 39% were experimenting, but no single business function had more than 10% of respondents reporting scaled agent use.
That gap matters. It suggests we are not in the “AI has already transformed everything” phase. We are in the “AI is everywhere, but institutions have not fully reorganized around it” phase.
That is exactly where “slowly, then all at once” changes tend to form. First, a technology appears as a tool. Then it becomes a habit. Then it becomes infrastructure. Then it becomes an expectation. Eventually, work that used to seem normal starts to look irrationally slow.
A person who can use AI well may already be faster. A team that redesigns its workflow around AI may become dramatically faster. An organization that redesigns its assumptions around AI may begin to need different people, different policies, different metrics, and different forms of accountability.
The important shift is from AI as assistant to AI as workflow layer.
—
From chatbot to workflow layer
The most important near-term AI systems are not just chatbots. They are systems that can take a goal, use tools, inspect files, call APIs, write code, summarize documents, open tickets, classify requests, and ask humans for approval when needed.
METR’s work on AI task horizons is useful here. METR defines a model’s 50% time horizon as the length of task, measured by how long it would take a human professional, that an AI system can complete autonomously with 50% reliability. Their research found that frontier AI time horizons have roughly doubled every seven months across software and research tasks, with possible acceleration during 2024.
This does not mean AI systems can simply replace reliable professionals. They still fail in jagged, surprising ways. They are often brittle, overconfident, and poor at knowing what they do not know. But the direction is clear: AI systems are moving from answering questions toward completing bounded tasks.
That matters because many jobs are not single tasks. They are bundles of tasks. If AI can take over 20%, 30%, or 50% of a role’s routine components, the job does not need to vanish for the workplace to change dramatically. The job can be compressed, rebundled, monitored, or made more demanding.
Anthropic’s March 2026 Economic Index report points in this direction. It notes that prior analysis found 49% of jobs reviewed had seen at least a quarter of their tasks performed using Claude, and its newer data showed more diverse work-related usage patterns, with augmentation increasing slightly.
The likely near-term future is not “everyone is replaced.” It is more like: fewer people are needed for routine production, while more value shifts toward orchestration, judgment, review, exception handling, and accountability.
—
The second automation: measuring the worker
This is where workplace monitoring enters the story. Once work becomes digital, AI-readable, and routed through platforms, it becomes easier to measure. Once it becomes easier to measure, organizations will be tempted to manage through those measurements.
Some monitoring will be defensible. A help desk should know ticket volumes, aging incidents, recurring problems, escalation rates, customer satisfaction, and documentation gaps. A security team should know whether accounts are being attacked, whether patches are missing, and whether suspicious behaviour is occurring. A software team should know whether tests pass, whether vulnerabilities exist, and whether deployments fail.
The danger is not measurement itself. The danger is bad measurement becoming management.
The OECD’s 2025 report on algorithmic management defines these systems as tools that can allocate, instruct, monitor, and evaluate work. Its employer survey found striking levels of monitoring and evaluation, especially in the United States. The report says 55% of surveyed U.S. firms monitored the content and tone of conversations, voice calls, or emails, compared with 6% in Europe and 8% in Japan. It also says 90% of U.S. firms used an algorithmic management tool to partially or fully automate at least one evaluation task.
That is the watchtower side of the AI transition.
A world of AI-enabled workflows can become a world of AI-enabled surveillance: tone scoring, keystroke inference, ticket-speed rankings, idle-time suspicion, screenshot analysis, sentiment tracking, meeting participation scoring, “responsiveness” dashboards, and automated performance flags.
The perverse incentive is obvious. If a company can measure a thing cheaply, someone will eventually ask whether it can be optimized. But not everything measurable is meaningful. A worker who resolves fewer tickets may be handling harder tickets. A person with fewer messages may be doing deeper work. A slower response may be a more careful response. A quiet employee may be preventing future problems that never show up in a dashboard.
AI does not remove this problem. It can intensify it by making weak proxies look authoritative.
—
The productivity trap
The promise of AI in the workplace is that it can reduce busywork. The risk is that it simply raises the expected output of each worker while increasing surveillance.
That is the productivity trap: every efficiency gain becomes a new baseline.
At first, AI helps a worker write a better email, summarize a document, draft a script, or troubleshoot an issue. Then managers notice that work can be done faster. Then faster becomes normal. Then normal becomes quota. Then the worker is no longer experiencing AI as assistance, but as pressure.
This is one reason the “AI will free us from drudgery” story is incomplete. It might. But only if institutions choose to treat efficiency as slack, resilience, quality, and human capacity. If they treat it only as extraction, AI becomes a ratchet.
This will vary dramatically by job. High-trust, high-skill workers may gain leverage. Low-autonomy workers may gain monitoring. Some professionals will become managers of AI systems. Others will become inputs into AI-managed systems.
The future workplace may divide into three broad categories. {Adrian notes: This part came out sounding pretty grim in general.}
First, there will be monitored workers, whose activity, pace, tone, and output are constantly evaluated.
Second, there will be AI-assisted workers, who use AI to produce more work but remain accountable for the result.
Third, there will be AI workflow owners, who design, audit, govern, and improve the systems themselves.
The practical career question is: which side of that divide do you want to be on?
—
Regulation is arriving, but unevenly
Regulation will shape this, but it will not eliminate the issue.
The European Union’s AI Act takes a risk-based approach and bans certain “unacceptable risk” uses of AI. The European Commission’s AI Act overview lists prohibited practices including harmful manipulation, social scoring, some biometric categorization, and emotion recognition in workplaces and education institutions, with limited exceptions such as medical or safety reasons.
That is significant because it recognizes that AI used on workers is not just a productivity tool. It can affect rights, dignity, privacy, and power.
But rules vary by jurisdiction. In Ontario, for example, employers with 25 or more employees must have a written policy on electronic monitoring, and the policy must describe whether and how employees are monitored. But the Ontario government’s own guidance is explicit that these requirements do not establish a right not to be electronically monitored and do not create new privacy rights for employees.
That distinction is crucial. Transparency is not the same as protection. A worker may be told they are monitored without having much practical power to refuse.
—
Authenticity collapse is part of the same story
At first, workplace monitoring and fake AI content may seem like separate problems. They are not.
Both are consequences of the same shift: AI makes production cheap, and cheap production creates a crisis of trust.
If anyone can generate a plausible article, image, video, résumé, animal photo, product review, legal summary, or corporate update, then the valuable thing is no longer mere production. The valuable thing is provenance, judgment, and trust.
This is already visible online. Generic AI content is everywhere, but much of it feels empty. People skim past it. It has the structure of meaning without the cost of experience. It is fluent, but not necessarily grounded. It is polished, but not necessarily true.
AI images are even more destabilizing. They are interesting, beautiful, and useful, but they also pollute the visual commons. Fake animal pictures, fake historical photos, fake disaster images, fake scientific diagrams, fake product photos, fake political images, and fake evidence all create a world where “seeing” becomes less reliable.
This matters especially for children and education. If search results, social feeds, and learning materials fill with plausible-but-wrong images of animals, places, people, and events, children may absorb a distorted visual model of the world. The issue is not only deception. It is ambient unreliability.
Provenance standards are one response. The Coalition for Content Provenance and Authenticity says Content Credentials are intended to help people understand where digital content comes from and how it changed. In February 2026, C2PA announced Content Credentials 2.3 and said more than 6,000 members and affiliates had live applications of the standard.
Provenance will not solve everything. Metadata can be stripped. Trust chains can be broken. Bad actors can lie. But provenance points toward the right principle: in a world of cheap generation, the history of an object matters.
The same principle applies to work. If AI agents produce outputs, organizations need to know what model was used, what prompt was used, what data was accessed, what tools were called, what files were changed, what human approved the action, and what uncertainty remained.
In other words: the future needs receipts.
—
Taste and judgment become economic skills
This is why “taste” and “judgment” are not soft luxuries. They are central future skills.
When production is expensive, the ability to produce is valuable. When production is cheap, selection becomes valuable. Editing becomes valuable. Verification becomes valuable. Taste becomes valuable. Accountability becomes valuable.
A person using AI badly can create infinite mediocre output. A person using AI well can create useful work faster. A person with judgment can decide which outputs should exist at all.
That distinction matters. In a world flooded with AI writing, the scarce person is not the person who can generate more text. It is the person who can say:
> This is wrong.
> This is unsupported.
> This sounds plausible but is not useful.
> This metric will harm workers.
> This image is misleading.
> This workflow needs a human approval point.
> This agent should not have that permission.
> This task should be automated.
> This task should not.
> This content needs provenance.
> This dashboard is measuring the wrong thing.
That is not anti-AI. That is mature AI use.
—
The energy and infrastructure constraint
There is also a physical side to this story. AI feels weightless because it appears in a browser or app, but the infrastructure behind it is enormous.
The International Energy Agency reported in April 2026 that electricity demand from data centres rose by 17% in 2025, with AI-focused data centres growing even faster. The IEA says capital expenditure by five large technology companies exceeded $400 billion in 2025 and was set to increase by another 75% in 2026. It is also important to note that power consumption per AI task is experiencing a rapid decline as efficiency improves at a rate unprecedented in energy history. However as more people utilize AI, and energy-intensive uses of it (eg AI agents) general use continues increasing. As a result, electricity consumption from data centres is set to double by 2030. It expects data-centre electricity consumption to double by 2030, while power use from AI-focused data centres is poised to triple.
This matters because it means AI’s future will not be determined only by clever models. It will also be shaped by chips, power, cooling, grid connections, regulation, cost, and geopolitics.
That may slow some hopes of unlimited AI. It may also intensify pressure to route tasks intelligently: small models for simple work, local models for private or low-cost tasks, frontier models for hard reasoning, and human review where consequences are high.
Again, this points toward a practical future skill: not just using AI, but operating AI systems responsibly and economically.
—
Forecast: what happens next
Six months
Over the next six months, AI will continue becoming normal in professional workflows. The biggest change will be expectation. More workers will be expected to summarize faster, draft faster, research faster, script faster, and respond faster. Organizations will keep experimenting with agents, especially in IT, knowledge management, customer support, software, and operations.
Twelve months
Over the next twelve months, workflow redesign will become more visible. More companies will ask why certain work is still manual. AI policies will become more common, but many will lag actual usage. Employee monitoring concerns will grow as AI becomes embedded in collaboration and productivity platforms.
Twenty-four months
Over the next twenty-four months, the divide between AI-native and AI-sprinkled organizations will widen. Some teams will use AI to eliminate bottlenecks and improve quality. Others will use it mainly to squeeze more output from workers. The same technology will produce very different cultures depending on governance.
Five years
By 2030 and beyond, the labor market will likely be heavily reshaped. The World Economic Forum’s 2025 Future of Jobs Report projected that structural labor-market transformation could affect 22% of jobs by 2030, with 170 million jobs created and 92 million displaced, for a net increase of 78 million. It also found that nearly 40% of job skills are expected to change.
Forecasts like that should not be treated as prophecy. But they are useful directional signals. The future is unlikely to be simple mass replacement. It is more likely to be churn, rebundling, compression, uneven opportunity, and uneven harm.
—
The better path: accountable AI workflows
The central challenge is not whether AI should be used. It will be used.
The challenge is whether it will be used as a tool for human leverage or as a tool for human compression.
A healthier AI workplace would measure workflows more than bodies. It would track systems, bottlenecks, errors, queues, failure modes, approval points, and customer outcomes. It would avoid fake productivity metrics like keystrokes, mouse movement, or simplistic response-time scoring. It would make AI actions auditable. It would require human review for high-impact decisions. It would distinguish between assistance and evaluation. It would protect privacy. It would allow workers to challenge automated conclusions.
A worse AI workplace would do the opposite. It would convert every digital trace into a performance signal. It would confuse speed with value. It would evaluate tone without context. It would punish invisible work. It would use AI to generate quotas, then use monitoring to enforce them.
The technology does not decide which version we get. Institutions do. Managers do. Vendors do. Regulators do. Workers do, where they have leverage. Technologists do, when they build the systems.
—
The career opportunity
For individuals trying not to be left behind, the goal should not be merely to “learn AI.” That phrase is too vague now.
The stronger goal is to become useful at the layer where AI meets real work.
That means learning how to:
– design AI-assisted workflows;
– log model use, prompts, tool calls, costs, and outputs;
– build human approval gates;
– evaluate outputs for accuracy and usefulness;
– protect sensitive data;
– distinguish good metrics from harmful metrics;
– explain AI risks without sounding anti-technology;
– preserve provenance for documents, images, and decisions;
– use small, local, or cheap models where appropriate;
– reserve expensive frontier models for tasks that justify them.
This is a practical lane between naive boosterism and blanket rejection.
The people most valuable in the next phase may not be the ones who produce the most AI content. They may be the ones who can make AI output trustworthy, useful, governed, and connected to real human purposes.
—
Conclusion: the future is not just automated; it is instrumented
The next AI shock will not be a single event. It will be a sequence of smaller shifts that compound.
A ticket queue becomes AI-triaged.
A meeting becomes AI-summarized.
A codebase becomes AI-maintained.
A document repository becomes AI-searchable.
A worker’s tone becomes AI-scored.
A photo becomes suspect until provenance is checked.
A dashboard becomes a manager’s view of reality.
A small team suddenly does the work of a much larger one.
A junior role quietly stops being hired.
A new kind of worker appears: the person who manages the AI-managed workflow.
That is the real “slowly, then all at once.”
The future will not simply ask whether humans can still produce. Humans will produce with machines constantly. The harder question is whether humans can still judge, verify, govern, and refuse.
AI will make work faster. The unresolved question is whether it will make work better.
The answer depends less on the models than on the systems we build around them.
—
Source links
– McKinsey — The State of AI: How organizations are rewiring to capture value
– METR — How does AI task horizon vary across domains?
– Anthropic — Economic Index, March 2026 report
– OECD — Algorithmic Management in the Workplace
– European Commission — AI Act regulatory framework overview
– Ontario — Written policy on electronic monitoring of employees
– C2PA — Content Credentials 2.3 announcement
– IEA — Data centre electricity use surged in 2025
– World Economic Forum — Future of Jobs Report 2025 press release