The Future of Work With Agentic AI: 2026-2030 Outlook

I have spent the last two years thinking about, building with, and writing about agentic AI, and I have grown deeply sceptical of confident predictions on either end of the spectrum. The doomers who insist that 70% of knowledge work will be automated by 2028 are usually selling something. So are the cheerleaders who claim AI will simply make every worker 30% more productive with no displacement. The reality, in my view, is messier, more uneven and more interesting.

This piece is my attempt to lay out what I actually believe about how work changes over the next four years, with explicit confidence levels attached to each prediction. I draw on the productivity research that I find credible, the regulatory signals I have been tracking, and the patterns I observe directly in the companies I work with. I try to distinguish what is already happening from what is plausible from what is speculative.

If you are a worker trying to make career decisions, my goal is to give you a clearer mental model than the doomscroll versions. If you are a leader thinking about workforce planning, I have tried to be honest about what I do and do not know. If you are early in your career, the sections on which skills compound matter most.

I will be wrong about some of this. Probably about more than I expect.

The history of technology forecasting is humbling. Experts in 1995 underestimated the internet’s economic impact and overestimated its speed of penetration in specific industries. Experts in 2015 overestimated self-driving cars and underestimated language models. The mode of failure is consistent: predictions either over-extrapolate from a single technology curve or under-account for the friction of real organisations.

I try to follow three rules to avoid these traps.

First, I attach explicit confidence levels to every claim. If I say “high confidence” I mean I would bet meaningful money on the outcome. “Medium confidence” means I lean that way but reasonable people will disagree. “Speculative” means I am offering a structured guess.

Second, I distinguish between “the technology can do this” and “organisations will deploy it widely”. The gap between capability and adoption is usually larger than people expect.

Third, I separate the question of what happens by 2027 from the question of what happens by 2030. Three years versus six years is a meaningfully different forecast horizon.

The most dangerous confidence is the confidence that has not specified a time horizon.

I will sometimes hedge. I am not trying to give you certainty I do not have.

The dominant debate in 2026 is whether agentic AI augments workers (raising their productivity) or replaces them (removing the work entirely). My view is that the framing is wrong.

The actual pattern, observable in companies that have deployed agentic AI seriously, is more nuanced. Tasks are unbundled from jobs. Some tasks become heavily automated. Other tasks become more valuable. The overall job changes shape, often substantially, but rarely disappears outright in a single year.

Concretely, I have seen first-line customer support agents go from spending 90% of their time on tier-one resolution to spending 60% of their time on edge cases that the agentic system escalates. That same agent now spends more time on the complex 40% of cases that genuinely require human judgement. The job changed; the worker did not disappear.

But, and this matters, the headcount required for the same business volume often does drop. If an agentic system handles 60% of tier-one volume previously handled by humans, a company that grows 20% per year will hire fewer support agents than it would have. That is not “no replacement”, but it is also not “mass redundancy”.

The honest framing, in my view, is “task unbundling with headcount compression”. Not a slogan, but a useful mental model.

The jobs that change first share three characteristics: their work is largely digital, well-documented and patterned. This is not the same as “their work is unskilled”. Many highly skilled jobs are also highly patterned.

The roles where I see the most change between now and 2027:

• Customer support, particularly tier-one
• Inside sales development representatives
• Junior software engineering tasks (not roles, tasks)
• Data analysis and reporting
• Recruiting sourcing and screening
• Legal contract review and basic drafting
• Financial analysis and research
• Marketing content production
• Bookkeeping and basic accounting
• Quality assurance for software

Confidence: high that these roles see significant change by 2027. Medium that the change manifests as net headcount reduction rather than role evolution. Lower that the change happens uniformly across geographies and industries.

I expect a fairly aggressive change curve in customer support and inside sales, where the deployment economics are clearest and the integration paths are well understood. I expect a more gradual curve in legal and finance, where regulatory friction slows adoption.

The order is determined by deployment friction, not by capability.

The same agentic technology may be ready for both customer support and legal work in 2026, but customer support deployments will outpace legal by years.

The flip side of the unbundling argument is that some roles grow substantially. These are roles that benefit from the same agentic technology, either by becoming more impactful or by being the supply side of the change.

The growth roles I see clearly:

• AI agent engineers, of course
• AI solution architects
• AI evaluation and trust specialists
• AI governance and compliance roles
• Human-in-the-loop reviewers (a category that did not exist three years ago)
• Domain specialists who can codify expertise into agent specifications
• Sales engineers selling agentic systems
• Senior engineers and managers who can leverage agentic tooling at scale
• Trainers and educators for agentic AI
• Mental health and human services roles, growing for demographic reasons but also as displaced workers seek transitions

The most under-discussed growth category, in my view, is the human-in-the-loop reviewer. As agentic systems become widespread, there is genuine economic value in the people who calibrate, audit and oversee them. This is a new class of work, and I expect it to absorb a meaningful fraction of displaced workers from adjacent roles.

Confidence: high that these categories grow. Medium that growth is fast enough to absorb displacement in net terms. Low that the geographic distribution of new jobs matches the geographic distribution of lost ones.

It is worth being honest that many roles will see relatively little change in the next four years.

Construction, electrical work, plumbing, nursing, teaching of young children, hospitality, frontline healthcare, agriculture and skilled trades will all change at the margins as agentic systems improve back-office and scheduling tasks. But the core work, the part that requires physical presence and embodied judgement, remains human work for the foreseeable future.

This is not a permanent guarantee. Robotic capability is improving, and longer time horizons may reshape this list. But on a 2026-2030 window, these roles are largely safe from the kind of disruption hitting digital knowledge work.

Confidence: high.

This matters for career advice. The narrative that “all jobs are at risk” is wrong and unhelpful. A meaningful fraction of the labour market is, for the foreseeable future, on stable ground. Workers in these roles do not need to panic-pivot into AI. They may, however, benefit from learning to use agentic tools that augment their existing work.

A nuance worth flagging. Roles that combine digital and physical work, such as field service technicians, fall into a middle ground. The digital portion will see substantial change; the physical portion will not. The job evolves, the worker continues.

The productivity research on AI tooling, particularly the Microsoft Copilot studies, the GitHub Copilot studies, and various enterprise pilot reports, is more ambiguous than headlines suggest.

The honest summary, in my view:

• For narrow, well-defined tasks, AI tools produce reliable productivity gains in the 20-40% range
• For broad knowledge work tasks, gains are smaller and less consistent, often 5-15%
• For complex coordination work, gains are sometimes negative in early adoption phases
• Variance across individual workers is enormous, often larger than the average effect
• Senior workers often benefit less in percentage terms than mid-level workers (they were already efficient)

This is not the “100% productivity gain” some marketing materials suggest. It is also not “no effect”. It is a real, measurable, modest-to-substantial gain in well-targeted domains, with high variance and non-trivial implementation friction.

For agentic AI specifically, the research is too thin to draw strong conclusions yet. The companies that have deployed agentic systems seriously, particularly in customer support and software engineering, report larger gains than the Copilot studies showed, but the sample sizes are small and the selection effects are large.

Confidence: medium on the headline numbers. High that the variance across workers and tasks is the most under-discussed feature of the data.

The skills that pay off in 2026 and beyond are shifting in ways that are not yet reflected in most education curricula.

The skills that are becoming more valuable:

• Ability to specify problems precisely (so an agent can be instructed)
• Ability to evaluate and audit AI outputs (so you can trust them)
• Domain depth that combines with AI fluency (so you can codify expertise)
• Systems thinking and orchestration (so you can compose tools effectively)
• Communication and stakeholder management (so you can lead AI-enabled change)

The skills that are becoming less differentiating:

• Routine document production
• Boilerplate code writing
• Basic data manipulation and reporting
• First-pass information gathering
• Rote knowledge recall

This shift has profound implications for education. Traditional curricula that emphasise rote production are increasingly out of step with the labour market. Curricula that emphasise specification, evaluation and orchestration are aligned with where the demand is going.

I do not think universities will pivot quickly. The lag between curriculum and labour market reality is significant, and most students entering university now will graduate into a different skill landscape than the one their courses were designed for.

The most valuable thing a young person can do in 2026 is build the meta-skill of learning new tools quickly and judging their outputs critically.

This applies more than any specific tool or framework.

The generational impact of agentic AI is uneven, and the framing matters.

Workers in their twenties face the most acute pressure on entry-level jobs but also have the most adaptation runway. The entry-level analyst, junior developer and graduate trainee roles that have historically served as career launch pads are compressing fastest. This is a real problem and one that the labour market has not solved.

Workers in their thirties and forties face a different challenge: their accumulated craft is valuable, but the tools they work with are changing fast. The risk is not displacement but skills depreciation. Workers who actively upskill remain competitive; those who do not find their relative value declining.

Workers in their fifties and beyond face the most variable picture. For those in roles that are largely safe (skilled trades, healthcare, education), little changes. For those in roles facing acute disruption (some white-collar professions), the runway to retirement may shape decisions more than reskilling.

The under-discussed generational effect is on people aged 16-22, who are choosing careers now. The career advice they received from parents and schools is increasingly stale. This is the cohort I worry about most.

Confidence: high on the direction of these effects. Medium on their magnitude.

Regulation will shape agentic AI deployment as much as the technology itself.

In the EU, the AI Act and its agentic-specific extensions create real friction in high-risk domains. By 2028 I expect a fairly mature compliance regime around agentic systems in healthcare, finance and critical infrastructure, with documentation and audit requirements that meaningfully slow deployment.

In the United States, the regulatory picture is more fragmented. Federal regulation remains slower than in the EU, but state-level regulation, particularly in California, New York and Texas, fills the gap unevenly. Sector regulation, particularly in finance through the OCC and SEC, has progressed faster than general regulation.

In India and other major Asian economies, the regulatory posture is more permissive on deployment with stronger emphasis on data localisation and access controls. I expect this to remain the case through 2030.

The regulation I most expect to matter for workers:

• Disclosure requirements when agents take consequential actions (high confidence)
• Audit trails for high-stakes decisions (high confidence)
• Worker notification requirements for significant AI deployments (medium confidence)
• Liability frameworks distinguishing AI errors from human errors (medium confidence)

For most workers, regulation creates a buffer of human oversight roles that did not exist before. The compliance cost of pure AI deployments creates economic pressure to retain human review in many contexts.

The geographic distribution of agentic AI’s impact will be deeply uneven.

Major tech hubs, including the San Francisco Bay Area, Seattle, Bangalore, Hyderabad, London, Tel Aviv and Singapore, will see the largest concentration of new high-value jobs. They will also see substantial disruption in their existing white-collar workforce.

Mid-sized cities with strong professional services sectors will see substantial disruption, slower job creation in agentic AI specifically, but potential gains in supporting industries.

Rural and small-town economies will see less direct impact in either direction, although the secondary effects through national economic restructuring will be real.

For workers, location flexibility becomes a strategic asset. Workers who can move toward concentrations of agentic AI activity have more options. Workers who cannot, or who choose not to, should plan more deliberately around roles that are physically distributed.

Remote work mitigates this somewhat. The companies hiring most aggressively for agentic AI roles in 2026 are disproportionately remote-friendly. This is genuinely good news for workers in non-hub geographies, although competition for those remote roles is global.

If I were giving career advice to someone in their twenties or thirties in 2026, here are the compound skills I would emphasise.

Domain depth plus AI fluency. Picking a domain (healthcare, finance, law, education, manufacturing) and combining genuine domain expertise with AI fluency is the single most valuable combination. Agents need someone who can specify the work; that person is the domain expert with technical literacy.

Evaluation and judgement. The ability to look at an AI output and reliably assess whether it is correct, appropriate and complete is enormously valuable. This skill is harder to acquire than it looks and rewards practice.

Systems thinking. The ability to compose tools, agents and humans into workflows that produce reliable outcomes is the core skill of the next decade.

Communication and translation. The ability to translate between technical and non-technical stakeholders is rising in value as agentic systems require ongoing calibration.

Adaptability. The willingness and capacity to learn new tools every six to twelve months is the underrated meta-skill.

These compound. A worker who has all five is enormously valuable. A worker who has one or two is competitive. A worker who has none is, frankly, vulnerable.

The good news is that none of these skills require a particular educational background. They reward sustained practice more than credentials.

I find this to be one of the more democratising aspects of the current moment, although you have to actively engage with it.

If you are reading this and trying to figure out what to do, here is the strategy I would adopt.

First, assess your current role honestly. Where does it sit on the disruption spectrum? What fraction of your tasks could be automated by an agent today, and what fraction in two years? Be honest. The hardest part of this exercise is being honest.

Second, identify the parts of your work that are most resistant to automation. These are usually the parts that involve genuine judgement, relationships, or accountability. Invest in deepening them.

Third, learn at least one agentic AI tool well enough to use it daily. Not as a passive consumer, but as a competent operator. This single discipline puts you ahead of most of the workforce.

Fourth, build optionality. A second skill, a side project, a network of peers in adjacent domains, all create alternative paths if your primary role becomes uncomfortable.

Fifth, do not panic. The pace of change is significant but slower than headlines suggest. You have time, more than you think, less than you would like, to adapt.

The workers who do well over the next four years will not be those who tried to predict the future perfectly. They will be those who built adaptive capacity and engaged with the technology directly.

Here are concrete predictions for 2027 and 2030, with confidence levels attached.

Some of these will turn out wrong. I expect the regulation predictions to be most reliable; the macro displacement predictions to be hardest.

The doomers tend to underestimate organisational friction. Even technologies that work perfectly take years to penetrate large organisations. The displacement curve is slower than the capability curve. They also underestimate the human appetite for human interaction, particularly for high-stakes decisions.

The cheerleaders tend to underestimate the unevenness of the productivity gains, the genuine disruption to entry-level career paths, and the geographic concentration of new opportunities. They also tend to conflate “no net displacement” with “no painful transitions”.

The truth, I think, is that agentic AI is one of the most significant labour market shifts in a generation, but not a discontinuity. It will reshape work in ways that benefit some and challenge others, with the burden of adaptation falling disproportionately on workers who are mid-career, lower-paid and geographically immobile.

The policy response to this matters enormously. The technology decisions matter less than the policy decisions about training, transition support and labour protections.

The future of work is not determined by the technology. It is shaped by the decisions we make about how to deploy it.

I find that framing more useful than fatalism in either direction.