Which Roles Are Changing, Which Are Growing, and Which Skills Hold Their Value
Predictions about artificial intelligence and employment have been unusually poor. The confident forecasts of a few years ago named the wrong occupations, got the direction of change wrong in several fields, and consistently underestimated how much of any job consists of things the technology cannot yet touch. It is worth understanding why those predictions failed before making new ones.
This article takes a different approach. Rather than listing jobs that will disappear, it examines which tasks are being absorbed, how that redistributes work inside existing roles, and which capabilities have become more valuable rather than less. The picture that emerges is less dramatic than the headlines and considerably more useful for planning a career.
Why the Early Forecasts Missed So Badly
The forecasts that aged worst shared one methodological flaw: they treated a job title as a single unit and asked whether a machine could do it. But almost no job is one task. A radiologist reads images, but also consults with referring physicians, handles ambiguous cases, manages liability, trains residents, and explains findings to frightened patients. Automating the image reading changes the job substantially without eliminating it.
The second flaw was assuming that automating a task reduces demand for the people who did it. Sometimes the opposite occurs. When a task becomes dramatically cheaper, the volume of work that becomes economically viable often expands faster than the efficiency gain, and total employment in the field rises. Spreadsheets did not reduce the number of accountants; they expanded what accounting was expected to cover.
Ask what a technology makes cheap, then ask what becomes newly worth doing at that price. The second answer predicts employment better than the first.
An economic rule of thumb
- The bundling error → treating a job as one automatable unit rather than a portfolio of dozens of tasks with different exposure.
- The demand error → assuming fixed demand, when cheaper capability usually expands the market for the surrounding human work.
- The last-mile error → underestimating how much of professional work is accountability, relationship, and judgement rather than production.
- The timeline error → confusing what is technically demonstrated with what is deployed, regulated, insured, and trusted at scale.
Tasks, Not Jobs: Using the Right Unit of Analysis
A more reliable method is to break a role into its constituent tasks and score each one on two axes: how mechanical it is, and how costly an error would be. That produces four quadrants with very different trajectories.
| Quadrant | Character of the task | Trajectory | Examples |
|---|---|---|---|
| Mechanical, low stakes | Repetitive, checkable, recoverable errors | Fully automated already | Transcription, data entry, first-draft copy, tagging, routine translation |
| Mechanical, high stakes | Repetitive but errors are expensive | Automated with mandatory human sign-off | Contract review, medical coding, compliance checks, payroll |
| Judgement, low stakes | Requires taste but mistakes are cheap | Machine-assisted, human-directed | Editorial choices, visual design iteration, campaign concepts |
| Judgement, high stakes | Ambiguous, consequential, relational | Largely untouched, rising in value | Negotiation, diagnosis under uncertainty, crisis management, leadership |
Run this exercise on your own work and the result is usually reassuring and clarifying at the same time. Perhaps forty per cent of your week sits in the first quadrant and is genuinely at risk. But that forty per cent is often the part you least enjoy, and the remaining sixty per cent becomes the whole job rather than the part you squeeze in around admin.
What Is Actually Happening, Role by Role
Software Development
The field where change has been fastest and where predictions were most wrong. Agentic coding tools now write substantial amounts of production code, handle migrations, and generate test suites. Yet demand for engineers has not collapsed, because the bottleneck moved rather than vanished. It now sits in system design, reviewing machine-generated code for subtle correctness, understanding legacy constraints, and deciding what to build. Junior roles have genuinely become harder to enter, which is a real problem the industry has not solved.
Writing and Content
Commodity content production has collapsed in value, and anyone whose offer was volume at a low price has been displaced. Simultaneously, demand rose for writing with a genuine point of view, original reporting, and domain expertise deep enough that a model cannot fake it. The market split into two: near-zero-value output and clearly premium work, with the comfortable middle largely gone.
Design
Routine production work, resizing, variants, basic layout, stock illustration, moved substantially to tooling. Identity work, art direction, and the judgement of what looks right for a particular audience did not. Designers who positioned themselves as production capacity struggled; designers who positioned themselves as decision-makers gained leverage, because they can now execute far more of their own vision.
Customer Support
First-line volume is now largely automated in well-run operations. The role that remains is more demanding and better paid: handling the escalations, the angry customers, the genuinely novel problems, and the cases where the answer requires someone to take responsibility. Headcount fell, average seniority rose.
Teaching and Training
One of the clearest gains rather than losses. Content delivery and assessment became cheap, which freed instructor time for the part that actually determines outcomes: noticing which student is lost, adjusting explanation in real time, and sustaining motivation. Demand for good teaching rose because personalised support became affordable at scale.
| Field | Tasks absorbed | Tasks that grew | Net direction |
|---|---|---|---|
| Software | Boilerplate, tests, migrations, documentation | Architecture, review, security judgement | Stable senior, harder junior entry |
| Content | Commodity drafting, SEO filler, basic translation | Original reporting, expertise, distinctive voice | Polarised into two markets |
| Design | Variants, resizing, stock imagery, routine layout | Art direction, identity, taste | Fewer producers, more directors |
| Support | Tier-one responses, tagging, routing | Escalation, retention, accountability | Smaller and more senior |
| Education | Content delivery, marking, drills | Diagnosis, motivation, adaptive explanation | Growing |
| Finance | Reconciliation, categorisation, basic reporting | Advisory, scenario judgement, client trust | Shifting upward |
The Skills That Gained Value
If production capacity is abundant, the scarce inputs are the ones that decide what to produce and whether the result is any good. Five capabilities have consistently appreciated.
- Problem definitionThe ability to look at a messy situation and articulate precisely what needs solving. Machines are excellent at answering well-posed questions and useless at deciding which question is worth asking. This is the single most transferable skill in the current environment.
- Verification and tasteKnowing whether a plausible output is actually correct, and whether a competent piece of work is genuinely good. This requires enough domain depth to spot a subtle error, which is exactly why the erosion of junior roles is worrying: taste is built by doing the work.
- AccountabilitySomeone has to sign their name to a decision and absorb the consequences. This is not a technical capability and it does not automate, which is why it increasingly defines senior roles across every field.
- Systems thinkingUnderstanding how components interact, where a change will have second-order effects, and which constraints are real. As individual tasks get automated, the value concentrates in whoever understands the whole.
- Relationship workTrust, persuasion, difficult conversations, and reading a room. Durable partly because it is genuinely hard for machines and partly because people prefer humans for it even when a machine would suffice.
The Uncomfortable Middle
An honest account has to name a real problem rather than ending on reassurance. The traditional path into professional expertise ran through exactly the tasks that are now automated. You learned to write by drafting things nobody much cared about. You learned to code by building simple features under supervision. You learned to design by producing variations. Those apprenticeships are disappearing, and nothing has replaced them.
The result is a hollowing at the entry level that particularly affects people early in their careers and those changing fields. Organisations that solve this, by deliberately creating learning work even when a machine could do it faster, will have a serious talent advantage in five years. Most organisations are not currently solving it.
We have automated the exercises that produced expertise while continuing to require expertise. That contradiction is the defining workforce problem of this decade, and it will not resolve itself.
Tooliqo Editorial
Practical Positioning: What to Do This Quarter
- Audit your own tasks → list what you did last week, sort it into the four quadrants, and note what proportion is mechanical and low-stakes. That figure is your exposure, and knowing it beats worrying about it.
- Automate your own first quadrant → before someone else does it for you. Being the person who removed the tedious part of your job is a much better position than being the person it was removed from.
- Deepen rather than broaden → shallow generalist knowledge is what models supply most cheaply. Genuine depth in one domain is what lets you verify their output and is therefore where your leverage sits.
- Take on accountability deliberately → volunteer for the work where someone has to decide and answer for it. This is uncomfortable and it is the most durable career position available.
- Build the relationships → in a market where output is abundant, trust is the scarce good. The people who know your work are your actual moat.
None of this requires predicting which model will win or which tool to bet on. It requires an honest inventory of what you spend your time on and a deliberate shift toward the parts that got more valuable rather than less. That is a manageable quarter's work, and it beats waiting for clarity that is not coming.
Key Takeaways
- Analyse tasks, not job titles. Almost every role is a portfolio with mixed exposure.
- Cheaper capability often expands demand for the surrounding human work rather than reducing it.
- Production is shrinking across nearly every field while judgement and accountability expand.
- Problem definition, verification, accountability, systems thinking, and relationships all gained value.
- The erosion of entry-level work is a genuine unsolved problem, not a transitional inconvenience.
- Automate your own routine tasks before the decision is made for you.
Frequently Asked Questions
Which fields are genuinely safest?
Any work combining physical presence, high-stakes judgement, and relationship trust: skilled trades, healthcare delivery, teaching, complex negotiation, and senior operational leadership. Note that safety here means the role persists, not that it stays unchanged. Every one of these fields is absorbing tools into its daily practice.
Should I learn to code if I am starting out now?
Yes, but for a different reason than five years ago. The value is no longer in typing syntax, which tools handle well. It is in understanding how software actually works so you can direct it, review it, and recognise when a plausible-looking solution is wrong. Learn enough to read and judge code even if you never write much of it by hand.
Is a degree still worth it?
For fields with licensing requirements, obviously yes. Elsewhere the calculation has shifted toward what the programme actually gives you: supervised practice on real problems, a network, and depth in a domain. Programmes that deliver those remain valuable. Programmes that mainly deliver information you could obtain free are much harder to justify than they were.
How do I compete when I am early in my career?
Deliberately seek the work that builds verification skill even when a machine could do it faster, because that skill is what makes you employable at the next level. Be visibly useful with the tools, since fluency still separates candidates. And narrow your focus sooner than previous generations did: depth in something specific is easier to demonstrate and harder to replace than general competence.
0 Comments