Concept

A useful choice of what to learn links your skills to work that still needs doing as tools change.

Issue

General advice to learn technical skills or people skills can leave you unsure what to practise next.

Solutions

Look at what your work really needs, then choose your next learning target.

Say you’ve used AI to draft a report. The page looks finished, but you still need to check the figures and explain what the recommendation means for another team.

You could spend your next learning session making the draft better, understanding the numbers or explaining the decision. Which would improve the work most?

Start with what the work needs, rather than hunting for a skill that technology can never replace. A useful skill helps you produce a result someone needs. But doing better work, finding more job openings and earning more are different things.

This module is for working adults choosing what to learn in the job they have now, or in a next job they could realistically get. It is about the learning you choose, alongside any training and qualifications you’re required to have.

What you’ll get from this module

In this module

By the end of this module, you’ll have a clearer way to choose what to learn next. We’ll link research on jobs over past decades, a training experiment and AI workplace studies that point different ways to the choices in your own work. You’ll pick a situation from daily life and use a written look at your tasks to choose a learning step you can afford, or to name what you still need to find out.

What makes a skill combination useful?

Labor-market research: analytical and social skills can work together.

David J. Deming’s 2017 study looked at trends in United States jobs from 1980 to 2012. It compared workers across two national surveys of young people. Jobs and pay grew especially fast in work that needed both maths skills and social skills.

The combination matters. The finding points to work that needs the ability to analyse things as well as to deal with other people. It gives a reason to look at how skills add up together, instead of treating technical skills and people skills as rivals.

The study observed what happened. It didn’t set anything up to test, so it can’t single out automation as the cause. Its measures of social skill included signs of being sociable, rather than measuring trust directly.

Sun Hyung Kim and Dohyoung Kwon’s 2024 analysis shows why education matters to this finding. Using the same surveys, they looked at men aged 23–33 with different levels of education. The growing pay boost linked to social skills was mostly among workers who had been to college.

Why it matters: Together, these findings from past decades suggest that combinations of skills can matter, but what they bring depends on a person’s circumstances. They don’t show the same pay boost for everyone, or that it will last as AI changes work.

Read the study of social skills in the labor market. Read the analysis across education groups.

Randomized training study: the comparison changes the lesson.

Adam Osman and Jamin D. Speer’s 2025 study tested different mixes of training with college graduates in Cairo. Across several intakes, 985 people were put by chance into one of four groups. Training was mostly soft skills, mostly technical skills, or a mix of both, and one group got no training to compare against.

The training lasted four weeks, with 120 hours of teaching. About 18 months later, the researchers interviewed 82% of the people. Monthly income in the mixed-training group was 27% higher than in the group trained mostly in soft skills.

That compares two kinds of training with each other. Against the group with no training, the benefits of training over the longer term were small on average. So the finding backs paying attention to the mix being taught. It isn’t a promise that any course will raise your income.

Why it matters: The result is about these graduates in Cairo and the training they had. Not everyone was followed up, and people could take other training too. It gives a reason to think about useful combinations of skills, while keeping your own local work in view.

Read the experiment comparing training mixes.

Workplace rollout study: AI can narrow performance gaps.

Erik Brynjolfsson, Danielle Li and Lindsey Raymond’s 2025 study followed AI help as it was brought in, bit by bit, for 5,172 customer-support agents at a business-software company. With access, agents solved 15% more problems an hour on average.

Agents with less experience and less skill gained more. Quality didn’t improve for everyone: the most experienced and most skilled agents got slightly faster, but the quality of their work slipped slightly.

Here, AI help let workers with less experience catch up on how well they did the task.

Why it matters: Access came at different times at one company; agents weren’t each put into groups by chance. The study didn’t look at pay or at how many workers employers wanted overall. So smaller gaps in performance don’t show smaller gaps in pay.

Read the workplace study of AI assistance.

What do the skill studies really show?

Randomized developer experiment: help must survive the complete task.

A 2025 study by Joel Becker and colleagues at Model Evaluation & Threat Research looked at 16 experienced developers doing 246 real tasks in open-source software projects they knew well. Each task was picked by chance to allow or not allow the AI tools of early 2025.

Allowing those tools made the tasks take 19% longer. That was more time spent finishing the work, not a 19% fall in productivity. The result belongs to its setting: these developers, these established projects and the tools available from February to June 2025.

A February 2026 update from Becker and colleagues said their later experiment couldn’t reliably measure the effect on productivity now. Who took part, which tasks they put forward, changed pay and time spent on overlapping tasks done with AI all muddied the comparison.

The authors suspected bigger gains in speed, but the later data didn’t reliably show how big. The earlier slowdown is still a useful reminder not to assume AI help always helps. It isn’t a verdict on every developer tool today.

Why it matters: The support and developer studies show why you need to look at the work itself. AI help has to help you finish work that meets the standard it needs, counting the time spent checking and fixing it. A result somewhere else can give you a reason to compare. It can’t tell you what will happen in your task.

Read the developer experiment. Read the follow-up on measurement limits.

Find the part worth learning

What this means for you

Choose what to learn by what it adds to the whole result. The studies give reasons to combine skills and to test AI help. Because their results differ, a single ranking of skills for everyone isn’t much use.

Start with a task someone needs completed

Describe a real piece of work, who it’s for and what a good enough result would be. Break it into parts, so you can see where AI help might fit and where something is still unsettled.

In the made-up report example, drafting the text, checking the figures and explaining the recommendation are different parts. A draft that looks finished doesn’t tell you whether the figures are right, or whether the reader understands the decision.

For each part, ask what a tool might do. Could it do that part, help you do it, or leave something that still needs a person? You may not know yet.

Treat those answers as possibilities to check. Replacing means the tool does that part. Helping means you still do it, with help. Before you rely on either, check the real result and keep any checking the work needs. If you don’t have the evidence, leave the question open.

Look for a useful combination, then identify its weak point

Look at how what you know about your subject and how you work with others both feed into the same result. That could mean analysing a result and explaining it, or finishing your part and sorting out a problem when you hand it on.

Go back to the made-up report. If the calculation is wrong, presenting it more clearly won’t fix the mistake. If the calculation is right but the other team can’t use the recommendation, learning to explain the result may be the better target.

Your choice follows the weak point you’ve seen. You don’t need to split your time evenly between broad kinds of skill. Choose one combination that matters to the work. Then put your next learning time, at a cost you can afford, where things need to get better.

Count checking and correction as part of the work

Where your employer’s rules allow it, compare typical work done with and without AI help. Keep the standard the work has to meet in view, and include hard cases, not just easy examples.

Write down how long the work took, counting the time spent checking and fixing it. Note whether the result meets the standard, what needs redoing and which odd cases are still unsolved. Keep any checking the work needs, and never get round your employer’s rules to do this exercise.

If a comparison isn’t possible, use evidence from work you’ve already done. If that evidence is too thin, work out what you need to find out before you change how you spend your learning time.

When AI help does one part reliably, test whether getting better at another part would add more to the whole result. When it fails, the next thing to learn may be the subject knowledge involved, or how to deal with the unusual case.

Keep better work and better opportunities separate

Doing a task better tells you something about your work. To judge a big course or a change of career, you also need evidence about jobs you could actually get.

Check job openings, what they ask for and feedback from people who judge the work. Keep any training and qualifications you’re required to have. If a promising job isn’t one you could really get, deal with what’s in the way before you commit to the move.

Don’t turn doing a task better into a pay rise you assume will come. Keep track of how the work turns out, apart from what employers want, the offers you get and what you’re paid. That keeps your choice about learning tied to what you can actually see.

Review the choice when the evidence changes

Choose a stretch of learning you can afford and a point to look at it again. Look for a change in the weak point you found: a clearer explanation, fewer mistakes left unsolved or a smoother handover. In the report example, ask whether the reader can use your explanation, and whether the figures meet the standard needed.

Change your next stretch of learning when what you see in the work calls for it. Test AI help again after big changes to the tool or the task. If you can’t tell yet whether the learning helped, say so, instead of calling it a success.

Where would learning help your work most?

How about your daily life?

Where do you finish a task but still need to check, explain or sort something out with others? You might prepare reports, answer customers, review an analysis or hand work on to a colleague.

Which part is holding back a useful result? Where has AI help changed what you need to do yourself?

Write down situations worth looking at. One good idea is enough. Add more if that helps you choose.

A situation in my daily lifeWhat I’d like to improve
________________________________________
________________________________________

My chosen situation: ____________________

Choose work in your job now, or in a next job you could realistically get. Take that situation into the exercise below. Choosing it doesn’t yet tell you which skill needs work.

Choose your next learning target

Start with your chosen situation. If you came straight here, pick a task in your job now or in a next job you could realistically get. This exercise is about the learning you choose. Keep any training and qualifications you’re required to have.

Never get round your employer’s rules or skip the checking the work needs. Only compare work where you’re allowed to, using a real piece of work you’re allowed to use. Use evidence from work you’ve already done if an experiment isn’t possible.

  1. Describe the complete result. Name the task, who needs it and the standard it has to meet. List its parts, including any checking, explaining and handing on.
    Task, recipient and acceptable result: ____________________
    Components: ____________________
  2. Note what assistance might change. For each part, write whether a tool might do it, help you do it, or leave something that needs a person. Mark anything you don’t know yet. These are possibilities to check, not findings.
    Component, possible role of assistance and uncertainty: ____________________
  3. Examine representative work. Where you’re allowed to, compare work done with and without AI help. Include hard cases. Write down the time spent checking and fixing, whether the quality was good enough, what needed redoing and which odd cases are still unsolved.
    If a comparison isn’t possible, use evidence from work you’ve already done. If an important answer is missing, write down the information or feedback you need before you change how you spend your learning time.
    Evidence and remaining gaps: ____________________
  4. Choose the weakness that matters. If AI help does one part reliably, test whether getting better at another part adds more. If it fails, think about the subject knowledge, or the handling of unusual cases, that the failure points to.
    If working with others is the weak point, practise it alongside the technical work. Choose one stretch of learning you can afford, taking an amount of time that fits your life.
    Specific learning target and affordable commitment: ____________________
  5. Set a reassessment point. Name what you’d need to see in the work to keep or change the target. Test AI help again after big changes. If the result is unclear, keep it marked as unclear.
    Before a big course or a change of career, check job openings you could get, what they ask for and feedback from people who judge the work. Write down anything you’re missing before you commit.
    Reassessment point, evidence to inspect and access checks: ____________________

Ask Jing

Become a Jingius.

Find the weak point, then choose what to learn

Loading interactive tool…

Your next step

Finish with one clear learning target tied to a weak point you’ve seen, or a clearly named gap that stops you choosing yet. Keep the next step affordable, and keep doing better work apart from evidence about jobs or pay.

A note on the evidence

What skills paid in past decades, what training led to and how well AI does tasks are different questions. The studies of past decades observed what happened, and the training result comes from one group of people and one comparison of courses.

The AI findings depend on the workplace, the task and the tools studied. Use their differences to guide what you look at in your own work, and keep claims about pay and job openings apart.

How we prepared this module

This module links research on jobs, a training study with groups picked by chance and AI workplace studies that point different ways, including a later update on how to measure. A made-up report example works through the learning choice before the written look at your tasks applies it to your situation.

Sources