How to Stay Valuable at Work as AI Gets Better

Why "learn AI" is the wrong goal

Every list of advice for the age of AI says to learn AI. It is true, and it is not enough, for three reasons the evidence makes plain.

First, basic fluency is becoming table stakes rather than an edge. Employers say AI skills matter: 74 percent of United States employers in ZipRecruiter's 2026 survey saw them as a strong advantage or requirement, and half expected candidates to be practical or advanced users already.¹⁷ Microsoft's 2024 survey of 31,000 knowledge workers found 66 percent of leaders saying they would not hire someone without AI skills, and 71 percent preferring a less experienced candidate with AI skills over a more experienced one without.³ Those are company surveys, but the direction is consistent, and employers surveyed by the World Economic Forum expect 39 percent of workers' existing skills to be transformed or become outdated by 2030.¹ When everyone is expected to have a skill, having it stops distinguishing you.

Second, "AI skills" mostly does not mean technical skills. The OECD's 2026 review of the evidence found that fewer than 1 percent of workers need advanced AI skills, while most need digital skills and the ability to use, analyse and interpret data, alongside managerial and human skills such as problem-solving and creativity.¹⁵ Lightcast's analysis of 1.3 billion job postings found 51 percent of postings requiring AI skills were outside IT and computer science.² For most people, fluency means using and checking, not building.

Third, claimed fluency is nearly as good as certified fluency, which means neither is a durable advantage. In a 2026 experiment where 1,725 recruiters in the United Kingdom, United States and Germany rated synthetic CVs, listing AI skills raised the chance of an interview invitation by around 8 to 15 percentage points, but formal certificates added only a moderate amount over simply declaring the skill.¹⁶ If employers cannot yet tell real fluency from claimed fluency, the person who can demonstrate results has an advantage that a course completion does not provide.

What the evidence says about who benefits and who does not

The research on AI and individual workers contains a paradox that most career advice ignores. AI helps the least experienced people most, and it is hurting the least experienced people most.

AI narrows skill gaps on tasks

• When around 5,000 customer support agents were given an AI assistant, productivity rose 14 percent on average, but 34 percent for the newest and least skilled agents and hardly at all for the most experienced. The tool spread the practices of the best agents to the rest.⁶
• In a randomised trial with 776 professionals at Procter and Gamble working on real product problems, individuals working with AI matched the performance of two-person teams working without it, and the usual gap between technical and commercial staff disappeared.⁵
• In the Boston Consulting Group experiment with 758 consultants, those given GPT-4 on tasks inside its competence produced work rated around 40 percent higher in quality, and the bottom-half performers gained most.⁴

And yet the youngest workers in exposed jobs are losing ground

Using payroll data on millions of United States workers through June 2026, Stanford researchers found that employment of 22 to 25 year olds in the most AI-exposed occupations was 19 percent below where it would have been had it kept pace with less exposed peers, while experienced workers in the same occupations showed no gap. The gap is relative to less exposed peers, it came through reduced hiring rather than dismissals, and it concentrated in occupations where AI substitutes for tasks rather than complementing workers.⁷ Our guide on whether AI will replace your job covers what that study does and does not show.

The resolution of the paradox is uncomfortable but useful. AI makes a novice better at the tasks novices used to be hired to do, and it also makes those tasks cheaper to do without a novice. What it has not yet replaced is what experienced workers bring: knowing what the task is for, recognising when the output is wrong, and being accountable for the result. PwC's 2026 analysis of over a billion job advertisements found that junior roles exposed to AI were seven times more likely than other junior roles to demand traditionally senior capabilities such as leadership and strategic thinking.¹⁹ The bar for entry has moved up to where experience used to sit. The International Monetary Fund's 2026 analysis of job vacancies in advanced economies found the same tension: postings demanding AI skills offer higher wages, but the spread of those skills is linked to lower employment in occupations with high exposure and low complementarity to AI, a challenge it singles out for young people.²⁰

Three ways fluent-looking AI use goes wrong

Before the framework, the failure modes, because each one is a way of using AI a great deal while becoming less valuable.

Unchecked output

In the 2025 KPMG and University of Melbourne study of over 48,000 people in 47 countries, 66 percent of employees said they had relied on AI output without evaluating its accuracy, and 56 percent had made mistakes in their work because of AI.¹⁰ A Microsoft Research and Carnegie Mellon survey of 319 knowledge workers found that higher confidence in the AI was associated with less critical thinking, while higher confidence in one's own ability was associated with more, and that the effort of knowledge work was shifting toward verifying and integrating output rather than producing it.¹¹

Workslop

Research affiliated with BetterUp and the Stanford Social Media Lab surveyed United States desk workers in 2025 and found 40 percent had received AI-generated work in the previous month that looked polished but lacked the substance to move the task forward. Recipients spent nearly two hours dealing with each instance, and 42 percent saw the sender as less trustworthy.⁹ This is company-affiliated research, but its lesson is hard to argue with: passing along unreviewed output spends your reputation to save your time.

Deskilling

A multicentre observational study of 19 experienced endoscopists in Poland found that after routine use of an AI detection tool was introduced, their detection rate in colonoscopies performed without the tool fell from 28.4 percent to 22.4 percent. The authors note other factors may have contributed, and it is one study in one field, but it is the first field evidence of a skill declining after regular AI assistance.¹² Anthropic's January 2026 analysis of its own usage data found that when the tasks its model could perform were removed from an occupation, the tasks left for the human often had lower educational requirements, a pattern it described as net deskilling.⁸ And a small, not yet peer-reviewed study of 54 students found those who wrote with a language model showed the weakest brain connectivity on EEG measures and struggled to quote their own essays.¹³ None of this means you should not use AI. It means using it in a way that keeps your own skill alive is part of the job.

The Stay Useful framework: Fluency, Leverage, Proof

This is Stay Useful's interpretation of the evidence above, organised so you can act on it. Each capability is defined by what an observer would see you doing, not by what you know.

AI Fluency, as behaviour

Fluency is the capability most advice stops at, and the one the evidence says is necessary but not sufficient. Here is what it looks like when done well.

• You brief, you do not prompt. Purpose, audience, inputs, output format, constraints, and what to do when uncertain. Our beginner's guide to using AI at work shows the anatomy of a good brief.
• You iterate. Anthropic's usage data shows experienced users are more likely to work in patterns of iteration and validation, and have a higher success rate, than new users who delegate and accept.⁸
• You know your own frontier. In the consulting experiment, the same tool lifted quality by around 40 percent on tasks inside its competence and left people performing 23 percent worse on a deliberately difficult one.⁴ Fluent users have tested the tool on their own tasks and know which side of the line each one is on.
• You verify against a source you open. Not by asking the tool to check itself. Every figure, name, date and citation.
• You know when not to delegate. Work where your own thinking is the product, or where you are building a skill you will need, you do first and use the tool second. The deskilling evidence is the reason.¹²
• You do not confuse fluency with knowledge. The survey evidence says confidence in the tool lowers scrutiny.¹¹ Fluent users treat their own domain knowledge as the check on the tool, not the other way around.

AI Leverage: redesigning work rather than chatting

The gap between "40 percent faster on this task" in experiments and "2 to 3 percent of my week" in population surveys is the gap between using AI on tasks and redesigning work around it. Most people are on the wrong side of that gap, which is why being on the right side is valuable.

Leverage is also where the automation-or-augmentation distinction lives. The Stanford payroll study found employment declines concentrated in occupations where AI usage substitutes for human tasks, and flat or rising employment where it complements workers.⁷ The person who redesigns a workflow decides which of those it does. That is a position of influence over your own exposure that a user of prompts does not have. The economist David Autor argues that AI's distinctive opportunity is to extend the reach and value of human expertise, letting more people do higher-stakes work well; whether that happens in your workplace depends on who designs the work.¹⁴

What redesigning work involves

1. Problem selection. Choose a piece of work that matters to someone: it is slow, error-prone, or nobody has time for it. Not the task that is easiest to automate.
2. Mapping. Write down the steps as they actually happen, including the checks people do informally and the exceptions.
3. Allocation. Decide which steps the tool does, which a person does, and where a person must confirm before anything proceeds. High-stakes actions stay behind a human.
4. Safeguards. What information may enter the tool, and what may not. What gets verified and how. What happens when the tool is uncertain.
5. Repeatability. Write the instruction and the checks so that a colleague could run the workflow and get the same result. If only you can run it, it is a habit, not a workflow.
6. Ownership. You are accountable for its output, its failures and its improvement. Ownership is what makes it yours in a review conversation.

Our guide to 25 ways to use AI to make your job easier gives each use case in the shape of task, AI role, human check and proof, so you can start from any one of them.

Proof: turning improvement into evidence

This is the part of the framework with the least direct research behind it, and we want to be honest about that. No credible study yet shows managers specifically rewarding documented AI impact. What the evidence does show is a labour market that cannot yet tell real fluency from claimed fluency,¹⁶ that is raising productivity expectations for everyone,¹⁷ that judges the quality of what you send harshly,⁹ and that rewards judgment and experience over tool use alone.⁷¹⁹ In that market, a record of what you improved, how you measured it and what you checked is a rational hedge, and the only way to make your fluency and leverage visible to anyone who was not in the room.

What a piece of proof contains

Keep these short, dated, and in one place. Three of them, built over three months, are a stronger answer to "how do you use AI?" than any list of tools, in a performance review or an interview.

The other things that matter, defined as behaviour

The generic advice is not wrong. It is unfinished. Here is what each piece means as something you could do this week.

A note on New Zealand

New Zealand's position makes this framework more valuable, not less. In the KPMG and University of Melbourne study, New Zealanders reported among the lowest levels of AI training of the countries surveyed: 24 percent had undertaken any AI-related training against 39 percent globally, and only 36 percent believed they had the skills to use AI appropriately, against 60 percent globally.¹⁰ At the same time, a recruiter survey of 500 New Zealand office workers found 87 percent believed developing generative AI skills was necessary for career success.¹⁸ PwC's New Zealand analysis of job advertisements found postings asking for AI skills rose from about 3,900 in 2024 to about 9,600 in 2025, around 2.7 percent of all postings, with most of the growth in roles that use AI rather than build it, and wage premiums for AI skills that remain low across most sectors.¹⁹

Read together: demand is rising, the skills are scarce, and the market is not yet paying much extra for them. That is the situation in which someone who can demonstrate fluency, leverage and proof stands out most, because there are few of them and the evidence is easy to check.

A 30-day plan

1. Days 1 to 7: fluency. Pick three tasks you do weekly. Brief the tool properly on each. Verify every specific. Write one line per task on where it helped and where it failed. This is your personal map of the frontier.
2. Days 8 to 14: choose a problem. Find one piece of work that is slow or error-prone and matters to someone other than you. Map its steps as they actually happen, including the informal checks.
3. Days 15 to 24: redesign. Allocate steps between the tool and a person. Write the safeguards. Build the instruction. Run it three times, fixing what broke. Have a colleague run it once from your instructions.
4. Days 25 to 30: proof. Write the record: problem, baseline, change, result, checks, limits. Share it with your manager, not as a boast but as a proposal for what to do next. Then pick the next problem.

At the end of the month, check the behaviour rather than the calendar. Our AI skills checklist of seven things to check at work is the short version of the three ideas above, written as questions you answer about one real piece of work.

If you are early in your career, do the fluency week twice and put extra weight on the judgment column: do the task yourself first, then compare with the tool, and write down what you would have missed. The hiring data says the bottom rung is where the pressure is, and that judgment is what the rung above requires.⁷¹⁹ Our guide on whether AI will replace your job explains why.

Related questions