AI Readiness & Strategy
The AI-Augmented Workforce: Why Adaptability Is the Ultimate Currency
AI takes tasks, not whole jobs. What the 2025 evidence says about upskilling, where AI helps and hurts, and how to build an AI-augmented workforce.
An AI-augmented workforce hands generative AI the tasks it does reliably and spends the time saved on judgement, research and harder problems. Adaptability is the currency because the World Economic Forum's Future of Jobs Report 2025 expects 39% of workers' core skills to change by 2030, and AI helps inside its frontier but misleads outside it (Dell'Acqua et al., 2023).
For most of the history of work, a professional skill meant knowing how to operate something: first a lathe or a ledger, later a spreadsheet or a reporting tool. You learned the tool, and it did what you told it. Generative AI, the class of models that produce text, code and images from an instruction, breaks that pattern. It interprets what you ask and hands back something you didn't type, whether that's a draft, a summary, a working block of code or an answer that may or may not be right.
That changes what a skill is. We now have to train people to work alongside a system that can do part of their job, sometimes better than they can and sometimes confidently wrong. Microsoft and LinkedIn's 2024 Work Trend Index, a survey of 31,000 people in 31 countries, found that 75% of knowledge workers already use AI at work, and that 78% of them bring their own tools instead of waiting for the company to provide one.
Failure mode
The gap is already open
In Microsoft and LinkedIn's 2024 Work Trend Index, two-thirds of leaders said they wouldn't hire someone without AI skills, yet only 39% of AI users had received any AI training from their employer. People are learning in the dark, on tools nobody chose.
The usual response is a fear story in which AI comes for the jobs. The evidence so far points somewhere more useful.
Tasks are what's moving to AI, and the people who do best with it know which ones to hand over and how to check what comes back. They also have a use in mind for the hours they get back. The rest of this piece is about making that work, and what a leadership team should do about it in the next two quarters.
What an AI-Augmented Workforce Actually Means
The phrase gets used loosely. People apply it to any team with a chatbot licence, and also to automation programmes that quietly remove roles under a friendlier name. Neither of those is what it means.
An AI-augmented workforce is one in which generative AI performs a defined share of each role's tasks, people stay accountable for the outcome, and the time released is deliberately redirected to work the AI cannot do: framing the problem, checking the answer, deciding, and researching the next harder question. The word that matters is deliberately. If augmentation happens by accident, one person at a time, and nobody changes what the role is for, what you get is unmanaged drift.
Because the unit of change is the task, you need to know which tasks each role contains. Accountability stays with the person, so every role needs the skill of checking AI output as well as producing it. And a productivity gain only counts if it's reinvested; otherwise the same inbox absorbs it at the same pace. We call the first two the two fluencies and the third the time shift, and together they're what adaptability means for a professional now: re-sorting your own work, task by task, as the frontier of what AI does reliably moves.
What the Evidence Says About AI and Jobs
What this means for you: the serious studies agree on direction, so the open question for your organisation is how to augment work with AI without losing the judgement that makes the output trustworthy.
The headline numbers are better than the headlines. The World Economic Forum's Future of Jobs Report 2025 draws on more than 1,000 employers representing over 14 million workers across 55 economies, and it expects 170 million jobs to be created and 92 million displaced by 2030, a net gain of 78 million. Skills are the harder read. The same report expects 39% of workers' core skills to change in that period, and estimates that 59 of every 100 workers will need training by 2030: 29 upskilled in their current role, 19 reskilled and moved to a new one, and 11 who need it but are unlikely to get it.
Field experiments find gains that are real but uneven. Brynjolfsson, Li and Raymond's study of 5,179 customer support agents was published as a National Bureau of Economic Research (NBER) working paper in 2023 and in the Quarterly Journal of Economics in 2025. In the 2023 paper, Brynjolfsson, Li and Raymond found that access to a generative AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers, with minimal effect on the most experienced. In effect, the AI spread the habits of the best agents to the newest ones.
| Source | Year | What it measured | Finding | What we take from it |
|---|---|---|---|---|
| World Economic Forum, Future of Jobs Report | 2025 | 1,000+ employers, 14 million workers, 55 economies | 170m jobs created, 92m displaced by 2030; 39% of core skills change; 59 in 100 workers need training | Budget to train every role; a pilot group won't cover it |
| Brynjolfsson, Li and Raymond, National Bureau of Economic Research (NBER) / QJE | 2023 / 2025 | 5,179 customer support agents, staggered AI rollout | 14% more issues resolved per hour; 34% for novices; little change for experts | Point AI at onboarding and the long tail of newer staff first |
| Dell'Acqua et al., Harvard Business School and BCG | 2023 | 758 consultants, 18 realistic tasks | 12.2% more tasks, 25.1% faster, higher quality inside AI's frontier; 19 percentage points less likely to be correct on a task outside it | Map which of your tasks sit inside the frontier before you delegate |
| Microsoft and LinkedIn, Work Trend Index | 2024 | 31,000 people, 31 countries | 75% of knowledge workers use AI; 78% bring their own tools; 39% have had employer training | Replace shadow AI (tools staff adopt unapproved) with sanctioned tools and training |
| Anthropic Economic Index | 2025 | About one million real Claude conversations | 57% of use augments a person's work, 43% automates a task; 36% of occupations use AI in at least a quarter of their tasks | Plan for collaboration as the dominant pattern |
| PwC, Global AI Jobs Barometer | 2025 | Close to a billion job adverts, company financials | 56% wage premium for AI skills, up from 25%; skills changing 66% faster in AI-exposed occupations | Treat AI fluency as a pay-grade skill, and plan for faster skill churn |
Usage data points the same way. Anthropic's Economic Index, published in February 2025 and drawn from about one million Claude.ai conversations, classified 57% of AI use as augmentation, where the person and the model iterate together. Anthropic classified the other 43% as automation, where the model completes the task with minimal human involvement.
So far AI is taking over tasks, and the job is turning into the part that decides what those tasks are for.
The labour market is already pricing this in. PwC's 2025 Global AI Jobs Barometer drew on close to a billion job adverts and found a 56% wage premium for roles requiring AI skills, up from 25% the year before. PwC itself notes that the premium is partly scarcity. We'd argue that's the point: people who can work with these systems are in short supply, and the gap is widening faster than training budgets are moving.
Where AI Helps and Where It Hurts: The Jagged Frontier
What this means for you: the tool that speeds a team up on one task makes it confidently wrong on the next, and how hard the task looks won't tell you which is which.
If you're planning an AI-augmented workforce, the single most important study to read is Dell'Acqua and colleagues' 2023 field experiment with 758 Boston Consulting Group consultants, published as Harvard Business School Working Paper 24-013. Its 18 realistic consulting tasks sat inside what the authors call the jagged technological frontier of AI capability. On those, Dell'Acqua and colleagues found that consultants with GPT-4 completed 12.2% more tasks and finished them 25.1% faster, and their work was of significantly higher quality.
Then the authors gave the same consultants a task chosen to sit just outside that frontier, a management decision that needed a careful read of interview notes and financial data together. Consultants using AI were 19 percentage points less likely to reach the correct recommendation than those without it (Dell'Acqua et al., 2023). They got there faster and felt more confident about it. They were also wrong more often.
Why the frontier is jagged, not a line
The frontier is jagged because AI capability doesn't track human difficulty. Drafting a strategy memo from a brief feels hard to a person, yet it can sit comfortably inside the frontier. Reconciling two sources that quietly disagree feels routine, and it can sit outside, because the model fills the gap with something plausible and doesn't flag it.
So "use your judgement about when to use AI" doesn't work as a policy. Judgement about the frontier has to be built from evidence, task by task, and then taught. Remember who got it wrong in the BCG study: experienced professionals who trusted a fluent answer, which is precisely the trust fluent answers are designed to earn.
The delegation test
In our engagements we sort tasks with three questions before anyone is told to hand them to an AI. We call it the delegation test. It's deliberately simple, so a team can run it on a whiteboard.
- Is the task inside the frontier? Has this kind of task been tried on your data, with the output checked by someone qualified, and did it hold up? A guess counts as a no.
- Can the person check the output? Is there an independent source, a test, a reference figure or a second reader that would catch a plausible wrong answer? If the only check is "it reads well", that's a no.
- Where does the saved time go? Is a named, harder piece of work waiting for the hours released? If the answer is "more of the same, faster", the gain will evaporate.
Readiness signal
You have found the frontier when
For each task a role performs, you can show from evidence whether AI output on it has been checked and held up. Until you can say that about a task, treat it as outside the frontier.
Tasks that pass all three are delegated with a checking step, and a task that fails the first stays with the person. The case worth slowing down for is a task that fails the second. It gets a check designed before it's delegated, and designing that check is often the most valuable work in the whole programme.
The Two Fluencies Every Professional Needs Now
What this means for you: AI fluency without sharper judgement produces fast, confident errors, and judgement without AI fluency leaves the gains on the table, so both have to be taught to everyone, well beyond a centre of excellence.
It's tempting to call prompt engineering the mandatory skill, but prompting is a mechanical layer that changes with every model release. What lasts is knowing what to ask a system to do and recognising when its answer is wrong, then deciding what to do instead. We describe this as two fluencies, and every role in the organisation needs a working level of both.
AI fluency
Knowing what current models do reliably, how to give them context and how to keep iterating past the first answer. Like spreadsheet literacy before it, this is a general skill for everyone.
Judgement fluency
Knowing how to check an output against something independent, when a fluent answer is hiding a gap, and which decisions stay with a person whatever the model says. It's the half most training programmes skip.
The time shift
Alongside the two fluencies sits a third requirement: deliberately moving the hours AI releases into research, problem framing and the harder questions that were never reached. Skip it and the gain gets absorbed into more of the same work.
AI fluency: the collaborator, not the tool
The clearest picture of what AI fluency does at scale is the Brynjolfsson, Li and Raymond result. The assistant in their study sat alongside the agents and replaced none of them. It watched the best ones and suggested their habits to the rest, and novices improved by 34% while experts barely moved. A good collaborator works like that, carrying the organisation's accumulated practice to whoever needs it next.
Day to day, AI fluency means a professional can describe a task well enough for a model to attempt it and can supply the context the model lacks. They can tell a first draft from a finished answer, and they keep iterating. It also means knowing which tools the organisation has sanctioned, which is why the choice of first generative AI use cases matters so much: people learn fluency on the tools in front of them.
Judgement fluency: the amplified human
The second fluency is what separates the BCG consultants who were helped from the ones who were misled. Judgement fluency is the trained habit of asking where an answer came from and what would be true if it were wrong, then looking for the independent evidence that would settle it. The World Economic Forum's 2025 report ranks analytical thinking as the core skill employers value most, ahead of any technical skill. In our experience, that ranking is exactly right.
The model makes the first draft cheap, which makes the person who can tell a good draft from a wrong one expensive.
Regulation now assumes this fluency too. Article 4 of the EU AI Act, the AI literacy obligation in the EU's regulation on AI systems (Regulation (EU) 2024/1689), has applied since 2 February 2025. It requires providers and deployers to take measures to ensure, to their best extent, a sufficient level of AI literacy among the staff who operate or use AI systems, taking account of the context in which the systems are used.
The UK has no equivalent duty, but Articles 22A to 22D of UK GDPR, in force since 5 February 2026, set the safeguards for decisions based solely on automated processing that have legal or similarly significant effects, including the right to human intervention. Both regimes assume a workforce that can tell when a system has reached the edge of what it should decide. That's why we treat the engineering view of AI governance and the people view in this article as the same control seen from two sides.
The time shift: where the hours go
What AI gives an organisation is time, and what the organisation does with that time is the strategy. In our experience the teams that benefit are the ones that name the waiting work in advance.
Think of the root-cause analysis nobody had a week for, or the data quality problem everyone works around. Then there's the customer question that keeps coming back and never gets researched. Those are the bigger problems, and they're the reason to take routine work off people in the first place.
The time shift is also where the sequencing of a data and AI roadmap meets the sequencing of a workforce programme. The hours released in the first quarter should go to the data foundation work that the second quarter's use cases depend on. Cut a workforce programme loose from a delivery roadmap and you end up with trained people and nothing changed for them to work on.
Why Upskilling Is an Enterprise Mandate, Not a Perk
What this means for you: you're already paying for not training people, through shadow tools and unchecked output today and through a wage premium at the next hiring round.
The Microsoft and LinkedIn numbers describe organisations that have adopted AI without ever deciding to. Three-quarters of knowledge workers use it and more than three-quarters of those bring their own tools, yet fewer than four in ten AI users have been trained by their employer. This is the present state of most enterprises we meet. In our engagements it means confidential data is reaching consumer tools and unchecked output is reaching decisions, with no record of either.
The World Economic Forum's estimate that 59 in 100 workers need training by 2030 gives the scale of the task. Its breakdown reads like a plan: 29 upskilled in their current role, 19 reskilled and moved, and 11 who need training and are unlikely to receive it. That last group is the one a leadership team can change. Upskilling everyone means every role, including the ones that look furthest from AI, because the frontier reaches them too.
Key idea
Treat training as a control
An AI literacy programme is the control that makes every other AI control work. A policy on checking output or a human-in-the-loop requirement assumes people who know what to look for. So fund it as governance, from the same budget as the controls it supports.
Cost matters here as much as risk. PwC's 56% wage premium for AI skills is what you pay to hire fluency from outside. Teaching it inside is cheaper and keeps the domain knowledge the organisation already has.
It also puts the gain where Brynjolfsson, Li and Raymond found AI helped most. In their study, AI assistance lifted the people with the most room to grow by the most, and AI-assisted agents stayed in their jobs longer. No training budget line usually gets credit for that retention effect.
The organisations getting this right run it every year, the way they run security awareness, and tie it to the tools and tasks people actually have. The judgement half gets as much time as the mechanics. The operating model for governing generative AI names the roles that own this, and the training is how those roles reach everyone else.
When Augmentation Is the Wrong Answer
What this means for you: some tasks should be automated outright with a downstream check and some left to people, and treating everything as a collaboration wastes the first kind and endangers the second.
Augmentation is the dominant pattern, but it has limits. Within Anthropic's 43% automation share there is real work where, in our experience, a person adds nothing by being in the loop. Reformatting and classification at volume are examples, and so is first-pass extraction from documents with a downstream check; insisting on a human for each of those adds cost without adding control. Those tasks belong in agentic systems with the approval steps designed in, and the point at which an agent is trusted to act is a separate decision from how you augment a professional.
At the other end, some decisions shouldn't be augmented at all until the frontier moves. The BCG study's outside-frontier task is the template: a decision that depends on reconciling sources that disagree, where a plausible synthesis is worse than none. Prudence puts decisions with legal or similarly significant effects on individuals here too, and where such a decision is solely automated, UK GDPR Articles 22A to 22D require safeguards. For these, use AI to prepare the evidence and leave the recommendation to a person.
In practice
In practice: start with the tasks nobody will miss
When a team is anxious about AI, we start the delegation test with the tasks they most resent, like the weekly status report or the reformatting. The first gain lands on work nobody wanted, and people learn the checking habit where the stakes are low. By the time harder tasks come up for discussion, the team has already seen the tool be wrong.
Augmentation is also the wrong first move when the data underneath can't be trusted. Ask a model to summarise reports built on inconsistent definitions and you'll get fluent summaries of inconsistent numbers.
The data governance that makes the numbers agree has to come first. That's why we argue that a transformation that works fixes the data foundation first, and it's why the first quarter of an AI workforce programme often looks like a data programme.
What to Do in the Next Two Quarters
What this means for you: a workforce programme that produces trained people and a changed set of tasks in six months beats a three-year academy plan that produces a curriculum.
Start with one team and one week: list the tasks the team actually performs, in the words they use, and run the delegation test on each one. In our engagements this takes a day and produces three lists: tasks AI can take now with a check, tasks that need a check designed first, and tasks that stay with people. The third list is usually shorter than the team expected. The second is where the interesting work is.
Then make the time shift explicit. For every hour the first list releases, write down the work it funds, whether that's a research question or a data quality fix. A gain with no destination is absorbed within a month, and the team concludes, correctly, that nothing changed.
Sort and sanction
Run the delegation test on one team and replace its shadow tools with sanctioned ones. Then train the two fluencies on the tasks that team actually has, giving the judgement half equal time.
Design the checks
Define the independent check for every delegated task, such as a test, a reference figure or a second reader, and record who signs off. It's the control the regulators and your auditors will ask to see.
Shift the time
Point the released hours at the named harder problems, starting with the data foundation the next use cases depend on. The measure is what got researched and fixed. Hours saved won't tell you that.
Extend by evidence
Take what held up to the next team, and promote a task across the frontier only when its checked output has earned it. Log the ones that failed; the frontier moves, and you'll revisit them.
On timing, with scope attached: getting one team sorted, sanctioned and trained on the two fluencies takes weeks. Extending that across a function, with checks designed and a delivery roadmap to absorb the released time, is multi-quarter work, and it should run alongside the AI readiness assessment that decides which use cases the organisation funds next.
The organisations that come through this well will be the ones whose people have that adaptability: they can tell, task by task, what to hand over and how to check it, and which harder problem to spend the afternoon on instead. If you want help running the first sort, our digital transformation and agentic AI teams do it as the opening week of an engagement, and you come away with the three lists above and the checks designed.
Frequently Asked Questions
What is an AI-augmented workforce?
An AI-augmented workforce is one in which generative AI performs a defined share of each role's tasks, people stay accountable for the outcome, and the time released is deliberately redirected to judgement, research and harder problems. Change happens one task at a time, inside each role. Anthropic's 2025 Economic Index classified 57% of the AI use it observed as augmentation and 43% as automation, where the model completes the task with minimal human involvement.
Will AI take my job or make me more productive?
The evidence to date points to tasks within jobs moving to AI. The World Economic Forum's 2025 Future of Jobs Report expects 170 million jobs created and 92 million displaced by 2030, a net gain of 78 million. Brynjolfsson, Li and Raymond's 2023 study of 5,179 support agents found an AI assistant raised productivity 14% on average and 34% for novices. In our experience the gains go to people who know what to delegate and how to check it.
What skills do professionals need to work with AI?
Two fluencies. AI fluency is knowing what current models do reliably, how to give them context and how to keep iterating past a first answer. Judgement fluency is knowing how to check an output against an independent source, when a fluent answer hides a gap, and which decisions stay with a person. The World Economic Forum's 2025 report ranks analytical thinking as the skill employers value most, ahead of any technical skill.
Why does AI make some work worse rather than better?
Because AI capability has what Dell'Acqua and colleagues call a jagged technological frontier: it does not follow human difficulty. In their 2023 experiment with 758 BCG consultants, AI users completed 12.2% more tasks 25.1% faster inside the frontier, but were 19 percentage points less likely to be correct on a management decision just outside it. The model fills gaps with plausible text and doesn't flag them, so tasks that reconcile disagreeing sources need a person or an independent check.
Is AI literacy training a legal requirement?
In the EU, in effect yes for organisations in scope. Article 4 of the EU AI Act, applicable since 2 February 2025, requires providers and deployers to take measures to ensure, to their best extent, a sufficient level of AI literacy among staff who use AI systems. The UK has no equivalent duty, but UK GDPR Articles 22A to 22D, in force since 5 February 2026, require safeguards, including human intervention, for solely automated decisions with significant effects.
Unolabs is a Data and AI first engineering consultancy, headquartered in the United Kingdom with engineering operations in Pune and active engagements across the UK, Australia, and Hong Kong. We help enterprises build the architectural foundation for autonomous AI execution — governed data platforms, semantic intelligence, and agentic systems that enterprises can stand behind.
If you are weighing how to prepare your teams to work alongside AI rather than around it, book a discovery call and we will work through it with you.
More Where
This Came From.
New architectural deep-dives land every two weeks. Pick your channels and we will send them as they publish.
Continue reading
- AI Readiness & StrategyEnterprise Generative AI Use Cases That Reach Production, Grouped by What Each One NeedsEnterprise generative AI use cases grouped by what each needs to reach production, where each one fails, and its likely EU AI Act risk tier.17 min read
- Data Governance Trends 2026The EU AI Act and Your GenAI Operating Model: A Compliance-Ready BlueprintWhat the EU AI Act asks of GenAI deployers, and the operating model that satisfies it: named roles, four lifecycle gates, and evidence at each step.14 min read
- Data Platform StrategyDo 70% of Digital Transformations Really Fail?No reliable study shows 70% of digital transformations fail. What BCG (2020) and McKinsey (2018) actually measured, and the six-month test to apply.13 min read
Find out where your teams should delegate first
Bring one team's week of work and we'll run the delegation test on it with you. You'll see which tasks AI can take today and where the saved time should go, so your people get more capable along with the tools.
Book a Workforce AI Review