What Does It Mean to Be Experienced in the Age of AI?
Artificial intelligence is changing what we can do. We need to give equal attention to how it changes what we learn, how we develop professionally, and how we recognise human capability.
On 16 October 2026, Anthropic published research into how work was changing inside the company. Among its findings, the share of junior technical hires had fallen from 27% to 9% since early 2025. Hiring managers were placing greater emphasis on candidates’ autonomy and judgement. [1]
One company cannot represent an entire economy. A smaller share of hires does not, by itself, establish a decline in their absolute number either. But the two observations together bring me to a question I consider important: how does someone acquire the judgement and independence we expect when the way they learn their work is changing?
My first formulation was: “If AI does the junior’s work, who will train the next senior?” As I thought it through, I realised that it captured only part of the issue. The experience paradox existed long before artificial intelligence. Employers wanted experienced candidates, while young people wondered where they were supposed to acquire the experience required for their first job.
Today, we need to examine a broader question: what does the word “experienced” mean when part of the work is performed alongside AI systems or delegated to them?
I have spent more than thirty years building technology. I find the new possibilities compelling: the speed at which we can test ideas, and the access to assistance that used to be difficult to obtain. Yet much of our discussion focuses on what the tools can do and what we can produce with their help. I would like us to spend more time considering the person using them: how that person develops, what they understand, and which capabilities they acquire through the collaboration.
What experience contained
When we described someone as experienced, we usually meant more than the knowledge they possessed. They had worked under real conditions. They knew the processes, rules, exceptions, recurring cycles, and particular characteristics of people and organisations. They had encountered unclear requirements and problems that could not be resolved by following a manual.
Experience could involve remembering a procedure, recognising a recurring pattern, mastering a physical movement, understanding a difficult case, or finding a way to work with someone who saw the same problem differently. The combination varied from one occupation to another.
Through that involvement, people gradually built an internal picture of their work. They began to recognise where attention was needed, what to expect, and which apparently minor deviation deserved investigation. Sometimes they sensed that something was wrong before they could fully explain why.
Years of service acted as an indication that someone had been exposed to enough of these situations. It was always an imperfect indication. Two people could have spent the same length of time working and developed very different depths of understanding. Repeating a process could produce familiarity without necessarily extending judgement. Equally, a demanding period with good guidance could provide more learning than years of stagnation.
The problem of assessing experience therefore predates AI. What AI invites us to reconsider is the relationship between time spent working, personal effort, and the result delivered.
As work shifts, learning shifts with it
In software development, a worker can now delegate parts of code writing, debugging, or test execution to one or more agents. Here, an agent means an AI system that can undertake a sequence of actions using tools to pursue a goal, within the limits it has been given.
Human involvement moves to different points in the process: defining the problem, supplying context, organising the work, assessing proposals, and intervening when the direction needs to change. The extent of that shift depends on the task, the tool, and the autonomy it is allowed.
This activity can generate experience of its own. You learn to give better instructions, recognise when important information is missing, coordinate parallel efforts, and decide what requires scrutiny. You may also encounter more approaches and attempt projects that previously exceeded your capabilities.
The question is how much you understand about the subject you are managing. Fluency in directing an agent and knowledge of the work delegated to it are related, but they are not interchangeable. Their combination can teach you a great deal, provided you can examine the intermediate steps, the decisions, and their real consequences.
Much of the work also remains outside the agent’s environment: informal agreements, a need that was poorly expressed, a client relationship, a practical difficulty in the field. Professional development requires contact with these dimensions too.
In occupations involving substantial physical presence, equipment handling, or direct human care, the change will take a different form. It would be premature to assume that every occupation will follow the same path. We can, however, already examine how learning changes where AI plays a substantial part in the working day.
When the result improves, what has the person learned?
For me, this is a crucial distinction. Better performance with a tool may reflect personal learning, the assistance supplied by the tool itself, or a combination of the two. We need to understand which is happening, especially when we use that performance to assess someone, promote them, or give them greater responsibility.
A small randomised study presented by Anthropic on 29 January 2026 illustrates the difficulty. Fifty-two developers, mostly juniors, worked with a Python library unfamiliar to them. In the subsequent comprehension assessment, the AI-assisted group averaged approximately 50%, compared with 67% for the group without AI. The difference in completion time was not statistically significant. [2]
The study was brief and does not describe years of professional development. The researchers also observed different patterns of use: actively seeking explanations was associated with better learning than passive delegation. That comparison between usage patterns does not establish causation, but it deserves attention.
There is evidence pointing in another direction too. “Generative AI at Work”, published in The Quarterly Journal of Economics in 2025, examined the introduction of an AI assistant among 5,172 customer-support workers. It found an average productivity increase of 15%, with larger benefits for less experienced workers and evidence of learning. [3]
These are different tools, tasks, and settings. The findings do not provide a universal rule about AI. They help us ask a more precise question: which forms of collaboration make people more capable, and which mainly improve the immediate result?
Independence does not require us to work without tools all the time. It does require us to assess what they provide, recognise their limits, and adapt when conditions change.
Will professional grades still mean the same thing?
Labels such as junior, mid-level, and senior serve as shorthand for professional maturity. In practice, they may combine years of service, technical knowledge, independence, responsibility, and the ability to collaborate. The weight assigned to each element differs between organisations.
If a less experienced person using AI can produce something faster that previously required greater familiarity, production speed becomes a less dependable indication of personal experience. If a more established worker understands the subject deeply but struggles to use the new tools, that person also needs development. Both cases require us to be more specific about what we recognise and reward.
In “AI Won’t Kill Software. It Will Kill Average Software”, I wrote about the value of understanding what is actually worth building. In “AI Made Software Easier to Build. Not Easier to Run”, I considered the demands that remain after something has been built. The same thinking applies to people: what can someone reliably take responsibility for, beyond what they are able to present?
Perhaps professional grades will retain their names but need different definitions. We may need to pay closer attention to understanding the subject, handling exceptions, working with people and systems, evaluating a decision, and following its consequences.
For an individual, that may mean being strong in one dimension while still needing guidance in another. Assessment becomes more demanding than counting years or deliverables. It can also become fairer, provided it recognises actual capabilities rather than assuming a title describes them adequately.
I would like this discussion to include both newcomers and experienced professionals. Prior experience can provide a valuable foundation for working with AI. Its value still needs to become evident in practice as new conditions emerge for all of us.
Who will teach us our work?
A new worker learned from colleagues and managers, formal education, clients and partners, attempts and corrections. The quality of that learning depended heavily on the environment they entered and on how much time other people could, or were willing to, give them.
AI offers a possibility here that I consider important. You can ask for an explanation again without feeling that you are holding someone up. Examine an example from several perspectives. Practise in a simulation before facing the real situation. Obtain assistance at the moment a question arises.
Human guidance can then focus more closely on matters that require knowledge of the particular organisation, discussion, and consideration of consequences. But that requires us to decide how these sources of learning will work together. Who explains the purpose of a process? Who checks whether the system’s instructions reflect actual practice? Who helps when two apparently sound choices lead to different consequences?
There is another dimension too. Experienced people may need to make part of their knowledge explicit so that it becomes available both to newcomers and to the AI systems supporting them. Exceptions, the reasoning behind a decision, and the lessons of a failure can then become more valuable to the organisation as a whole.
Learning can become more reciprocal. A newcomer brings fluency with tools and different approaches. An established colleague supplies context and knowledge of consequences. AI supports exploration, experimentation, and explanation. Realising that possibility requires room for genuine exchange and for questioning things that appear obvious.
Where might we lose the advantage?
We could achieve better onboarding and faster capability development. We could also limit those benefits through the way we organise everyday work.
If every minute saved becomes another deliverable, explanation and experimentation will remain things to do “when there is time”. The worker will have an incentive to finish and move on. The business may see more output without knowing whether the person’s capability is growing alongside it.
If the presence of AI becomes sufficient reason to withdraw human guidance, knowledge that has never been documented will remain inaccessible. If the system does not understand the practical peculiarities of the work, its fluent answers may create more confidence than is justified. And if experienced colleagues must review an ever-increasing volume of output, they may have less time left to help someone understand it.
We cannot place the entire responsibility on a newcomer with the instruction to “get better at using AI”. They need access to suitable assistance, time to apply what they learn, feedback, and opportunities to take real responsibility within safe boundaries.
In “AI and the New Aristocracy of Thought”, I considered the distance between access to information and the ability to use it meaningfully. A related possibility arises here: people with similar access to tools may have very different opportunities to develop, depending on the environment in which they work.
One organisation may use AI to develop its people’s capabilities. Another may be satisfied with the immediate output it secures. The difference may become more apparent over time, when someone needs to understand a new situation or take responsibility for a more demanding decision.
The progress that happens within us
I think about a young person trying to understand what is worth learning. Someone with years of experience wondering which parts of their knowledge retain their value. A team leader deciding who is ready for greater responsibility. Their questions are connected, even though they stand at different points in their working lives.
AI gives us an opportunity to change how we acquire capabilities. I would like us to approach that opportunity with curiosity and ambition: better guidance, access to knowledge, and more scope to experiment and participate. At the same time, we need to understand what progress is taking place in the person, and how to distinguish it from the improved performance of the person-and-tool combination.
I do not believe there is a definitive answer already. Occupations differ, and capabilities continue to evolve. But the question has immediate significance for education, hiring, mentoring, promotion, and how we see ourselves at work.
When we say that AI makes someone a better worker, we should be able to explain what we mean. What do they understand more deeply? Which new decision can they take responsibility for? What can they create, and how can they contribute more meaningfully to their work?
That is the progress I would like us to learn to recognise: expanding what is possible for us while developing our own ability to make meaningful use of it.
Sources and further reading
- Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory, “How employment is changing at Anthropic”, Anthropic, 16 October 2026. Internal company research; it is not a representative sample of the wider labour market.
- Anthropic, “How AI assistance impacts the formation of coding skills”, 29 January 2026. Full research paper: “How AI Impacts Skill Formation”. A small, short-duration randomised study; it does not measure professional development over several years. The difference between 50% and 67% is 17 percentage points.
- Erik Brynjolfsson, Danielle Li and Lindsey Raymond, “Generative AI at Work”, The Quarterly Journal of Economics, 140(2), May 2025, pp. 889–942. First published online on 4 February 2025. The figures cited here come from the final journal article rather than the earlier working-paper version.