Internships are dead.

That is the thought I keep returning to when I imagine a company designed around AI from the beginning.

Give a capable agent the right tools, a clear playbook, access to the relevant information and a way to check its work. My bet is that it will outperform an intern on much of the routine digital work we have used to introduce people to a profession: collecting information, assembling a first draft, comparing documents, preparing a recurring report.

The claim is about a model of work. It is a prediction about where that model is going, rather than a finding that every agent already beats every beginner. Internships involving physical practice, relationships or specialist training raise different questions. But for the internship whose main proposition is an extra pair of hands at a screen, I think the change is fundamental.

The difficult part comes next. Where does a person begin?

The bargain inside the internship

Consider the familiar arrangement: a beginner does useful but limited work, and the organisation provides exposure, correction and a first professional reference. The work helps pay for the learning.

Imagine a small company asking an intern to prepare a supplier comparison. The assignment involves finding product specifications, extracting prices, checking delivery terms and turning the findings into a recommendation. There is real learning in that process: discovering which sources to trust, spotting missing information, realising that two similar prices can mean very different things.

Now imagine a well-equipped agent doing the same assignment. It has access to the source documents, a template that captures the relevant differences and a rule that every quoted term must link back to its source. It can be asked to identify missing evidence before making a recommendation.

This is a hypothetical workflow, not a result measured at Earthrise. Its attraction is easy to see. If the output becomes faster to produce and easier to verify, hiring someone mainly to assemble it becomes harder to justify.

A company can still choose to train a beginner. But it has to recognise training as an investment in its own right.

The tools and playbooks matter

The condition in this argument does a lot of work: given the right environment.

A model without access to the actual documents may write a fluent answer to the wrong question. An agent with outdated instructions may execute an obsolete process very efficiently. A system that cannot tell whether its changes saved has only attempted the job.

Anthropic’s engineering guidance describes agents that use tools and feedback from their environment, and emphasises clear tool design and evaluation. That supports the importance of the surrounding system; it does not establish that an agent is a universal substitute for a junior employee. Building effective agents.

For the supplier comparison, quality would mean checking whether the prices refer to comparable quantities, whether delivery charges are included, and whether the recommendation changes when an uncertain lead time is corrected. A beautifully formatted table is a weak measure of success.

I would compare the complete cost of getting to an accepted answer: setup, execution, checking and correction. The agent earns the work when that whole process is better. The claim should survive that test.

The evidence complicates the headline

There is an important counterargument: AI can make a beginner more useful.

In Generative AI at Work, Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the introduction of an AI assistant in customer support. Less experienced and lower-skilled workers improved both their speed and the quality of their output. The study also found evidence of learning. This was assistance for human workers, not a test of autonomous agents replacing interns. Research paper.

So the relevant comparison includes a beginner equipped with AI. That person may become capable of useful work much earlier.

Nor should we assume that AI makes every task faster. METR’s early developer experiment found a slowdown in its particular setting. Its subsequent update said newer results were affected by selection and measurement problems and could not reliably establish the current productivity effect. METR nevertheless considered greater speedups likely, based on conversations with participants; it cautioned that the data offered only weak evidence about their size. These findings are a reason to test the actual workflow, rather than borrow a universal speed claim. METR’s research update.

I still expect the routine-output internship to lose ground. The stronger version of that argument leaves room for a different kind of beginner: someone who can use agents well and is deliberately being taught the judgment needed to direct them.

The missing first rung

There is a problem hidden inside that optimistic answer.

How does someone learn to recognise good work if they never struggle through making any of it?

An experienced person may notice that a supplier’s attractive price assumes an unrealistic order quantity. A beginner may see a polished recommendation and accept it. Telling the beginner to “review the AI” does little if we have never taught them what to look for.

If we remove routine work and keep hiring only people with judgment, we risk consuming experience without replenishing it.

That is the part colleges and employers need to design together. Requiring students to arrive with agent skills does not settle who will teach them domain knowledge, how they will encounter consequences, or where they will learn to challenge a plausible answer.

If the first rung becomes automated, we have to build another way up.

What I would put in its place

I would rebuild the internship around a small piece of real responsibility.

For the supplier assignment, the beginner would help define what the company is trying to optimise: delivery reliability, flexibility, total cost or something else. They would explain which evidence can answer that question. Then they could use an agent to gather and organise it.

Their review would include tracing a sample of claims to the original documents, testing a recommendation under a changed assumption and naming what remains unknown. A mentor would challenge their reasoning and correct the work with them.

Occasionally, the beginner should work through a case manually. The point would be to learn the structure of the problem well enough to recognise when the automated version has gone wrong.

The final assessment should examine the decision and the learning: what did they catch, what did they misunderstand, what would they do differently? A folder full of generated documents would tell us very little.

This requires protected mentoring time, access to suitable tools and a clear boundary around decisions the beginner can make. Calling someone an agent supervisor without giving them those things merely moves the burden.

For colleges, a useful placement could become a supervised project with a real organisation: an actual question, inspectable evidence, feedback from a practitioner and a student who can defend the result. Employers would have to help create that opportunity rather than assume it will emerge elsewhere.

A different beginning

My conviction is that internships built around cheap, routine digital output will become increasingly difficult to defend in an AI-first company. The title may survive. The job underneath it will have to change.

That leaves a choice for the organisations building this future. We can use agents to remove the work beginners used to do and leave the question of experience unresolved. Or we can use the same tools to give beginners a wider view of a problem, with enough supervision to turn that view into understanding.

The second path asks more of the employer. It also makes the purpose of an internship clearer.

The work should leave someone more capable of carrying responsibility than when they arrived. If agents change how that happens, we should redesign the beginning.

Continue reading: Manager of Agents.