AI systems /
When doing gets cheap, defining the job becomes the work
AI has made it cheaper to produce code, reports, research and first drafts.
That is a meaningful change. It is not the whole change.
When execution gets cheaper, the valuable work moves towards choosing the right job, giving it the right context and judging whether the result is good enough.
A prompt is not a job
Most people still use strong models for prompt-sized tasks. They ask for an email, a summary or a function. Those are useful, but they hide the difference between a model that writes a paragraph and one that can carry a complete piece of work.
Consider customer reviews. A small request is: 'Summarise these reviews.'
A real job is different. Give the model the reviews, product catalogue and support history. Ask it to identify recurring problems, separate defects from misunderstandings, show the evidence and prepare a report for the product team. Send uncertain cases to review.
The difference is not a longer prompt. It is a defined outcome.
For a model to carry a job well, it needs what a capable person would need:
- the source material and an authoritative version of it
- rules for the decisions it can make
- boundaries on what it may change or send
- evidence a reviewer can inspect
- a clear definition of done
The scarce skill is seeing useful jobs
Many valuable jobs already exist inside a company. They are too large, messy or unclear for someone to pick up between meetings.
The compliance folder nobody trusts. The customer records full of duplicates. The field reports that should be checked against current procedure.
AI can help with that work once someone turns it into an executable job. Someone still has to know which source is correct, which exception matters and which decision carries risk.
This is where domain knowledge becomes more valuable, not less. A person who understands the work can turn a week of manual effort into a repeatable process. A person who does not is likely to hand the model a vague brief and accept a polished guess.
Let the model work inside a visible system
A capable model should not need supervision for each step. It should operate independently within a workflow that has clear limits.
Keep the sources it used. Record important decisions. Test what can be tested. Put approval gates before consequential actions. Make it easy for a person to inspect the final result.
Those controls make larger assignments safer. They also make them easier to improve after the first run.
The company work matters here too. Models cannot reliably carry a complete job when the organisation cannot explain where its knowledge lives or who is allowed to decide what.
The rule
Do not judge a model by how much better it makes a small prompt.
Give it a real job with sources, boundaries, evidence and a definition of done. Then judge the completed work.