Intelligence work, judgement work
First course in the track. Before choosing any tool, you need to know what kind of work is at stake. That separation decides where AI goes first and where it should not go yet.
- 01 · You are here
- 02 · Where you already outsource
- 03 · Why your pilot never shipped
- 04 · Who answers when AI gets it wrong
- 05 · The agenda your budget holder approves
Sort any task in your area into two kinds: the ones that follow a rule you can write down, and the ones that depend on judgement. With that, you decide what is worth automating now, what is worth only supporting, and what stays with people.
The wrong question costs the whole budget
The conversation almost always starts in the same place: which tool to buy. It turns into a feature comparison, a price per user and a rollout deadline. Months later the licence is paid, a few people use it, and nobody can say which work is now done differently.
The reason is simple: a tool is an answer, and nobody had asked the question. The question is this: which work, done one way today, no longer needs to be done that way?
A hard task is not the same as a judgement task
Think of these four: turning a specification into code, checking a standard contract against a clause list, classifying an invoice, reconciling a bank statement. All of them take real expertise and cost money when done badly. But all of them follow rules — complicated rules, yes, but rules. They can be written in a document, reviewed, and checked afterwards.
Now compare with these: deciding what to build next quarter; accepting a technically worse solution so as not to miss a deadline; granting an exception to a supplier who missed a date but matters to the business. There is no list here. There is a criterion the person built over years of practice, knowledge of the context, and the experience of answering for the consequences.
The more a task is done by following a rule you can write down, the more likely the current way of doing it changes with AI. What matters is the proportion inside the task, not the job title of the person doing it.
A contracts analyst spends part of her day checking clauses against a standard template and part negotiating exceptions with other teams. The first part follows a rule. The second is pure judgement. The work does not leave her desk: what changes is how much time she spends on each half.
Reading this split as a list of who gets made redundant leads to the wrong decision. It exists to find where a person is being used as if they were a rule, and where their attention could go instead: exceptions, decisions, creation, negotiation and responsibility.
Sell a tool and you compete with the next launch
When what you deliver is a tool, every new model that comes out threatens your business. When what you deliver is finished work, every better model makes your delivery faster and more controlled.
The difference is about money. What a company spends on the work of a department is far more than what it spends on software licences per user. That is why it makes more sense to charge for the work delivered than for the number of active seats.
If your next decision is which subscription to buy, you are buying a tool. If it is about which work gets done differently, and who answers for it, you are changing the way the work is done. The two cost about the same and the results are completely different.
“Our operation is far too complex for this”
It probably is. But complexity is almost never spread evenly. Inside any complicated operation there are tasks whose rule is long, dull and always the same. Those are exactly the ones that eat up hours of expensive people and resist every attempt at improvement, because nobody ever stopped to redesign them.
The point here is not to classify the whole company. It is to take five concrete tasks and see which side each one falls on.
Which part of your judgement nobody has written down yet
If part of your team's work follows a rule, part of yours does too. Approving what you have already approved ten times, for the same reasons, is not judgement: it is a rule nobody put on paper.
The uncomfortable question is this: how much of your own criteria could you write down for someone else to decide in your place? What is left after that exercise is the judgement that stays yours — and the responsibility that stays with a person.
Take the candidates into your next meeting
Take the tasks you marked as candidates for automation and answer three questions together with whoever does the work: how many hours a week they take, who reviews the output, and what happens today when they come out wrong. Those answers are what course 02 starts from.
Two possible relationships between a system and an activity
The split is not between professions. It is between kinds of activity inside the same role — and most people have both in the same week.
Substitution on one side, complement on the other
Read it as a test you can run on any task on your desk, not as a list of who stays and who goes.
Activity carried out by an explicit rule
- The criterion can be written down before looking at the case.
- Two competent professionals reach the same result.
- Exceptions are rare and already have a defined treatment.
- Checking a format, sorting by a given criterion, moving data between systems.
Non-routine problem-solving and complex communication
- The criterion only becomes clear after examining the actual case.
- Two competent professionals may reasonably disagree.
- Exceptions are frequent, and that is where the value sits.
- Deciding under ambiguity, negotiating, explaining to someone who resists, carrying the consequence.
Why does reading this as a list of job titles lead to the wrong decision?
What is the quick test for which column an activity belongs to?
What changes in the budget when you sell completed work instead of a tool?
Try to answer before opening. An attempt that fails sustains more retention than rereading the explanation.
Where the argument comes from
- Julien Bek, “Services: The New Software”, Sequoia Capital, 5 March 2026. Supports the distinction between selling a tool and selling completed work. The ratio of six dollars spent on services for every dollar spent on software is a claim of the text itself, not our inference. Verified 6 Aug 2026: authorship, date and URL confirmed on Sequoia's page and in three independent outlets.
- David H. Autor, Frank Levy and Richard J. Murnane, “The Skill Content of Recent Technological Change: An Empirical Exploration”, The Quarterly Journal of Economics, vol. 118, no. 4, November 2003, pp. 1279–1333. The authors' formulation: computer capital substitutes for workers in cognitive and manual tasks that can be accomplished by following explicit rules, and complements workers in nonroutine problem-solving and complex communications tasks. Method: representative data on task input from 1960 to 1998. Verified 6 Aug 2026: title, journal, volume, pages and DOI confirmed on Oxford Academic and NBER.
When the classification becomes an investment decision
Prumo Discovery is TheNeil's paid diagnosis: fixed scope and timeline, run with your teams, and it can recommend not proceeding. It is the next step when there is a candidate operation and the decision to invest is still open.