Chapter 26: AI and the Future of Work
Beginning with honesty
Nobody knows.
I want to be plain about that before offering anything, because confident prediction is this subject’s dominant mode and its record is poor. Seventy years of AI history show serious people getting timelines wrong in both directions. The people forecasting mass unemployment by a specific date and the people forecasting that nothing much will change are both extrapolating past the evidence.
What follows is the part I think is reasonably well supported, and I will flag where it stops.
Tasks, not jobs
The single most useful correction to public discussion: AI does not automate jobs. It automates tasks.
Almost every job is a bundle of tasks, and AI’s effect on your work depends on the proportions in your particular bundle. Two people with the same job title can have very different exposure.
Tasks with high exposure: producing routine written material, summarizing and synthesizing information, retrieving and organizing information, first-pass analysis, routine translation, transcription and note-taking, initial drafting of structured documents, basic code.
Tasks with low exposure: physical work in variable environments, work whose product is a relationship, work requiring accountability someone will accept, judgment under genuine ambiguity, negotiation where reading a room matters, coordinating people who disagree, anything requiring presence.
Look at your own week honestly against those two lists. That ratio tells you more than any national statistic.
What is happening now
Three things I would say with reasonable confidence.
Task compression is real and uneven. People doing text- and information-heavy work are completing certain tasks substantially faster. The gain is concentrated in specific tasks rather than spread evenly.
The gains often appear as greater volume instead of shorter workdays. Freed capacity gets absorbed by reading material previously skipped and attempting work that was not worth the hours before. That expansion can be valuable even when it does not reduce working time.
The bottleneck moves to verification and judgment. When producing a draft costs minutes, the constraint becomes deciding whether the draft is any good and what to do with it. That is a human activity, and it is where value is accumulating.
Which brings me to an observation I want to make carefully, because this book draws a firm line between what is known and what is guessed. The internet all but eliminated the travel agent and other occupations. My expectation (and this is a prediction, not a fact) is that AI will do something similar to work built largely on searching, summarizing, and drafting. Paralegal work is on that list. I do not expect the profession to vanish; I expect fewer people to do more of it, with better tools.
Reasonable people disagree about how far and how fast this goes and will go.
Time will tell.
The problem I find genuinely worrying
Here is the thing I have no comforting answer for.
Junior work has traditionally been how people become senior. The first-year associate reads the documents. The junior analyst builds the model. The apprentice does the repetitive part badly and then less badly. That work was economically marginal and developmentally essential — it is how judgment gets built, by doing the thing enough times to develop instincts about it.
Much of that work is exactly what AI now does. And if organizations stop hiring for it, or stop having juniors do it, the pipeline that produces experienced people has a gap in it that will not become visible for a decade.
I do not know how this resolves. I notice that the people most confident it will work out are rarely the ones who would bear the cost. If you manage or train people, this is worth thinking about deliberately, because the incentive to cut the junior work is immediate and the cost is deferred and diffuse.
If you are early in your career, the practical implication is that you may need to seek out the developmental repetition rather than expecting a job to supply it. Do the analysis yourself before checking it against AI. The friction is the point.
What strengthens your position
Technology eventually consumes many specific skills. The following capabilities are more durable:
Judgment in a domain you know deeply. Knowing which answer is right, why the plausible one is wrong, and what the situation actually calls for. This gets more valuable as production gets cheaper, not less.
Accountability. Being the person who signs, who answers for it, who is trusted with the consequence. Organizations cannot delegate this to a system that cannot be held responsible.
Relationships and trust. Built over years, not transferable, and largely unaffected by any of this.
Directing the work. Increasingly, competence means knowing what to delegate, what to check, and how to combine machine output with human judgment. That is a skill, this book has been teaching it, and it is in short supply.
Adaptability. The willingness to keep learning tools that keep changing, which may be the most durable capability in this book.
Practical moves
Become the person in your organization who understands this technology with sound judgment about what works, what fails, and where the risks lie. That role is vacant in many organizations and available to anyone willing to do the reading.
Deepen rather than broaden. Generic competence is where the pressure is greatest. Specific expertise in a domain, with judgment about its edge cases, is where it is least.
Keep your judgment sharp. Anything you always delegate, you will eventually be unable to evaluate. Do some work manually on purpose — this is the professional version of the advice above about juniors.
Document what you actually do. Most people underestimate how much of their work is judgment and relationship rather than production. Knowing your own ratio is the beginning of managing it.
Build optionality, not a five-year plan. Nobody can see far enough to plan against. Skills that transfer, relationships that persist, and a habit of learning quickly are what work under uncertainty.
What I would not do
Panic. The historical record on technology-driven mass unemployment predictions is not good, and panic produces poor decisions.
Assume immunity. “AI can’t do what I do” is being said in a great many fields, and it is partly right in most of them and completely right in fewer than the people saying it believe.
Wait to see how it shakes out. The earlier argument still holds: the gap between people using these tools well and people not using them is a gap in judgment, and judgment only accumulates through use.
Exercise 26.1: Audit your own bundle
List everything you did at work last week. Mark each item high, medium, or low exposure using the two lists above.
Calculate the rough proportion. That number is your actual exposure, and it will differ from your intuition.
Exercise 26.2: Find the judgment in your work
Look at the low-exposure items. What makes them low-exposure? Is it judgment, relationship, accountability, or physical presence?
Those are the parts to invest in. Write down one way to deepen each over the next year.
Exercise 26.3: Ask the uncomfortable question
I’m a [role] doing [description of actual work]. Which parts of this are most likely to be automated in the next five years, and which are most durable? Be direct rather than reassuring.
Then apply Chapter 9. It does not know the future, and its answer is a useful prompt for your own thinking rather than a forecast.
Learn More About Artificial Intelligence
Learn more about the CLEAR, TRUST and LEARN methods for Artificial Intelligence