What Will Be Left for Us to Work On? An Operator's Read
Arvind Narayanan argues AI transforms work over decades, not overnight. After 17 years running IT through five tech waves, I think he's mostly right.
Arvind Narayanan gave a keynote recently titled "What will be left for us to work on?" — part of his ongoing AI as Normal Technology work with Sayash Kapoor at Princeton. It's the most grounded thing I've read on AI and jobs, and I want to summarize the argument and then add what 17 years of running IT operations taught me about how technology actually lands inside a company.
His argument, in three points
1. Capability is not reliability. Narayanan's team tracked frontier AI agents over two years. Raw capability climbed fast; reliability — consistency, robustness, calibration — barely moved. This matches what every operator using coding agents already knows: impressive on the happy path, untrustworthy without review. It's why AI works today as a collaborator and fails as an unattended replacement.
2. The job is a sandwich, and AI only eats the middle. He frames knowledge work as decide → execute → deliver. AI compresses the execution layer. But deciding what to build and verifying that it actually works expand to fill the freed-up time. The work doesn't disappear; it moves up the stack, from doing to judging.
3. Cheaper work creates more work. James Bessen documented this in Learning by Doing (Yale, 2015): ATMs made bank branches cheaper to run, so banks opened more branches, and teller employment went up for decades. Software engineering headcount grew enormously through fifty years of relentless productivity gains. When a task gets cheaper, demand for it usually grows faster than the labor it displaces.
His conclusion: no lab announcement will put everyone out of work on a Tuesday. Diffusion — organizations, workflows, trust, regulation — takes decades, and that lag is where humans adapt.
Where I've seen this movie before
I ran IT for a remote-staffing company that grew from 60 people to 1,500. In that time we lived through five "this changes everything" waves: open-source ERP, VoIP replacing telephony hardware, cloud replacing our racks, BI replacing gut-feel reporting, and now AI.
Every single wave followed Narayanan's curve, not the hype curve. The technology was "ready" years before we could deploy it. Not because we were slow — because deployment is the hard part. VoIP was mature in 2005; getting 1,500 concurrent agents on it with call quality a client would accept took years of unglamorous work on networks, failover, and training. The capability existed. The reliability, integration, and organizational trust had to be built, one painful quarter at a time.
That gap between "the demo works" and "the business runs on it" is where careers are made. It's also why I don't believe the overnight-displacement story.
My one pushback: the lag is real, but it's shrinking
The historical analogies assume adoption friction stays constant. It doesn't. ATMs needed physical installation in every branch. An AI model upgrade arrives through an API — the same day, everywhere, to every company already integrated. For purely digital work, "decades of adjustment" may compress into years.
The direction of Narayanan's argument is right. The timeline is the part I'd hold loosely.
What I'm doing about it
Narayanan says he spends about ten hours a week learning new AI workflows, and warns against what he calls the dependence spiral — offloading to the machine before you've built the judgment to check it. That's the advice I'd give any IT leader or founder right now:
- Automate the execution, keep the judgment. Let agents write the code, the config, the first draft. Never let them own the decision or the verification.
- Reinvest the saved time upward. The hours AI gives back should go into the decide and deliver layers — architecture, evaluation, client trust — not into doing more of the same task faster.
- Build the skill before you delegate it. Same rule I applied with every wave: run open source before you buy enterprise, understand the PBX before you outsource telephony. You can only supervise what you could, in principle, do yourself.
The question was never "what will be left for us to work on?" Every wave I've lived through left more work, just higher up the stack. The people who got hurt weren't the ones whose tasks got automated — they were the ones who refused to move up.
References: Arvind Narayanan, "What will be left for us to work on?" (AI as Normal Technology, 2026); Narayanan & Kapoor, "AI as Normal Technology" (Knight First Amendment Institute, Columbia, 2025); James Bessen, Learning by Doing: The Real Story of Technology, Wages, and Wealth (Yale University Press, 2015).