Research

Frontier AI capability is compounding at roughly five times a year while 1–5% of U.S. work hours actually use it. The steam and electricity records read that gap as a stage that closes through slow organizational co-invention; the skeptics read it as a sign the gains will be small. What the verified sources support on each side.

18 verified sources

Through 2024, AI revenue pooled in infrastructure, thinning by an order of magnitude at each step toward applications. In 2025, enterprise spending split almost evenly between the two. This paper maps where the value pools, what defends it, and the two dissents that contest the consensus.

18 verified sources

Redesigning work around AI captures gains three to five times larger than inserting tools into an unchanged process. It is also the move that most often falls short of expectations, so the strategy question is a priced bet between a bounded gain and a larger, riskier one.

19 verified sources

A job is a bundle of tasks. AI takes them up one at a time, and what happens to the job depends on the tasks it leaves behind, which is why occupation-level predictions of mass displacement keep missing.

25 verified sources

Enterprise vendors now sell agentic AI as digital labour, a workforce a firm builds, prices and manages. Three claims travel under that banner, and the evidence treats them very differently.

23 verified sources

Organizations bought the technology; most have yet to capture the value. The distance between the two is human and organizational. Trust, identity, workflow, and incentive decide whether a tool in hand becomes a tool in use.

22 verified sources

AI covers the codifiable majority of work, while the residual it handles worst, tacit judgment, is built by doing the entry-level tasks AI absorbs first. A firm that automates its bottom rung saves cost this quarter and may be dismantling the process that produces its senior bench.

14 verified sources

Why a well-run organization absorbs a new technology rather than being transformed by it — and the half-century of theory that explains the gap between AI adoption and AI value.

17 verified sources

Agents change what a software company sells, what the selling costs, and who answers when the work goes wrong. All three move with how much control the buyer cedes to the machine.

27 verified sources

AI answers are absorbing the search click, citable reputation is displacing the keyword, and the payment incumbents have shipped rails for agents to complete the purchase. Consumer trust gates the shift, and ownership of the customer is the prize.

20 verified sources

Frontier safety now works as a stack of partial mitigations governed by capability-triggered frameworks. The scaling recipe beneath that stack faces a projected data wall, and the newest documented harms are relational ones the frameworks' printed thresholds do not name.

28 verified sources

Adoption is near-universal and spending is real; measured impact is still early and small.

21 verified sources

Separating the agent from the chatbot, the copilot, and the script — and why the line matters when every vendor is blurring it.

13 verified sources

A term pulling in four directions at once — and what a disciplined, project-level definition of it actually contains.

15 verified sources