Eighty percent embedded, thirty-one percent deployed
The most important business statistic in AI right now is a pair of numbers that almost never appear together. Eighty percent of enterprise applications shipped or updated in Q1 2026 embed at least one , up from thirty-three percent in 2024. And thirty-one percent of enterprises have an AI agent actually running in production.
That gap: between shipped-with and running-in, is the entire business story of this year. Everything else is commentary.
The vendors closed the feature gap in eighteen months. The buyers have not closed the deployment gap and are showing no particular sign of closing it soon. Twenty-three percent report significant ROI from agents. Twelve percent of CEOs report both revenue growth and cost reduction. Fifty-six percent report no significant financial benefit yet. Gartner expects more than forty percent of agentic projects to be cancelled by the end of 2027.
I want to be careful here, because this data is mostly tier-C survey material, much of it vendor-sponsored, with inconsistent definitions of "agent," "production" and "significant." Any individual number could be off by a lot. But every source points the same direction, and the direction is what matters: capability is abundant and deployment is scarce.
Which is why Palantir is the most instructive company in the sector
Palantir grew revenue 93% year-over-year to $1.94B last quarter, earned roughly $1.1B, and raised full-year by nearly a billion dollars mid-year. It did this in the same market where most enterprises report no measurable return on AI.
Palantir does not have a better model. It does not have a model at all in any meaningful competitive sense. What it has is the ability to walk into an organization with bad data, political dysfunction and legacy systems, and make something work anyway. For years that was described dismissively as a consulting business wearing a software multiple. That description is correct and it turns out to be the best position in the industry.
The generalizable claim: in a market where intelligence is commoditizing toward the price of Chinese , the scarce input is not intelligence. It is the organizational capacity to absorb it. Model quality is converging. Deployment competence is not, and appears not to be something you can buy.
If you are choosing where to build, build in the gap. The unglamorous work, data plumbing, evaluation harnesses, permissions, change management, the human processes around the agent, is where the durable margin is. The model is increasingly the cheapest component in the system.
Google is the most interesting business story and it is being covered as gossip
In roughly one quarter, Alphabet lost Jeff Dean and Sanjay Ghemawat after twenty-seven years, along with Oriol Vinyals and Quoc Le, to a new venture. It lost Noam Shazeer to OpenAI and John Jumper: the AlphaFold lead, a Nobel laureate, to Anthropic. Demis Hassabis moved from running DeepMind to chairing it. Its flagship Gemini release slipped.
Meanwhile Alphabet has the best infrastructure position in the industry, that is years ahead of any other in-house program, distribution to billions of users, and Google Cloud growing 63%. The stock was up 12.4% in H1 before falling 5% on the Dean news.
So which is it? Both, and the tension is the point. Google's problem is not capability or capital. It is that the returns to being inside a large organization have fallen sharply relative to the returns to leaving one. Compute is rentable. Talent is mobile. The capital is standing by. What a big company uniquely provided (access to scale) is now a purchase order.
The detail that gives this away: Google is a founding investor in Discovery Loop, its cloud partner, and is supplying first-year compute. That is not a blessing. It is the structure you accept when the alternative is a hostile competitor built by the people who designed your infrastructure. Alphabet chose a minority stake in the thing it could not prevent.
I do not think this is a Google-specific pathology. I think it is what the org chart looks like everywhere for the next several years, and Google is simply first because it has the most people worth poaching.
The labor story is being told wrong
The coverage frames AI job displacement as layoffs. The data says otherwise. Roughly 50,000 cuts in 2026 have been attributed to AI, about 17% of announced cuts. Meaningful, not apocalyptic.
The real number is elsewhere: employment for software developers aged 22 to 25 is down nearly twenty percent since 2024, and workers aged 22 to 30 are being displaced at roughly three times the rate of workers aged 40 to 55. AI was cited in 0.6% of US job cuts in 2024, 4.5% in 2025, and 13% by Q1 2026.
This is not a firing wave. It is a hiring freeze at the bottom of the ladder, which is quieter, slower, and considerably harder to reverse. Companies are not replacing junior staff with AI so much as declining to hire the next cohort because the marginal junior is no longer obviously accretive.
The consequence nobody is pricing: a profession that stops hiring juniors has no seniors in a decade. Senior engineers are not manufactured. They are juniors who were paid to be mediocre for three years while they learned. Every organization currently optimizing away its entry level is drawing down a stock of expertise it has stopped replenishing, and the bill arrives long after the executives who authorized it have moved on.
I would not bet against this being recognized as a serious strategic error by 2030. I would also not bet on anyone acting on it before then, because the costs are deferred and the savings are immediate, and that asymmetry has never once been resisted.
Three calls
Deployment beats models. The companies capturing enterprise AI value in 2027 will look more like systems integrators with software margins than like model vendors. The model layer commoditizes; the integration layer does not.
Security is the most under-built category in the stack. Prompt injection up 340%, 520 tool-misuse incidents this year, no architectural fix, and agents shipping anyway. Alphabet paid $32B for Wiz and Palo Alto paid $25B for CyberArk, the strategics can see it. The first catastrophic enterprise AI incident will be a security event, and it will set enterprise adoption back by quarters.
The talent diaspora accelerates. Discovery Loop is the template, not the exception: senior researchers leaving incumbents with the incumbent's money, because scale is rentable and autonomy is not. Expect at least two more spin-outs of this shape within twelve months.
What would change my mind
- If enterprise ROI numbers improve materially in the next two survey cycles, the deployment-gap thesis weakens and model quality matters more than I am crediting.
- If Palantir's growth decelerates sharply, the deployment-competence argument loses its best evidence and may just have been government contracting all along.
- If entry-level technical hiring recovers, the ladder-collapse argument was a cyclical artifact of rates and post-2021 over-hiring, not an AI effect. This is the one I am least certain about, and the confound is genuinely hard to separate.
- If Discovery Loop struggles to ship, the "scale is rentable" thesis is weaker than it looks and incumbency is worth more than I think.
Predictions from this editorial are logged in the ledger with resolution dates and falsifiable criteria.