AT&T goes open-weight, EY prices its agents: the AI story is now cost control

AT&T plans to shift the bulk of its AI workload away from proprietary models from firms including OpenAI and Anthropic and toward open-source and open-weight alternatives, according to a Wall Street Journal report summarised by Charter ^[1]. EY, meanwhile, is hiring a head of "agent economics" to run a new AI "value realization office" ^[3]. Neither move is about what models can do. Both are about what they cost and who controls the invoice.

The AT&T decision is described as part of a broader pattern of businesses trying to control costs and reduce dependency on AI vendors ^[2]. That is a procurement posture, not a technology thesis: open weights move spend from a per-token contract you do not control to infrastructure and staff you do, and they remove the single-supplier risk that comes with building products on a frontier lab's roadmap and price list.

EY's version of the same problem is organisational rather than contractual. The firm's consulting vice chair told Bloomberg that the new unit is needed to manage a workforce of AI agents that do not naturally fit the existing corporate structure ^[4]. Read plainly, that is an admission that nobody in the current org chart owns agent unit economics. Agents consume a budget like software, are supervised like staff, produce output attributed to a team, and appear in no headcount plan. Somebody has to decide what an hour of agent labour costs, which line it sits on, and who is accountable when the bill grows faster than the output.

The labour-market signal is moving in the same direction, and it is worth reading carefully. An updated version of a widely cited paper from Stanford's Digital Economy Lab found the employment gap for young workers in occupations highly exposed to AI has widened to 19% ^[5], up from 15% a year earlier ^[6]. That is a four percentage point move, roughly a 27% increase in the size of the gap ^[9]. The figure measures how far young workers' employment in those jobs has fallen behind less-exposed peers ^[7]. The same researchers state that "we do not see widespread, economy-wise job displacement associated with AI" ^[8]. Both things can hold at once: hiring at the entry rung compresses while aggregate employment does not visibly break.

The cost frontier is also shifting to data that cannot be scraped. Workers in India are being paid to mount cameras on their heads while sorting recycling or packing warehouse goods, producing training footage for robotics companies ^[10]. Humanoid robots, some of which are starting to appear in automotive plants, need real video of how human hands fold items, lace shoes, or stack boxes ^[11]. That is a supply chain with wages in it, not a scaling law.

Against all this sits a useful corrective from Oxford philosopher and AI ethics professor Carissa Veliz ^[13], who writes in her book Prophecy that "predictions are not facts" ^[12] and calls AI "the new Oracle of Delphi" ^[14]. Her operational warning to leaders is that forecasts are often built on assumptions from people with a financial stake, who are not glimpsing the future so much as constructing it ^[15]. The failure mode is not being wrong; it is being wrong while certain ^[16].

What to watch: whether AT&T publishes any cost delta from the open-weight shift ^[1], whether EY's agent economics role produces a published pricing method or stays an internal cost centre ^[3], and what the next revision of the Stanford gap does after two consecutive widenings ^[5][6].