The words we use

What would change a decision, as opposed to what merely happened. Most AI news is real and still below the cut, because knowing it would not alter what a reader does on Monday.

The three labels every event in an issue carries. Primary means the lab, paper, regulator, or filing said it first-hand. Corroborated means an independent source confirmed it. Reported means someone said it and we have not been able to confirm it, so the wording stays hedged and attributed.

No link ships unverified. Every external URL is opened before send and confirmed to load and to contain the specific fact cited, not merely to be about the topic. The check runs as a mechanical gate that blocks the issue, and we link the deepest primary artifact rather than coverage about it. Aggregators and search-results pages are never cited.

A claim of AI value made without a number a reader could underwrite. It is the most common shape of AI announcement: real activity, real spend, and nothing an outsider could check or a CFO could book.

The account for what an AI investment actually produced, and where the value went. It answers two things at once: what value AI created, and who partakes in it.

Gross AI Yield = release + expansion. The two can be added only because both are money against the same baseline, for the same operation, in the same period. Net AI Yield = Gross AI Yield less the whole AI investment assigned to that same operation and period.

The whole AI investment is not the model bill. It is technology and compute, data, integration, evaluation, controls, change effort, and transition.

The two halves of the yield, and they arrive on different clocks. Release is cost that stopped: a position leaves the plan, a vacancy stays unfilled, a contract ends. It lands in next quarter's accounts and it is easy to prove. Expansion is a bet: the company keeps the capacity, points it at new or better work, and waits to see whether customers reward it. It may arrive late, or not at all.

Which one produced a company's number tells you more than the number does. A yield built almost entirely on release is a cost story. Expansion carries more uncertainty even when it creates the more valuable future.

A freed hour is capacity, not cash. When AI improves existing work, management then chooses: release the capacity, so the position, contract, or licence actually stops, or redeploy it, giving the time another job.

Not every part of the account is knowable the same way. Investment, capacity choices, and release can be recorded: a company knows what it spent, what it decided, and which costs stopped. Expansion and the overall result can only be observed against a baseline, because demand, pricing, competitors, and the wider market move at the same time. Attribution to AI is then a stated judgment, not a measurement.

The other half of the question, and the half a single financial number hides. Who received the gain, who carried the cost, and how fast did either arrive. When release drives most of a yield, the money tends to reach the enterprise, its owners, and the leaders rewarded for delivering it.

Customers benefit only if the company decides to lower prices or improve service. Pass-through is a decision, not a trickle. Workers receive a share only if someone designates one. If nobody names another destination, the result stays in earnings.

A pilot running in production with no effect on the P&L. Deployment is evidence that something works, not evidence that it paid.

Progress gated by reliability and control rather than by capability. When a system is good enough to be useful but not dependable enough to be relied on unsupervised, the binding constraint is trust, and more capability does not move it.

Excess returns that persist after competition catches up. The everyday sense of the word is irrelevant here; this is the economic sense. The test is whether the underlying advantage is excludable, meaning rivals cannot simply acquire it, rather than whether it is scarce today.

A real AI gain that never reaches the P&L because every rival deploys the same capability and the benefit passes through to customers. The technology worked exactly as promised and the company is no better off than before.

AI spending undertaken to stay in the game rather than to get ahead of it. Necessary, unavoidable, and not a source of advantage. Where durable advantage exists, it usually lives in scarce complementary assets a rival cannot copy, such as proprietary data, distribution, regulatory position, or workflow lock-in, rather than in the model itself.

The order diligence has to run in. Verify that a system works and is controlled before scaling it, because that is testable now. Reason about whether the advantage survives competition before committing capital, because it cannot be tested until later. Then judge the result against what was argued.

A human-led, reconfigurable way of organizing people, AI, software, data, rules, tools, memory, checks, and judgment around a responsibility or mission. It may be standing or temporary, and it can be nested inside another.

It is not several AI personas sitting side by side. It is a picture of the work rather than a headcount of software: the roles the work actually needs, each built with the simplest thing that does its job reliably, whether that is a separate agent, an instruction, a rule, or a person.

The attention it costs to hand work to AI responsibly: framing the task, assembling context, steering, inspecting the return, correcting it, and being able to explain why the result deserves confidence. The handoff is cheap. Accountability is not. When the tax is not paid by the sender, it is simply transferred to the recipient.

The point at which work has earned its next level of reliance. Work crosses it by one of three routes: a credible person can grade it, evidence or a test can expose the important failure, or reality can answer while the stakes stay small and reversible. Absent all three, the output may still be useful for exploration, but it has not earned the status of an answer.

The point where further gains stop depending on how well one person uses AI and start depending on whether the work around that person changes. It is the signal that the next problem is shared context, workflow, standards, review, decision rights, or team design. My work gets better before our work gets different.

AI is smart but far from wise. Left unsupervised it invents facts, agrees with a false premise, drops the detail buried on page forty, and quietly applies an incomplete standard, fluently and at scale. A sharper model makes the fluent answer more fluent, not more accountable, which is why quality has to be designed for rather than waited out.