Actionable insights — Does going all-in on AI actually pay?

1. The revenue-vs-headcount test — the only AI claim that can't be faked

1. Ignore AI product announcements and "AI-powered" language in the release. Pull two series instead: revenue and period-end headcount, from the year AI adoption began to the latest reported year.
2. Compute revenue per employee at both ends. Real operating leverage from AI shows up as revenue rising while headcount is flat or falling — not as a margin footnote management attributes to AI.
3. Check the denominator honestly: a headcount fall that came from selling or closing a business, or from a plain cost-cutting layoff into a shrinking top line, is not AI leverage. The signature you want is growth on a smaller base.
4. Then ask what mechanism produced it — a company that can name the specific policy (below) has a repeatable process; one that can't has had a good year.

• Revenue per employee across the AI cohort vs their own pre-AI baseline and vs peers; whether headcount reductions are disclosed as restructurings (cost cuts) or simply as slower hiring (leverage); the first year the trend breaks.

2. Distinguish an AI mandate from AI tools — look for the hiring gate

1. Sort companies into two buckets: those that provide employees AI tools to experiment with (the default, and near-worthless as a differentiator), and those that have made AI use a condition of how work is done.
2. Look for the three hard mechanisms that separate them, all of which are observable from memos, filings and interviews: (a) AI use scored in performance evaluations; (b) AI-first prototyping on new projects; (c) a hiring gate — teams must demonstrate the job cannot be done with AI before a headcount request is approved.
3. The hiring gate is the highest-signal item: it is the only one that mechanically converts AI capability into the headcount line, which is where the leverage in test 1 comes from.
4. Treat a mandate without those mechanisms as marketing, and a mandate whose mechanisms are in place as a reason to expect the revenue-per-employee series to bend.

• CEO memos and shareholder letters that name a hiring gate; job-posting counts as a live proxy for the gate biting; whether the mandate survives a growth quarter, when the temptation to hire returns.

3. Make management disclose an AI-attributed operating metric — and then track its rate

1. On the earnings call, find the metric that is specific to the AI channel rather than a total — AI-driven traffic, AI-originated orders, agent-sourced sessions — and its year-over-year rate.
2. Prefer a metric that is small-but-tripling to one that is large-but-vague: an early channel growing at a triple-digit rate is a leading indicator; "AI is embedded across the platform" is not measurable and cannot be falsified.
3. Log the number each quarter and watch the second derivative. A new channel decelerating from 3x to 1.4x while still described in the same language is the tell that the narrative has outrun the business.
4. Separately ask what share of the total it is — a tripling channel that is 1% of orders changes the story only if the rate persists for several more quarters.

• Whether the company repeats the metric next quarter (a metric quietly dropped is a negative datapoint); the base it is measured off; comparable AI-channel disclosures from competitors to tell company-specific execution from an industry-wide tide.

4. In a platform shift, ask who becomes the supply the new interface must call

1. When the interface to an industry changes — here, from a human searching to an AI agent recommending — do not just ask who builds the new interface. Ask what the new interface needs to consume.
2. Identify the party that owns the aggregated, structured supply the agents cannot easily rebuild (inventory, catalog, pricing, fulfillment) and check whether it is making itself easy to integrate with every model developer rather than betting on one.
3. Neutrality is the test: a supplier wired into OpenAI, Anthropic and others is positioned to win regardless of which model wins — the picks-and-shovels position in an agent world.
4. Then verify with test 3 that the position is producing measurable volume, not just a press release.

• Which catalogs/data sets the leading assistants actually cite in product answers; exclusivity clauses (a red flag for the neutrality argument); take-rate on agent-originated orders versus conventional checkout.

5. Test an AI-disruption narrative against the maintenance half of the job

1. When the market de-rates an incumbent because "AI can now do what they sell," decompose the job into creation and ongoing maintenance (updating, patching, integrating, adapting to changing requirements).
2. Estimate what share of the customer's spend and the incumbent's revenue sits in each half. AI is strongest at the first draft and weakest at continuous upkeep.
3. If the maintenance half dominates — as it does in enterprise software — the disruption thesis is at best years out, and the de-rating is a sentiment overshoot to fade, not a terminal repricing.
4. Wait for the confirming evidence before sizing up: one or two earnings reports that fail to show the predicted damage is the trigger, because that is when the re-rating starts and the narrative traders leave.

• Net revenue retention and seat counts at the supposedly-disrupted incumbents (the number that would actually break first); customer references for AI-built replacements reaching year two of maintenance; the multiple at which the narrative is fully priced out.

6. Split an "AI is helping our prices" story into mix and pass-through

1. When average selling prices rise in a shrinking unit market and AI is credited, decompose the increase into mix/premium (a genuinely better product customers pay more for) and input-cost pass-through (a higher bill of materials handed to the buyer).
2. Pass-through carries no margin and depends on continued pricing power; mix does. Only the mix component supports a re-rating.
3. Check whether the same AI wave is on both sides of the P&L — driving the premium feature and bidding up the component the product needs. If so, the AI story is a wash until units recover.

• Gross margin, not ASP, in the next PC-maker quarter; memory contract pricing and AI data-center memory demand; whether the AI-PC premium survives once memory costs normalize.