AI’s Missing Control Group

If You Only Read One Thing

Flat profits can coexist with useful AI. Apollo’s challenge to the AI payoff deserves attention precisely because its evidence supports a narrower conclusion than its headline. Ukraine’s restored satellite service presents a separate test: demonstrated usefulness without assured continuity. Both stories expose what customers still need to establish after buying a powerful technology: the improvement it causes, and the conditions under which that improvement survives.

AI’s Payoff Needs a Better Comparison

The strongest bear case against enterprise AI is that spending can grow without customers retaining enough benefit to justify it. Apollo’s October 24 note makes that question urgent. Its claim of no financial payoff outside technology goes further than the evidence presented.

Torsten Slok compares the Magnificent Seven with the remaining 493 companies in the S&P 500. Apollo’s charts show rising revenue per worker and profit margins for the seven, but flat measures for the rest. That is a useful warning about where financial gains are visible. It is not a comparison between otherwise similar businesses with and without AI.

Think of a retailer using AI to reduce administrative work while its freight bill rises. An unchanged margin could conceal a useful saving; the same margin could also reflect an expensive failed rollout offset by stronger sales. The missing comparison is what would have happened without the deployment. Aggregate earnings alone cannot supply that answer.

The spending evidence reinforces the need to separate adoption from intensity. Ramp’s September data show 55.9% of businesses in its sample paying for AI. Among paying firms, median monthly spending was $15.70 per employee; the median within the top-spending 1% was $9,567.53. Those are very different customer economics behind the same adoption label.

There is credible evidence that individual tasks can improve. An earlier study of 5,172 customer-support agents found a 15% average increase in issues resolved per hour after AI assistance arrived. That result neither establishes an economy-wide gain nor deducts every implementation cost. It does show why “no visible aggregate margin expansion” cannot settle whether useful work is happening.

The investment implication is selective, with medium confidence over the next two years. Businesses that own recurring customer work and can verify savings are better positioned than suppliers whose revenue depends on customers continually expanding experimental budgets. This qualifies our standing distribution-and-workflow thesis: owning the customer relationship helps only if that relationship supports a repeatable economic benefit.

The strongest objection is that implementation takes time and today's spending creates tomorrow's returns. Fair. But delayed returns still require financing, and customers can cut budgets before the learning process finishes. The decisive evidence would be several named non-tech adopters reporting sustained operating-cost reductions after including software, integration, and review costs. That would weaken the spending-without-payoff concern more convincingly than another adoption percentage.

Ukraine Regains Service, Not Independence

Ukraine has recovered an important military capability without demonstrating that it controls the conditions of access. The restoration shows why a functioning satellite connection and a dependable procurement arrangement are different assets.

On October 24, Ukraine’s defence ministry reported stable Starlink/Starshield connections at movement speeds above 400 kilometers per hour. That measures how fast a connected vehicle can travel, not its internet download speed. The ministry said it had previously recorded restrictions and would continue monitoring the service.

This advances our October 15 coverage of satellite switching costs. The new information is observed restoration after restrictions, rather than another promise that restoration is coming. The ministry credits high-level dialogue, but its notice does not disclose new contractual rights preventing another interruption.

The distinction matters because a drone operation depends on several inputs working together. Hardware, software, trained crews, and connectivity are complements: the loss of one can sharply reduce the usefulness of the others. A satellite supplier’s importance therefore exceeds its share of the customer's equipment bill. That dependence persists after service resumes.

The commercial advantage and the political liability come from the same source. SpaceX remains difficult to replace quickly; customers have reason to keep using an effective network. Yet an interruption gives governments a reason to finance alternatives even when those alternatives initially cost more or perform less well. The customer may value control over availability separately from technical performance.

My medium-confidence investment read is that SpaceX retains the stronger near-term operating position, while competing satellite and terminal suppliers gain a more credible case for public funding over several years. That is an opportunity to win funded procurement, not proof of imminent revenue or a reason to treat every satellite venture as a beneficiary. Restored service can also reduce the political urgency to pay for redundancy.

This extends the familiar bottleneck pattern into communications, with a qualification: the scarce input is reliable permission to use capacity, not simply capacity in orbit. A supplier can retain technical superiority while making customers more willing to subsidize its replacement.

The strongest counterargument is that diplomacy worked and duplicate systems would consume scarce wartime resources. That may preserve today's supplier relationship. The next meaningful test is whether a published service agreement changes who can authorize restrictions or requires advance notice. Another day of normal operation would establish continuity for that day, without answering the control question.

The Contrarian Take

Everyone says: If AI were economically useful, corporate profit margins would already show it.

Here’s why that’s incomplete: A productivity gain and a shareholder gain are different outcomes. A company can save an hour and use it to provide better service; rivals can then force the improvement into lower prices. The customer-support study establishes a bounded output gain, while Apollo identifies the absence of a broad margin gain. Both can be true. The harder investment question is who retains the saving after competition, implementation costs, and supplier bills. Evidence of useful AI should not automatically become evidence for a richly valued AI supplier.

Under the Radar

Ukraine is asking factories for predictability, too. In a separate October 24 industry meeting, the defence ministry emphasized predictable public procurement and realistic assessments of manufacturers’ capacity. The economic problem runs in both directions: buyers need dependable supply, while manufacturers need dependable demand before expanding. This is an announced discussion of purchasing conditions, not a disclosed set of funded orders. Its value will depend on whether future commitments become specific enough for factories to finance production against them.
An AI payment is a low threshold for adoption. Ramp’s methodology counts a business as an adopter when it records a positive payment for an AI product or service during the month. Its sample covers more than 85,000 American businesses, but is not the entire economy; free use and personal accounts can be missed. That makes the index valuable for tracking commercial reach, while leaving the depth and profitability of use unresolved. A subscription receipt cannot answer either question.

Quick Takes

Canada’s tariff relief redistributes the burden. The October 23 White House announcement removes some plywood, paper, machinery, and refrigeration products from the tariff list while adding valves, carpets, waste products, and others, effective October 30. This is selective substitution, not a bilateral settlement. Some importers gain relief while others acquire a new cost; remaining duties can still stack with separate sectoral tariffs. The immediate winners are product-specific, not Canadian exports as a whole. (Source)
Replit’s reported terms price the customer relationship. The Information reports (paywalled) that OpenAI offered Replit over 30% off in a contract exceeding $100 million annually. Those newly reported terms, summarized by AI Weekly, are not publicly confirmed contract disclosures. The implication is bargaining power for an application that can move substantial demand between suppliers. It does not establish that the application passes its savings to users, or that the model provider earns an attractive margin. (Source)
LSEG is buying information that must be gathered. In an October 22 agreement worth catching this weekend, LSEG offered $3.535 billion cash for credit-data provider Octus. The seller’s announcement says external debt will finance the purchase and adjusted earnings per share will be neutral in the first full year after completion. Proprietary credit intelligence can remain valuable as analysis gets cheaper; the acquisition price and financing still decide whether that value reaches LSEG shareholders. (Source)

The Thread

The most revealing test of a technology may be its absence. Ukraine’s restrictions exposed which operations depended on satellite access. AI’s economic debate lacks an equally clean comparison: an aggregate margin chart cannot reveal what the same businesses would have earned without their tools. My inference is that credible comparisons will become a commercial advantage. Vendors able to demonstrate the cost of withdrawal can defend renewal prices; vendors selling impressive usage counts face a harder negotiation. That does not make interruptions desirable or justify deliberately creating dependence. It explains why a customer’s measured avoided loss can support pricing more reliably than a vendor’s claimed potential. The next stage of enterprise AI competition may reward suppliers that can document a customer's counterfactual, the outcome without the service, as clearly as they document its usage.

Predictions

I predict: Ramp’s next monthly AI Index release will keep the median spending of its top 1% of paying firms above $5,000 per employee per month, while the median paying firm remains below $30. The claim is about concentration in this sample, not national adoption or profitability. (Confidence: medium; Check by: 2026-11-30)

2026-10-25 · 03:16 AM ET