Fourteen surveys and research reports hit the wire on October 22. Most were marketing with a sample size attached. A few carry real numbers. Read together, the clearest finding comes from almost all of them: companies are deploying AI faster than they can say what it does for them, and the one hard market number of the day shows the AI build-out already squeezing everything else.
Memory costs hit the phone market
Omdia’s preliminary count has global smartphone shipments down 20% year on year in the third quarter of 2026, the weakest third quarter since 2013. In Q2 the decline was 6%. The cause is memory. DRAM and NAND prices keep rising as AI servers take the supply, and that leaves phone makers no room on cheap models, so they’ve cut the low-margin entry and mass-market devices first, especially in price-sensitive emerging markets. Samsung led with 23% share and Apple followed at 22%. Together they now hold 45% of the market, about 8 points more than a year ago. Xiaomi had 12%, OPPO (with realme and OnePlus) 9% and vivo 8%. Omdia doesn’t expect fourth-quarter launches and holiday demand to make up the gap.
For memory investors this is the other side of the AI trade. Pricing power for Samsung, SK hynix and Micron shows up as unit losses for the Android brands that live on thin margins. The premium vendors are taking share precisely because their buyers can absorb a higher price.
The AI trust problem has numbers now
Israeli startup Hirundo tested Alibaba’s Qwen3-Coder-30B by asking it to build a web service with sign-up, login and secret storage for four different stated clients. It checked 1,840 resulting apps with Semgrep and ETH Zurich’s BaxBench. Apps built for a U.S. defense contractor were hackable 70% of the time, against 59% for a Chinese one. Apps for a Hong Kong democracy group leaked secrets 47% of the time, against 37% for a neutral football club. Adding an explicit security instruction made the U.S. case worse, at 74%. The same access-control flaws showed up when the model ran locally, which cuts against the common argument that self-hosting a Chinese open model removes the risk. Hirundo says the bias sits in the weights and that its unlearning process removed it. CrowdStrike and Booz Allen Hamilton reported similar patterns earlier.
PAN Communications tracked 31,500 answers from ChatGPT, Claude, Gemini and Perplexity to questions posed as nine kinds of executive over four weeks. Answers to the same question changed substantially 50% to 60% of the time, from 39% for HR chiefs to 70% for chief medical officers. Only 42% of sources cited in week one were still cited in week four. Authoritative sites held up much better: NIH kept 76% and Deloitte 68%, against 27% for Medium and 21% for Reddit. PAN sells AI visibility services, so read it as a pitch with data attached. The data is still useful.
FIS surveyed 100 finance directors at UK companies with more than £250 million in revenue. 90% say they have a clear business case for AI in cash and liquidity management, yet 29% can’t measure what it delivers. Their teams still spend about 16.2 hours a week gathering and reconciling cash data, and 65% admit investment decisions are sometimes made without a full view of available liquidity.
Hitachi’s survey of more than 200 industrial decision-makers in energy, manufacturing and transport finds 97% using physical AI somewhere, but only 8% at the stage of continuous, repeatable scale. Three quarters say leadership backs AI, while 38% say they have the resources to adopt it. Limited AI expertise is the top barrier for 43%. Korn Ferry’s talent outlook lands in the same place, naming AI and capability gaps as the force that will reshape hiring in 2027. Life sciences firms are reported to be scaling “trusted” AI decision systems. The word that keeps coming back across these reports is trust, and the number that keeps not coming back is return.
Work, money and what people actually do
Data from 8am, drawn from invoices at more than 60,000 mostly small U.S. law firms, shows the billable hour isn’t dying. 47.7% of firms billed only by the hour in the first half of 2026, against 10.3% that billed only flat fees, a gap of more than four to one. Larger firms billed nearly 12% more hours while logging slightly fewer, and earnings per billed hour rose 3% to 5% across firm sizes. If AI is making lawyers faster, so far they’re keeping the time.
The IHL Group and Sensormatic study of 915 retail decision-makers puts external theft and organized retail crime at 25% of loss. That makes it the biggest single source, but far from the majority. Supplier fraud, warehouse shrink and receiving errors add up to 33%, return fraud rose for 68% of respondents, and an admin error costs about as much per incident as employee theft. The shoplifting story is real for big chains: 71% of retailers with more than $1 billion in revenue saw it rise, against 28% of the smallest. The bigger money leaks out through process.
Toast’s survey of 702 U.S. restaurant workers found pay satisfaction at 67%, up 12 points from a year earlier, and benefits satisfaction at 62%, up 16. 42% of tipped staff said tips rose. Only about one in five is extremely open to using AI at work, and about one in ten rejects it outright.
The rest were thinner. 81.5% of U.S. marketers told one survey that mindset-based advertising can beat identity-based targeting, a useful line for anyone selling a cookieless future. TOURISE and LIONS Advisory argued that creative identity gives tourism destinations a commercial edge. Another report found gamers using their avatars as real-world style inspiration. Two Canadian surveys covered small business digital banking, where customers still want better personal service, and the admin load after a family death, which more than half of Canadians end up handling.
Strip out the vendor pitches and two numbers stand. Phones fell 20% because memory went to AI. And close to a third of UK finance chiefs backing AI can’t say what it returns.
That gap won’t stay open forever.