Nearly nine in ten CEOs see value from AI. One in five industrial firms runs it at scale.

The piece of the week, because it puts on record what the CEOs it asked usually say in private. BCG surveyed 152 chief executives of companies with at least $500 million in revenue: nearly nine in ten see cost or revenue benefits from AI in targeted areas, nearly two thirds run pilots, 26 per cent have embedded AI in a broader transformation, and 14 per cent have defined the P&L impact for all their initiatives. BCG’s four moves are organisational, not technical: the CEO orchestrates and the business is accountable, few high-value areas, value tracking before launch, people and change funded in full. Two things to read against the source. "High performers" are defined by outcome, at least 10 per cent cost reduction or 5 per cent revenue growth, so "seven times more likely to redesign workflows" is a correlation among 152 answers, not a measured lever. And the sentence that will be quoted most, 10 per cent of the value from algorithms, 20 from data, 70 from the operating model, opens with "In BCG’s experience". It is a house rule BCG has published since October 2020 and spelled out in 2022 ("Five Rules for Fixing AI in Business"), not a result of this survey. The direction holds, and I would apply it to any technology programme: name the person before you choose the tool.

The denominator the CEO survey does not have. Every twelve weeks the Census Bureau asks a sample of about 1.2 million businesses with employees, roughly 200,000 in each two-week wave, whether they used AI "in any business function" in the past two weeks; in 2023/24 the response rate was about 16 per cent, according to the Bureau’s own working paper, and who declines to answer cannot be weighted away. As of 3 May 2026 the answer was yes for 19.8 per cent. Among firms with 250 or more employees it was 37 per cent, in the information sector 39.7, in finance and insurance 33.9. From December to May the national rate stayed between 17 and 20 per cent. Set next to BCG’s: nearly nine in ten of 152 large companies see some benefit, one in five of all businesses with employees uses the technology at all. The questions differ too. BCG asks whether AI produced a benefit anywhere; the Census asks whether it was in use in the last fortnight, which is the more concrete question, not the higher bar. Whoever sells into the mid-market should carry the Census numbers, not BCG’s.

Same pattern, one industry. Deloitte Switzerland asked more than 140 manufacturing organisations and reports that 84 per cent already generate measurable value from AI, while only around one in five use cases has been scaled consistently across sites or the enterprise. Where the value sits: quality 62 per cent, production 57, logistics and supply chain 49. The barriers respondents name are implementation cost (43), lack of technical expertise (35), resistance to change (35), regulatory and compliance concerns (34) and data availability or quality (30). None of them is called the model; it hides inside three of them, accuracy in the data line, drift in the expertise line, lock-in in the cost line. Twice against the source: "measurable value" is what respondents say about themselves, nothing Deloitte measured, and the study gives a four-way regional split but no survey dates. What survives is the ratio. One in five scaled, in a sample that has nothing to do with BCG’s.

Now the OT side. Cisco surveyed more than 1,000 operational technology decision-makers across 19 countries and 21 industrial sectors. 61 per cent are using AI in live industrial operations; 20 per cent report scaled, mature deployments. 83 per cent plan to increase spending and 87 per cent expect meaningful outcomes within two years, so the spending plans are ahead of the results. The sentence to keep, from the head of Cisco’s industrial IoT business in the release: success "is no longer determined by models alone, but by whether networks, security, and teams are ready to support AI at the edge, in motion, and at scale". Cisco sells the network that sentence recommends, and the release does not hide it, so the list of prerequisites is written from the vendor’s side; the same list from the buyer’s side starts with who owns the use case and puts the network second.

The supply side. IoT Analytics had twelve analysts at Hannover Messe 2026 and logged 29 industrial agentic AI solutions: 72 per cent already commercially available rather than proof of concept, 78 per cent model-agnostic, with Siemens, Schneider Electric, ABB and Rockwell among the sellers. Siemens offers its engineering agent on a fixed subscription per user and year, a seat licence, which grows with headcount, not with sites. The author’s conclusion is that the question has moved from "Can we do this?" to "How do we scale, govern, and charge for it?". Set against the four surveys above, that is the gap of this week in one line: the vendors are at pricing, the buyers are at one in five. A sample of 29 products at one fair is small; the piece offers it as an observation, not a survey. It is still the only number in this list that counts products in a hall instead of answers on a form.