In R&D, a drugmaker rented AI for about a day's research budget, while a Japanese factory group paid 14% of a year's sales to buy it. A health group cut its clinicians' time on patient records by 33% with AI. A publisher lost 20.3% of its digital revenue in a quarter as Google sent fewer readers.
Investment, in R&D: rent, build or buy. On Thursday 15 October AstraZeneca agreed to pay insitro, an AI drug discovery company, $40m now and up to $1.46bn more if the work succeeds. The same day Recursion, which builds its own AI to find drugs, sold $200m of new shares. Building that AI is costly: Recursion spent $475m on research last year, more than six times its revenue. Omron of Japan chose to buy instead, paying $675m, about 14% of a year's sales, for a US maker of industrial computers.
Return, in the back office. UnitedHealth Group says AI has cut the time its clinicians spend reviewing patient records by 33%. It is spending nearly $1.5bn on AI this year, but gives no figure for the return.
Risk, at the customer end. Reach, publisher of the Mirror and the Express, lost 20.3% of its digital revenue in the third quarter as Google sent it fewer readers.
"AstraZeneca announces new strategic collaboration with insitro to use advanced AI systems to develop transformative therapeutics for difficult to treat cardiometabolic diseases"
On Thursday 15 October AstraZeneca agreed to work with insitro, an AI drug discovery company, on new medicines for a fatty liver disease and for diseases of the heart muscle. insitro receives "an upfront payment of $40 million" and up to about $1.46bn more if the work reaches agreed stages, plus royalties on sales.
AstraZeneca proposes the targets, the proteins in the body a medicine is designed to act on. insitro's AI tests whether each one causes the disease and works towards a drug. AstraZeneca runs the clinical trials and owns the worldwide rights.
For AstraZeneca the sum is small. It spent $14.2bn on R&D in 2025, about $39m a day, so the $40m up front buys roughly one day of its research. Even if every payment is made, the total of about $1.5bn would equal about five and a half weeks of its R&D, spread over many years. The large payments come only if the work succeeds: development milestones as the drugs pass each stage of testing, then sales milestones and royalties once they are on the market, paid out of the revenue those drugs bring in. If the programmes fail, AstraZeneca's cost stays close to the $40m.
The AI companies on the other side carry the cost. Recursion builds its own AI to find drugs and spent $475m on research in 2025, more than six times its revenue of $75m, and lost $645m. On 15 October it sold $200m of new shares, adding about 11% to its share count, and said its cash now lasts "into 2029". The same day Sanofi paid Recursion $10m for reaching an agreed step in their partnership, roughly one week of Recursion's research spending.
So there are three ways to pay for AI in R&D: rent, build or buy. Omron, the Japanese factory automation group, chose to buy. On 13 October it agreed to pay $675m in cash, Reuters reports, for OnLogic, a US maker of industrial computers, to "build a platform for Physical AI": AI that runs machines on the factory floor. That is about 14% of Omron's yearly sales of ¥767.4bn ($4.85bn). Omron said it had concluded that going further than an alliance would let it move faster. Where AstraZeneca rents for a day's research budget, Omron buys for seven weeks of sales.
In R&D, pay for AI as AstraZeneca does: a small sum up front, the rest only when your own people can test a result. Before you sign, check how long the supplier's cash will last.
See also: Vir Biotechnology Enters Multi-Program Collaboration and License Agreement with Gilead Sciences [Gilead Sciences, US, pharmaceuticals, revenue $29.4bn, 2025]: $90m up front, $25m of it in shares, and up to $1.74bn more for three cancer drug candidates; Third quarter 2026 results [TomTom, Netherlands, maps and software for carmakers, revenue €129m in the quarter]: reorganising around automated driving, where "AI is helping us innovate faster", affecting about 200 roles, as automotive revenue fell to €68m from €85m; Lantmännen: AI can strengthen European agriculture and improve farmers' profitability [Lantmännen, Sweden, farming and food, turnover SEK 60bn, about €5.4bn]: SEK 100m, about €9m, for a new plant-breeding facility, where "AI already plays an important role".
"We have expanded AI-enabled chart summarization to more than 40% of our employed clinicians, reducing chart review time by 33%"
With its third-quarter results on Tuesday 13 October, UnitedHealth Group gave two figures for AI in its administration. AI summaries of patient records now reach more than 40% of the clinicians it employs and have cut the time they spend reviewing records by 33%. When a patient or doctor appeals against a refused claim or treatment, AI now gathers the records and rules needed to decide, and has cut the time to resolve appeals by 26%, "and we expect this to accelerate in 2027".
The return UnitedHealth reports is time, not money. Cutting 33% of the time spent on patient records and 26% of the time spent on appeals are real gains, but they turn into money only if the same staff handle more patients and appeals, or fewer staff are needed. The company does not say which. The spending, by contrast, is in money: in his letter to shareholders in April, the chief executive, Steve Hemsley, said UnitedHealth expects "to invest nearly $1.5 billion in these efforts in 2026 and at least that much in 2027", meaning what he called "intelligent technologies". That is about 12% of last year's net earnings of $12.1bn. Bloomberg reported in June that executives say they see a 2-to-1 return, a figure the company has not published itself. Operating costs fell to 12.8% of revenue from 13.5% a year earlier, worth about $790m in the quarter, which shows how much a small efficiency gain is worth at this size, but the company credits efficiency in general, not AI.
UnitedHealth does not say who built the tools, but it is worth noting that its own technology arm, Optum Insight, develops AI and sells it to other health companies, and works with Microsoft, Google and Anthropic on terms it does not disclose. So the return can come twice: as savings inside the group and as sales to others.
"Across the businesses, AI-enabled capabilities are beginning to help clinicians spend more time with patients, improve service speed and experience, streamline administration and reduce friction for both consumers and clinicians."
Steve Hemsley, Chairman and Chief Executive Officer, UnitedHealth Group, 13 October
In the back office, ask three questions about each AI tool: who uses it, how much time it saves, and what that time is worth in money. UnitedHealth has published what it spends and answered the first two, but not the third.
See also: Bank of America third-quarter earnings call transcript [Bank of America, US, banking, net income $30.5bn, 2025]: expects AI benefits of $700m next year, 2.3% of last year's net income; Interim Results [Vertu Motors, UK, car retail, revenue £4.8bn, year to February 2026]: in a pilot at 18 dealerships, an AI tool called Opportune, which analyses customers' phone enquiries, lifted sales conversion to 24.8%, against 22.0% at comparable dealerships without it.
"Q3 digital trading continues to see the impact of ongoing industry-wide disruption in referrals, primarily from Google."
Reach publishes the Mirror, the Express and the Manchester Evening News. In its trading update on Tuesday 13 October it reported digital revenue down 20.3% in the third quarter, because Google sends it fewer readers.
Reach saw this coming. Back in March, its annual report listed among its principal risks a fall in digital audience that "could be caused by changes in major platforms' support and referrals to our content, changes to search and disruption from AI". The risk has a named owner, the Chief Digital Publisher, and was marked as increasing. Traffic from Google in the second half of 2025 was 46% lower than a year earlier, and Reach wrote down the value of its business by £222.8m, citing lower digital revenue expectations "attributable to the decline in referral traffic". Digital revenue was £128.9m last year, a quarter of the total.
Reach is not trying to win the Google readers back; it expects the audience on its own websites "will not see a recovery to its former peaks". Instead it is building revenue that does not depend on search. It has launched digital subscriptions at six titles, including the Express and the Manchester Evening News, "targeting in excess of 75,000 subscribers in 2026", and is moving into video, with page views on other platforms up 20% in the second half of 2025. The October update gives no subscriber number.
AI licensing is one more piece: a deal with Amazon's cloud business that "opens up our content to support answers Alexa gives to users", paid each time it is used, and now "a recently secured agreement with a major tech platform", which Reach does not name. It gives a value for neither, so readers cannot tell how much of the lost revenue they replace. Meanwhile it is cutting costs, at a cash cost of £35m to £40m this year, 33% to 38% of last year's adjusted operating profit.
"We are making strategic progress in an evolving digital landscape with a focus on original content, distinctive brands, and revenue diversification."
Piers North, Chief Executive, Reach, 13 October
At the customer end, Reach had the risk on its list and still lost 20.3% of its digital revenue in a quarter. It now plans on the basis that those readers will not come back, and is building revenue it controls itself. Ask how much of your revenue depends on someone else's search or AI answers, and what you are building that does not.
See also: The Getty Family and Koch Equity Development Commit to Equity Investment of up to $533 Million in Getty Images [Getty Images, US, image licensing, revenue $981m, 2025]: a court-supervised restructuring, while AI licensing brought in 4.4% of last year's revenue; CoreWeave bond rout tests market's math on AI buildout [CoreWeave, US, AI cloud, revenue $5.13bn, 2025]: Reuters reports its bonds now yield about 15%, a sign of what lenders charge for AI built on debt.
Bookings through AI search engines rose 70% in the first half. Unlike a news reader, a hotel guest still has to book the room. Whitbread
Plans $400m more for AI factory capacity in China and Malaysia, with no timeline. That is more than the $292m of cash its operations brought in last year, and it borrowed $199m in the first half. TTM
Microsoft will pay for all the power plants and lines its Indiana data centre needs, under a 20-year deal. Other utilities will be asked for the same terms. NiSource
The CMA proposes special status for Microsoft's business software and will check whether rival AI tools can compete with Copilot; Reuters expects a decision by February. Before you renew, check what your Microsoft agreement says about switching. CMA
Three questions for your next board meeting.
- R&D: for our most important AI capability, are we renting, building or buying it, and how much do we pay before the first result?
- Back office: for each AI tool in administration, how much time does it save on which task, and what is that worth in money?
- Customer: how much of our revenue depends on someone else's search or AI answers, and what revenue are we building that does not?
Ilona
Wider this week.
This issue drew on a scan of 33 industries and 17 regions. Reply and tell me which item was most useful, and which company I should watch.
AI, in board terms.
AI terms explained for the people who decide, with what to ask management for under each. Version 1.1 is on the website.
Let's talk.
I am interested in connecting technology with the P&L. Nine months ago I would have said AI was like teenage sex: everyone talking about it, hardly anyone doing it. Today we can see the divide between the companies that are doing it and those still talking. I believe we learn faster together if we go beyond surveys and look under the hood at what companies report in their own numbers. That is what this newsletter tries to do. If a question in this issue is on your board's agenda, or you see it differently, write to me.
The content is checked against the companies’ own filings before publishing, but mistakes can slip through. If you spot one, please let us know and it will be corrected in the next issue. For information only, not investment, legal or tax advice. Published by Agent & Capital Ltd, 71-75 Shelton Street, Covent Garden, London WC2H 9JQ. Subscribe here.
Get AI ROI by email every Monday at 07:00 UK time. One story each on investment, return and risk. Unsubscribe in one click.