AI ROI
Every Monday. What European boards and global investors can verify in reported numbers. Subscribe by email
WTW, run from London, has given the market the clearest account yet of what an AI programme costs and what it is meant to return: $650m to secure $400m of annual savings. Notably, most of the money spent so far has gone on restructuring. Santander Bank put a figure on AI in its results, €194m of value in nine months (0.4% of its income) in a week when its European peers all talked about AI and none of them put a number on it. Reddit grew revenue by 53% and warned, in the same letter, that it is losing more than a million search referrals a week because Google’s AI now answers the question before anyone clicks through. Reddit expects to lose more than a tenth of its 130 million daily users this quarter as a result.
WTW sets the benchmark: $625m in cash to unlock $400m in annual run-rate savings
Most companies announce an AI programme with a productivity story and a single cost figure. WTW did something different with its third-quarter results on Thursday 29 October: it gave the market the full arithmetic behind Propel, its AI Acceleration Plan, which runs to the end of 2028. To reach roughly $400m of run-rate savings it expects to spend about $625m in cash and $25m in non-cash charges, which it calls “a cash-cost-to-achieve ratio of approximately 1.6 times”. Read the other way round, once the savings reach full run-rate the programme pays back its cash cost in about 19 months. After reinvesting about $50m in growth, WTW expects $350m of net savings and an adjusted operating margin of about 30% in 2028; on the net figure the payback is about 22 months.
Some context is worth noting. The $650m is 6.7% of last year’s revenue, and the $400m target is 25% of last year’s net income. Spread across WTW’s 47,000 employees, the company is spending about $13,800 a head to save about $8,500 a head every year from 2028. The detail an audit committee would look at is where the early money has gone: of the $21m spent to 30 September, $14m was restructuring charges, which in most companies means people. Annualised savings so far stand at $18m, so one quarter in the cost-to-achieve ratio is running at about 1.2 times, better than planned, on numbers too small yet to prove the case.
WTW’s target cash cost to achieve each dollar of annual saving: $625m in cash for $400m of run-rate savings by 2028, reported against every quarter.
Hardly anyone reports this way. Of the ca. 90 result statements reviewed this week, WTW’s was the only one to give the cost, the run-rate target and the ratio between them, and then to report every quarter how much has been spent and saved so far. HSBC, by contrast, said it had taken $1.6bn of restructuring charges since the start of 2025 for about $1.9bn of annualised savings, and its chief executive, Georges Elhedery, said AI “will change how work is done across industries, impacting some jobs, transforming others”, without attaching a dollar of the saving to it. WPP, whose Elevate28 plan is meant to deliver £250m of gross run-rate savings by the end of this year for about £190m of cash restructuring costs, said on the same day that its growth is “powered by our agentic marketing platform WPP Open”, and gave no figure for it.
“While Propel, our AI Acceleration Plan, is in its early days, I’m excited by the initial impact we’re seeing as we scale AI to enhance client service.”
Carl Hess, Chief Executive Officer, WTW, third-quarter results, 29 October
If management is asking for an AI budget, ask for the four numbers WTW has published: the cash going in, the run-rate saving coming out, the ratio between them, and the date the saving is due. Then ask the fifth, which WTW has not answered: how much of the saving comes from automation and how much from headcount. If management can only give you the first, the business case is not finished.
See also: Guidewire Software agreed to buy London-based hyperexponential for $1.7bn in cash plus up to $400m in earn-outs, for underwriting software used on more than $75bn of commercial P&C premium at insurers including Aviva, Allianz and Markel. GE Vernova, the US power equipment maker spun out of General Electric in 2024 (revenue $38.1bn, 2025), is buying Trench Group, a Berlin-based maker of high-voltage grid parts, from the private equity firm Triton for $9.0bn. That is 13 times Trench’s expected 2028 earnings, and GE Vernova names data-centre demand among the reasons for paying it. Put simply, the AI build is now setting the price at which private equity sells grid suppliers. American Electric Power added $22bn to its five-year plan, now $100bn, after signing 15 GW of new load in one quarter, and makes data-centre customers sign take-or-pay contracts, worth up to $21bn in fixed cost offsets, before it builds.
Santander isolates €194m in AI value, while European peers leave the ledger blank
Santander has started reporting AI like any other item in its results: a single figure for the value it produced, a definition of what counts, and a target to measure it against. In its nine-month results on Tuesday 28 October the bank said its “Data & AI initiatives” generated €194m of business value, which it defines as incremental revenue, cost savings and lower loan-loss provisions. For a bank that earned €46.8bn of income over the same nine months, €194m is modest, 0.4%, or about €1,040 per employee. The target it is measured against is more than €1bn a year by 2028, on a run-rate basis against 2025, so the two figures are not on the same footing and the run-rate is the number to ask for.
In a week when much of corporate Europe reported, Santander was the only one to print a figure. Volkswagen, Shell, Airbus, Sanofi, Unilever, Lloyds, Deutsche Bank and BBVA all reported without putting a number on AI. BBVA reported costs of 37.9 cents for every euro of income, one of the lowest ratios among European banks, and said AI is central to its plan, but did not say how much of that efficiency, if any, came from AI. Deutsche Bank named AI as one of the levers behind €0.5bn of operating efficiencies this year, with no breakdown. Where banks and insurers did give numbers, they came from the same three places: customer service, fraud and code. NatWest’s AI fraud agent has handled 52,000 credit-card conversations, and staff use of AI “rose 66% in the quarter”. UBS counts 608 AI applications in production and 12,700 “AI personal agents” used by staff, against 6,800 fewer employees than a year ago, and does not put a money figure on any of them.
Santander’s nine-month AI business value, €194m, as a share of its income in the same period, €46.8bn. The 2028 target is more than €1bn a year.
“The group also continued to deploy AI across its businesses at scale. This generated €194 million of business value (higher revenue, lower costs and lower loan-loss provisions) during the year through better customer service, higher productivity and commercial growth.”
Banco Santander, nine-month results press release, 28 October
Santander deserves credit for reporting a return at all, but the attribution needs testing. Lower loan-loss provisions in a benign credit year are hard to separate from the credit cycle itself. And a return needs a denominator: ask management not only what value AI produced, but what the cloud, software and consultancy behind it cost.
See also: Capgemini raised its full-year revenue growth guidance to 9.0-9.5%, with UK and Irish bookings up 25.2%, as clients pay consultants to sort out their data before AI can return anything.
Reddit’s warning: 53% revenue growth collides with the zero-click search trap
By the usual measures Reddit had an exceptional third quarter. Revenue rose 53% to $895m, net income reached $283m, and global daily active users grew 12% to 130.3 million. The letter to shareholders of Thursday 29 October then carried a warning that wiped out 23% of its market value the next morning: total audience could fall by a low-double-digit percentage in the fourth quarter. The primary catalyst is Google, whose AI Overviews and AI Mode now answer queries directly on the search results page, without an outbound click. During October Reddit lost more than a million search referrals a week on average, and US daily users fell from 53.2 million in Q2 to 50.7 million, the first domestic contraction since the company listed in 2024.
The warning for boards sits in how the trap was built. First came the licensing cheque: Reddit sold Google access to its forum data for AI training. Its non-advertising revenue, which includes that licensing, was $48m in the quarter, about 5% of the total. Then came the cannibalisation: Google now answers on its own page many of the questions that used to send people to the forum, while, in the words of Steve Huffman, Reddit’s chief executive, it is still “crawling us as hard as ever”. As he conceded on the call: “the AI mode does not drive traffic, not for us and not to anybody.” Finally came the retreat: with logged-out search visitors falling away, Reddit will test a logged-in-only experience in selected markets outside the US this quarter, because logged-in users “make up over 80% of monetization”. Reach, publisher of the Mirror, lost 20.3% of its digital revenue when Google referral traffic dropped (Issue 5). Reddit shows the wider dilemma: selling your data to AI models creates a revenue line today, while feeding the machine that takes away your customer acquisition tomorrow.
Search referrals Reddit lost each week in October, on average, as Google’s AI answered queries directly on-page without sending users through.
“During October, we have seen that search traffic has dropped by more than 1 million referrals on average per week. If this rate of decline continues, current trends would suggest that total DAUqs could fall in Q4, by a low-double-digit percentage from our current base of 130 million DAUqs.”
Steve Huffman, CEO, and Drew Vollero, CFO, Reddit letter to shareholders, 29 October
Ask your commercial directors how many customer journeys begin through an intermediate search or aggregator, and what happens when an AI answers the query on-page without sending the user to your domain. If your company is selling proprietary data or product feeds to AI model developers, ask the follow-up: are you generating licensing income, or are you financing the tool that will disintermediate your primary sales funnel?
Read Reddit’s letter to shareholders →
See also: The Bank for International Settlements counted $222bn of bonds issued by the five largest hyperscalers by September, more than 10% of long-term US Treasury issuance this year, and $1.2tn of data-centre leases, “around $820 billion” of them not yet started. Finnish police opened a criminal investigation into Google’s $15bn Finnish data-centre programme at Muhos, halting site works over suspected breaches of the Water Act; in Europe the constraint on the AI build is as often a permit as a power line.
- The benchmark European insurers are missing. European insurers said little about AI this week. Ping An, in China, reported it as a line item: RMB10.04bn (about £1.1bn) of claims avoided by its fraud detection in nine months, equal to 3.8% of its property and casualty premium income, 82% of customer service handled by AI, and an 81% fall in its cost per token. That is the level of disclosure a European board could ask its own management for.
- Faster cars, fewer people. Mercedes-Benz announced PLM.AI, which it expects to shorten vehicle development times by up to 30%, in the same release as a renewed voluntary severance programme in Germany and €1.1bn of restructuring cash outflows in nine months. Mercedes has not yet said what the time saved is worth in money.
- A software engineer on the board. Novartis will propose Rajesh Jha, who spent 36 years at Microsoft and, in Novartis’s words, “played a key role in the integration of artificial intelligence across Microsoft’s product portfolio”, for election to its board at the annual meeting on 5 March 2027. Few large-company boards include someone who has built the tools they are asked to fund. Novartis is adding one, and it is worth asking who on your own board could do the same.
- Grid first, planning second. In the 28 October UK Budget the Treasury offered priority grid connections and fast-tracked planning for data centres it deems strategically critical, in exchange for binding local investment. It follows Microsoft’s withdrawal from a planned Nscale data centre in Essex after grid connection delays pushed power delivery into the 2030s. In Britain the thing holding up the build is the connection, not the capital.
- 45 days to explain the price. The US Federal Trade Commission issued Section 6(b) orders to nine companies, among them Delta, United, Uber, Lyft and DoorDash, requiring full documentation of their automated pricing within 45 days. Its draft policy says undisclosed personalised pricing could violate the FTC Act. Worth reading for any European retailer pricing dynamically under the AI Act and GDPR.
- On capital and restructuring: For our AI programmes, what is the cash cost to achieve each pound or euro of run-rate saving, and how much of the saving comes from AI itself rather than from job cuts we would have made anyway?
- On attribution and cost: When management reports AI “business value” to the board, what is the fully loaded cost behind it (cloud, inference, integration), and can we show that lower provisions or better margins are AI and not the credit cycle?
- On search and disintermediation: What proportion of our customers reach us through someone else’s search or comparison engine, and if an AI answers their question before the click and cuts us out of the exchange, what share of our revenue comes from customers we hold directly?
The underlying data. The scan behind this issue read the third-quarter statements, filings and transcripts of about 90 listed companies across the UK, Europe, the US and Asia in the busiest reporting week of the quarter. Reply if you would like the full list.
What I am building. Accelr8™ helps a business recover enterprise value and reduce risk. It checks what the business has committed, in contracts, plans and entitlements, against what actually happens in invoices, results and use, across six Value Pools, from Revenue Quality to Margin Integrity to Capital Velocity, and shows the board where value is lost, what has been recovered, and the evidence for both. All three questions above are the kind it answers from a company’s own records. Patent pending; a founding pilot is open.
Let’s talk. The boardroom conversation about AI is moving on from vendor demos to capital allocation and attribution, and we learn more from comparing what companies file than from reading conference brochures. If what you see on your own board or in your portfolio matches or contradicts this, I would like to hear it.
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.