AI ROI
Every Monday. What European boards and global investors can verify in reported numbers. Subscribe by email
Correction, 9 November. The version of this issue sent this morning said NAB wrote off A$1.35bn of software “because AI can rebuild it”. That was wrong. The write-down results from a change in NAB’s software capitalisation policy, announced on 20 April 2026 to align the bank with its peers in a period of rapid technological change. It is not an AI cost or an AI saving. The link to AI is a comment by NAB’s chief executive on a longer-term trend, not the reason NAB gives in its filings. The Risk section below has been corrected. Thank you to the reader who pointed it out.
Investment Booking Holdings expects $650m a year of savings from its transformation by the end of 2027, for a one-off cost of less than one year’s savings, and says customers who use its AI assistants book twice as often as those who do not.
Return MercadoLibre grew advertising revenue by 51% with AI selling the ads, while net income grew 6%; a gain in gross margin “was fully offset by higher hosting costs related to AI” and by traffic growth.
Risk NAB wrote down A$1.35bn of its own software, 19% of a year’s cash earnings, after shortening the life it assumes for software to match its peers; it is an accounting change, not an AI cost or saving, though its chief executive has said “the value of software assets is likely diminishing as AI advances”.
Booking Holdings: “remain on track to achieve the full $650 million in annual run-rate savings in 2027”
Booking Holdings, the world’s largest online travel company by revenue, reported third-quarter results on Tuesday 3 November. Its Transformation Program, a cost programme begun in November 2024, now aims to save $650m a year by the end of 2027, equal to 12% of last year’s net income. Ewout Steenbergen, the chief financial officer, said in February that “the aggregate transformation costs will be less than 1x the ultimate annual run rate savings”, in other words less than one year’s savings. Booking does not say how much of the saving, if any, comes from AI; it says the capacity freed is reinvested in “key strategic priorities such as GenAI”.
Booking also reports what AI adds to sales, which few companies do. Customers who use its AI assistants, Priceline’s Penny and Agoda’s AI Assistant, “have converted at approximately twice the rate of non-users”, meaning they go on to book about twice as often. Booking does not say what share of its customers use the assistants, and a doubled rate on a small share moves little. On the risk that customers find their hotel through someone else’s AI, Booking gives a measure: room nights referred by ChatGPT and other AI assistants “still account for significantly less than 1% of Booking.com’s room nights”, though such referrals “have accelerated over the last several weeks”.
Booking’s annual savings target for 2027, for a one-off cost of less than one year’s savings. Customers who use its AI assistants book about twice as often as those who do not.
“Priceline’s Penny and Agoda’s AI Assistant users have converted at approximately twice the rate of non-users.”
Ewout Steenbergen, Chief Financial Officer, Booking Holdings, answering an analyst on the third-quarter results call, 3 November
Ask for the same on your own AI programme: the annual saving and when it arrives, the most it may cost, whether customers who use the AI buy more than those who do not, what share of customers use it, and how many of your customers now reach you through someone else’s AI. Booking has published four of the five, though it does not say how much of the saving is AI. If management can give you only the first, the programme is a cost plan, and the board should say so.
Read Booking Holdings’ third-quarter results →
See also: BT Group [UK, telecoms] put its savings at “£1.7bn at a cost to achieve of £0.8bn”, with AI named as the enabler but given no figure of its own. Schaeffler [Germany, automotive and industrial supplier] will spend about €480m to save about €300m a year in a conventional restructuring, a cost of more than one and a half years’ savings. Zalando [Germany, online fashion] launched Zalando Extended, which lets customers buy from other shops inside its app, with an AI agent handling the order from 2027.
MercadoLibre: “Advertising revenue grew 51% YoY on an FX-neutral basis (58% in USD), well ahead of the market”
MercadoLibre, Latin America’s largest online marketplace, reported on Wednesday 4 November that its advertising revenue grew 51% in the quarter at constant currency, and its letter to shareholders says why: “AI is now central to how we sell ads.” Its AI advisor “now manages campaigns for 56,000 sellers, almost 4x the number three months ago”. Those merchants now have MercadoLibre’s AI running their advertising campaigns on MercadoLibre’s own marketplace.
The same letter shows what that revenue costs. Net income rose 6%. A gain in gross margin “was fully offset by higher hosting costs related to AI and strong growth in traffic across our ecosystem”, and AI investment “grew by almost $30mn” on the previous quarter, with no total given. What MercadoLibre does not show is how much of the advertising growth comes from AI, or what the AI behind it costs. The only link it offers is that the revenue gain from AI in search “more than offsets the incremental cost of the models behind it”, with no number attached.
MercadoLibre’s advertising revenue and its net income, in the same quarter, with AI running campaigns for 56,000 sellers.
“Our AI advisor now manages campaigns for 56,000 sellers, almost 4x the number three months ago.”
MercadoLibre, letter to shareholders, third quarter 2026, 4 November
Ask which revenue lines AI lifted this quarter, by how much, and what the AI behind them cost, in the same table. MercadoLibre shows the revenue growth and says the cost is rising, but not the link between them. If management can show only savings, treat the AI programme as a cost programme until it shows the revenue.
Read MercadoLibre’s letter to shareholders →
See also: Snap [US, social media] reported that “AI search now resolves over 92% of help-center sessions”, with costs almost flat. Expeditors [US, freight forwarding] grew revenue 42% with almost no extra staff, citing demand for air freight from the large cloud companies building AI data centres. Lemonade [US, insurance] credits AI for growing its business with fewer employees, and still made a loss in the quarter.
NAB’s chief executive: “the value of software assets is likely diminishing as AI advances”
National Australia Bank’s full-year results on Thursday 5 November include a A$1.35bn write-down of software the bank built and carries on its balance sheet as an asset, equal to 19% of the previous year’s cash earnings. The write-down comes from a change in accounting policy announced on 20 April: NAB shortened the useful life it assumes for software and raised the threshold above which a project is capitalised, “to align to peer practice” in a period of “rapid technological change”. It is an accounting change. It is not an AI cost, and it is not an AI saving.
The AI question came from an analyst. Asked about the change on the May results call, Andrew Irvine, the chief executive, said he saw “a macro trend” in which “the value of software assets is likely diminishing as AI advances”, because software is becoming faster and cheaper to build or replicate. That is the point for a board. If software becomes cheaper to build, software already paid for and sitting on the balance sheet may be worth less than its carrying value. NAB has also changed how it treats new spend: 48% of technology investment is now expensed as it is incurred, against 38% a year earlier, so the cost of change reaches profit sooner.
NAB’s write-down of its own software, 19% of the previous year’s cash earnings, after it shortened the life it assumes for software to match peers.
“I do think there is a macro trend here that over time, the value of software assets is likely diminishing as AI advances.”
Andrew Irvine, Group Chief Executive Officer, NAB, half-year results call, 4 May, answering an analyst’s question on the policy change, as reported by iTnews and in call transcripts; the write-down is reported in the full-year results of 5 November
Ask your CFO two things: what useful life the company assumes for its capitalised software, and what the balance sheet would look like on NAB’s new assumptions. Then ask where the cost of the AI programme itself sits, in which line, and whether it would be visible to a shareholder reading the accounts. If the answer is a footnote, the board cannot judge whether the programme pays for itself.
Read NAB’s full-year results →
See also: MetLife [US, insurance] disclosed for the first time $4.1bn of exposure to hyperscalers and data centres, under 1% of its general account. ABB [Switzerland, industrial] saw China’s planning regulator block the sale of its robotics division to SoftBank, reportedly over concern that the robots’ operating data would leave the country to train AI models. Ziff Davis [US, digital media, owner of CNET and IGN] booked a $29.6m goodwill impairment and a $28.7m quarterly loss; on the results call, as reported by PPC Land, its chief executive said organic Google referrals fell about 37% in a year and the company is no longer planning on a search recovery.
- Nestlé Vietnam has published 27 AI use cases and invited anyone to solve them. The list is on a Vietnamese venture platform, open to anyone with a working solution, from a single developer to a global vendor, until 22 November. The problems are operational: foreign bodies on sealing lines, trade-spend forecasting, workload balancing in distribution centres, moisture control in coffee drying. Each comes with the measure Nestlé will judge it by; none comes with a budget. I have not seen a large consumer company set out this plainly what it wants AI to do. The question for your own board is whether management could produce a list like this for internal use: every AI use case, the measure it will be judged by and the person who owns it.
- Teleperformance expects sales to fall, partly because of automation. The world’s largest customer-service outsourcer cut its 2026 outlook on 5 November to a fall in sales of 1.5-2.5%, with one business line “reflecting both accelerated automation and the ramp-up of offshore delivery”. The shares fell about 15% the next morning. In Issue 6 Infosys reported slower growth for the same reason; Teleperformance now expects its revenue to shrink.
- “It might not show up in the existing expense ratio.” Asked on 5 November what Allstate’s AI build will cost, Tom Wilson, its chief executive, answered, according to the call transcript published by Motley Fool: “It might not show up in the existing expense ratio. Maybe we will break it out separately. Maybe we will footnote it.” When one of the largest US insurers says the cost of its AI programme may sit in a footnote, any board that tracks its expense ratio should ask where its own AI costs are booked.
- Westpac halved the cost of a data migration with AI. Westpac’s results presentation on 2 November puts the migration of its St.George data warehouse at “initiative cost ~$35m, down from ~$70m”, and says “AI accelerated migration by >2 years”. It was the only project in this week’s results reported with its cost before and after AI, and it is the kind of figure a board can ask for on every legacy migration on its own plan.
- On the balance sheet: What useful life do we assume for our capitalised software, what share of technology spend do we capitalise rather than expense, and if AI lets us or a competitor rebuild that software for a fraction of its cost, is the carrying value still right?
- On margin: Where our products or workflows use AI, what is the full cost of running the AI, computing and model fees included, per transaction, and is that cost growing faster than the gross margin the AI adds?
- On contracts: How much price have our service and outsourcing suppliers conceded this year because of AI, and on the other side, is our own pricing falling faster than we can replace the revenue with volume?
The underlying data. The scan behind this issue read the results, filings and transcripts of about 70 listed companies across the UK, Europe, the US, Australia and Latin America in the second week of the quarter’s reporting season. 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.
The glossary. AI, in board terms. AI terms explained for the people who decide, with what to ask management for under each, including three in this issue: inference, run-rate and capitalisation. Version 1.1 is on the website.
Let’s talk. Board discussions about AI are moving from vendor demos to where the money goes and what it returns. If what you see on your own board or in your portfolio matches or contradicts this issue, 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.