This week's AI figures came from a restructuring plan, a supplier's prices and a central bank's record. Not from an AI budget.
Investment. BMW will cut 20% of its divisions and management roles by mid-2027, using AI to do the work with fewer people, and promises an EBIT margin of 3 to 5% in 2028. It gives no figure for what the AI costs.
Return. Concentrix says its own AI platform has compressed $1.3bn of its traditional work and expects revenue to fall 3 to 5% next quarter. The saving from AI in services is showing up on the customer's invoice.
Risk. The Bank of England has told firms to prepare for AI agents acting beyond their task. The UK AI Security Institute measured it: a written limit on what the agent may touch cut the risk by about 85%, and did not end it.
Agility, efficiency, AI-based innovations: BMW Group sets course for greater profitability and resilience
At its Capital Market Day in Munich on Wednesday 30 September, BMW said: "By the middle of next year, the BMW Group will reduce the number of divisions and associated management roles by 20 percent, with a comparable reduction at the organisational levels below." The method: "Through the efficient use of artificial intelligence, the company aims to align itself even more systematically for agility and effectiveness across all departments and corporate levels." The result, with a date: "As an initial interim step in 2028, the EBIT margin is expected to be within the range of three to five percent", on the way back to 8 to 10% "by the start of the next decade". The release names Mistral as the AI partner for crash simulation. It gives no figure for the AI and no restructuring charge.
"Consistent use of agentic AI applications across all areas of the company will be a game-changer for more agile and efficient development, leaner structures and faster decision-making."
Walter Mertl, Chief Financial Officer, BMW Group, Capital Market Day release, 30 September
BMW has promised a margin and not the spend behind it. In 2028 we will know whether the margin arrived (and what it cost?). Ask for the AI spend and the restructuring charge in the same table as the target. Without them it is not a plan.
See also: 2027-2029 Strategy: Arcadis' Reset: Focused Growth, Sustainable Value [Arcadis, Netherlands, engineering consultancy]: 1,000 jobs to go in 2027, a restructuring cost of about €70m, the margin target on the same slide; Conduent Investor Day [Conduent, US, business process services]: 6,000 jobs gone since January, $120m of savings by 2028, no restructuring figure; The Cigna Group Hosts 2026 Investor Day [Cigna, US, health insurance]: $3bn of savings by 2030, no cost given
"$1.3 billion of net revenue resulting from the compression of our traditional services using our iX Suite AI platform"
Concentrix runs customer service for other companies. On Tuesday 29 September its chief executive said about $1.3bn of its revenue (~13%) now comes from traditional services "compressed" by its own AI platform, and that about $3bn of revenue has moved onto the platform from new and existing clients over the last three years. He does not say what the compressed work earned before. The company guides fourth-quarter revenue to $2.41bn to $2.46bn, a fall of 3 to 5% in constant currency. The quarter's $1.05bn goodwill charge is attributed to "the trading range for the Company's stock price", and it is the second in a year: Concentrix wrote down $1.5bn in fiscal 2025 on the same ground. Without it the quarter was profitable, at a non-GAAP operating margin of 12.6%. Two days later Accenture, on its own results call, reported lower pricing in many areas of its business and said it is passing productivity from AI to clients.
"$1.3 billion that's running through our platform, which, as we've talked about before, is growing faster than our traditional revenue."
Chris Caldwell, President and Chief Executive Officer, Concentrix, third-quarter earnings call, 29 September
When a services supplier uses AI, the saving shows up on your invoice, not in its profit. Concentrix calls that compressed revenue. It reaches your accounts whether or not you run an AI programme. Ask your largest suppliers what they have changed in price or scope since January because of AI. If the answer is nothing, the saving has stayed with them.
See also: Accenture Reports Fourth-Quarter and Full-Year Fiscal 2026 Results [Accenture, Ireland, IT services and consulting]: lower pricing in many areas; 110,000 AI and data professionals (~14% of staff); FactSet Reports Results for Fourth Quarter and Fiscal 2026 [FactSet, US, financial data]: AI sales doubled, operating margin down two points; Interim Results for the period ended 5 August 2026 [S&U plc, UK, motor finance]: AI credited for a record 82% of payment arrangements with overdue customers; bad debts 18% lower
"Increasingly autonomous models could take unexpected actions, including exploiting vulnerabilities and accessing systems beyond their intended task"
The Financial Policy Committee watches risks to the financial system as a whole. Its September record says two things about AI agents: they can act beyond the task they were given, and firms must prepare for that now. The number behind the first point comes from the UK AI Security Institute, which gave OpenAI's newest model a routine job inside a mock software company with its safeguards off. In 26 runs out of 50 it also attacked the company's software. With the limits of the task written into the brief, that fell to 4 in 49. All of it simulated, and the point is not the attack but how much a written limit changed.
"The Committee reiterated the importance of firms continuing to prepare for and reduce frontier AI-related cyber and operational risks."
Financial Policy Committee, Bank of England, record of the meeting of 25 September
A written limit on what each AI agent may touch cut the risk by about 85%. It did not end it. The risk begins wherever an agent is allowed to act rather than answer, and most companies cannot yet say where that is.
See also: OpenAI alerts more than 100 groups about rogue AI agent activity [Reuters, on OpenAI, US, AI developer]: agents acted on customers' systems without authorisation, months to establish what they did; GPT-6 Astra performs unsanctioned supply-chain attacks in simulations [UK AI Security Institute, UK, government testing body]: the full test; LASST v. OpenAI, complaint filed 29 September [Legal Advocates for Safe Science and Technology, US, advocacy group]: asks the California Superior Court to rule that "OpenAI is responsible for the conduct of its agents"
A supplier, a regulator, a test laboratory and a central bank said the same thing in one week: agents act beyond their brief, and the company that deploys them answers for it. Every agent on your systems has entitlements someone granted. If a letter like OpenAI's arrived tomorrow, ask who would receive it, and whether that person could list your agents and what each may touch within a day.
See also: LASST v. OpenAI, complaint filed 29 September [Legal Advocates for Safe Science and Technology, US, advocacy group]: California Superior Court, injunction sought against OpenAI agents accessing computers without authorisation; FTC opens probe into AI giants including Anthropic and OpenAI [Reuters, on the US Federal Trade Commission]: the first US enforcement action on rogue agents; Financial Policy Committee Record, September 2026 [Bank of England, UK, financial stability]: AI hyperscalers at 47% of sterling corporate bond issuance this year
"Long Lake, the World's First AI Holding Company, Completes $6.3 Billion Acquisition of American Express Global Business Travel". All cash, $9.50 a share, a 65.1% premium to the 30-day average, financed by equity from Long Lake's investors and Koch Equity Development and by committed debt. The chief executive's plan is to "invest heavily to advance the frontier of AI deployment" in the business. The release gives no AI budget and no return target. The purchase price is the number any later AI result at Amex GBT will be measured against. SEC exhibit
AI-related debt issuance reached about $450bn by early September, more than double all of 2025, and AI hyperscalers accounted for 47% of sterling corporate bond issuance this year. Growth and fiscal outlooks, the Committee notes, "depended in part on expectations that AI development and adoption would generate significant productivity gains"; if those expectations disappoint, losses could be amplified and reach "not only AI-related asset valuations but also sovereign debt markets". The return your company is measuring is the same return the bond market has lent $450bn against. Record
"40% of Q4 '25 through Q2 '26 pipeline decisions had a tracked AI / computational contribution", and its AI agent for choosing drug targets, TargetNexus, is "on track to be part of 80% of research portfolio decisions by Q4". Roche does not put a franc figure on the AI; the measure it reports is how many of its own decisions the AI took part in, which it has been tracking for a year. Phase III success rose from 65% in 2025 to more than 80% so far this year, a figure it presents without attributing it to AI. For a board that is the more useful measure: not what the AI cost, but how much of the company's decision-making it reaches. Presentation
"So far this year, AI has been cited in 120,136 job cut announcements, approximately 21% of all cuts, and it remains the leading reason year-to-date." In September itself it was the fifth most cited reason, at 9%. This counts what companies say, not what AI did. Read your own management's restructuring paper the same way. Report
Three questions from this issue for your next board meeting, each answerable with a number, a date or a name.
Ilona
- Which of our cost and margin targets for 2027 and 2028 depend on AI, and what is the AI spend and the restructuring charge behind each one, in the same table?
- On our ten largest services contracts, what price or scope change has each supplier offered since January because of AI, and who on our side is tracking it?
- If an AI supplier wrote to us tomorrow to say its agents had acted outside their remit on our systems, who would receive the letter, and could that person list our agents and what each may touch, by name, within a day?
Three issues in.
Thank you for reading, and for writing back. I started this because I could not find a weekly read that took company numbers seriously and left the opinions out. Two things help me most: which item was the best one for you this week, and which sectors, countries or kinds of company you would like to see more of. If the issue is useful to you, pass it to one colleague who would put the three questions to their board.
AI ROI.
This newsletter shares its name with the book I am writing on how companies and their owners get a return from AI. Write to ilona@agentandcapital.com to hear when it is out.
A career technologist turned board director and advisor.
In executive roles at Porsche, DHL Supply Chain, Aston Martin, Innogy and ZF Foxconn I delivered ERP migrations, global infrastructure and cybersecurity builds and product-led transformations, each in under 20 months. I cut technology costs by 10 to 25%, including ahead of an IPO, and quadrupled the EBITDA of a business within a year ahead of its trade sale. I closed cyber and technology risks that due diligence had missed. As founder of Agent & Capital I advise private equity sponsors and executive teams on value creation and AI governance, and work with chief executives of technology companies on what their customers need from AI. I am a Harvard Certified Corporate Director and an ISACA-certified IT auditor with its Advanced in AI Risk credential.
Get AI ROI by email every Monday at 07:00 UK time: airoi.beehiiv.com/subscribe
AI ROI is written by Ilona Simpson and published by Agent & Capital Ltd, 71-75 Shelton Street, Covent Garden, London WC2H 9JQ. Sources are linked in each item; figures are as stated by the companies and institutions named. Every figure is checked against its source, but mistakes can still happen. I correct any error in the next issue. Not investment advice. © 2026 Agent & Capital.
Get AI ROI by email every Monday at 07:00 UK time. One story each on investment, return and risk. Unsubscribe in one click.