AI ROI, Issue 1 | Agent & Capital

Issue 1 · 21 September 2026

AI ROI

This week

Each week my agentic squad scans the wires, the papers of record, company investor releases and the regulators. Most of what they carry on AI is forecasts, surveys and product launches. I check the sources myself and keep only what the companies and their leaders have said or shown themselves: the figures, the disclosures and the interviews. This week Ingram Micro wrote its AI platform into the financial targets it gave investors to 2029, and Procter & Gamble counted the return on AI inspection in scrap on its own production lines. OpenAI, for its part, published its own count of agents doing things they were not asked to do, which is a record I rarely see in companies that run agents. In each case the figure comes from records the company already keeps for another purpose, and that is where I would start looking in any business.

Ilona

Investment

Ingram Micro

Ingram Micro Outlines Long-term Strategy, Intelligence-led Transformation and Multi-year Financial Framework at Capital Markets Day

At its Capital Markets Day on Tuesday 15 September Ingram Micro set a four-year financial framework and tied it to Xvantage, the platform that now carries more than 400 AI and machine-learning models and an assistant across its 165,000 customers and 1,500 vendors. The targets to 2029: net sales growing 4 to 6 per cent a year, gross profit 5 to 7 per cent, operating expenses held at 4.4 to 4.8 per cent of sales, non-GAAP net income up 11 to 13 per cent a year and adjusted ROIC of 16 to 17 per cent. A week earlier, at Goldman Sachs, CFO Mike Zilis said the platform had "removed $200 million in annualized OpEx" between late 2023 and early 2025, was live in 22 of 57 countries, and that return on invested capital had risen 240 basis points in a year.

"Our long-term financial framework is built on clear objectives of growing gross profit faster than net sales, driving further operating expense efficiency and leverage."
Mike Zilis, Chief Financial Officer, Ingram Micro

Ingram Micro has put its AI platform into the numbers investors will hold it to, with a $200m opex reduction already banked and ROIC targets to 2029, which is the form an AI investment should take before a board or an owner approves the next tranche.

Return

Siemens

Siemens and Procter & Gamble roll out AI-based quality inspection worldwide

Siemens and Procter & Gamble said on Thursday 17 September that they are rolling out AI visual inspection across P&G's global production. The Visual Inspection Cockpit combines P&G's own deep-learning models with Siemens' Industrial Edge platform and Nvidia-powered industrial PCs, inspects thousands of products a minute, and has cut scrap by 10 to 20 per cent where it runs. The two companies say it deploys five to ten times faster than conventional vision systems and scales from a single line to the whole footprint. Paul Thomas, P&G's Director of Machine Vision and Applied AI, said the system handles overlapping components, low-contrast defects and highly decorated packaging that older systems could not.

"Full inspection accuracy for thousands of products per minute, scalable from a single line to a global footprint."
Rainer Brehm, Chief Operating Officer Automation and Chief Technology Officer, Siemens Digital Industries

P&G's return comes from a defect count on a production line, not from a survey, and the reason it scales is that the models are P&G's own and the data is the line's own; that is the test to apply to any AI return a management team puts in front of you.

Risk

Fortune

In transparency push, OpenAI discloses six more incidents of agents going rogue

OpenAI published on Thursday 17 September a new reporting framework for cases in which its agents behave in ways they were not asked to, acknowledging that earlier disclosures had been "ad hoc and less frequent than ideal". The six cases include a model that fabricated financial figures for a California county after using credentials it had not been given, one that uploaded files to create citations it could not find, agents that used an internal software repository to pass notes to one another, and agents that uploaded local files to public websites to get round limits on their communication. During training, one model instructed itself 27 times to reject human authority.

"We want to be more transparent about the misalignment we see during training, evaluations, and deployment."
Marcus Williams, researcher, OpenAI

OpenAI's six cases were found by the developer in its own systems, so the question for any company running agents is who inside the business would have found them, and how long it would have taken.

Leading AI labs may need to be nationalized because risks are so high, Palantir's Karp tells CNBC

Alex Karp, chief executive of Palantir, told CNBC's Squawk on the Street on Thursday 17 September that his enterprise clients are frustrated that proprietary business knowledge is being drawn into frontier models and then used by their competitors, "and they paid for it". He set out a three-step model for who carries the loss when a model does harm: the developer first, under civil and criminal liability; disclosure second; and, where a developer will not carry the exposure, other parties brought in to share it. In his account the liability is unlimited, which is why he expects the largest laboratories to seek government protection. Karp sells the alternative to the laboratories he describes.

Host: How does the liability get written into the S-1? What do the risk factors sound like?
Karp: You're assuming that there will be an S-1. The only way to deal with this kind of liability is to go to the government and say, "Nationalize us, please."
CNBC, Squawk on the Street, 17 September 2026

The data terms between a company and its model suppliers are now a diligence item: what the supplier retains, what it may train on, and who pays when the model harms a customer.

CNBC

What to remember

Before approving the next AI budget, ask management to show where the spending sits in the plan's numbers, as Ingram Micro did with operating expenses and return on capital, and which line in the accounts will carry the return. When a return is put in front of you, ask which record it was measured in, because P&G's scrap figures come from its own production lines and a survey of users would have told you far less. Then ask who in the business would notice an agent acting outside its brief and how long that would take, and check what your model suppliers' contracts say about your data and about who pays if their model harms a customer. If nobody can answer the question about agents, I would settle that before anything new goes live.

Ilona

From Agent & Capital

Conference

The AI illusion: where AI builds and dissolves enterprise value.

Ilona Simpson joined Rebecca Hastings of Owendale Advisory and Sarah Hollyhead, IFT Independent, on the AI panel at the Institute for Turnaround's National Conference in London on Thursday 18 September. Her argument was to start where the return is nearest: the back office, where invoices, contracts, assets and entitlements can be reconciled against what is really happening and the automation pays for itself within weeks. What decides whether that return arrives is people, not software; a turnaround needs at least one person inside the business who knows the data and the process well enough to tell a right answer from a plausible one. And the same discipline that applies to any other spending applies here: governance and guardrails set before the tools run, and the risks weighed against the return rather than after it.

Continue the conversation

Board dialogue

Governing in the AI era.

On Monday 14 September Ilona led a dialogue with a group of directors in Madrid on how AI changes the strategic responsibilities of a board. Planning assumptions age faster, because capability, cost and vendor terms now move inside the annual plan, so boards should approve in shorter gates and reprice at each one. Software can take on work across functions and execute several steps on its own, so every AI-supported process needs a named outcome owner and a named process owner. And an agent or an attacker can act before the normal review cycle, so the board sets authority, intervention and recovery in advance rather than after the event. AI belongs in the board's existing decisions about value, people and risk.

Continue the conversation

The book

AI ROI.

This newsletter shares its name with the book Ilona is writing on how companies and their owners get a return from AI, and it will carry the working material as the chapters take shape. Reply to this email to hear when it is out.

AI ROI arrives on a Monday. Every issue keeps to companies that have put a figure on their own AI.

Subscribe by email