Mostly Promises, Very Little Evidence
Agentic AI in FMCG
Most of what is currently being presented as evidence for the agentic AI revolution in FMCG is vendor marketing dressed as research. Not because the technology has no future. It almost certainly does. But the case for transformation today rests on a body of “evidence” produced almost entirely by the companies selling the tools.
That distinction matters. It should change how you read every report, briefing, and case study landing in your inbox right now.
Your Evidence Source Is Also Your Salesperson
Pick up any serious-looking report on agentic AI in consumer goods. Follow the footnotes. You will find Microsoft blogs, Google case studies, Pactum white papers, and Snowflake posts. These are not independent researchers. These are the infrastructure vendors, the software providers, and the negotiation platform companies with a direct commercial interest in the conclusion.
The financial projections look authoritative. A projected three-year ROI of 124% to 282% for a typical five-billion-dollar enterprise. Between 7.7 and 17.6 million dollars in net present value. Where do these numbers come from? A Microsoft blog post. About Microsoft products.
This is not a minor disclosure issue. It is the whole problem.
The issue is not that these companies fabricate results. The issue is that they select which results to publish. Every vendor has a structural incentive to surface only the wins. There is no mechanism in a company blog post for publishing the pilots that ran for eighteen months and produced nothing. There is no audit trail for the implementations that cost more than they saved. You see the curated set. You do not see the base rate.
When Architecture Fills the Gap Where Outcomes Should Be
When the commercial results are thin, the language shifts. Architecture fills the gap.
L’Oréal’s AI beauty advisor synthesises 150,000 dermatologist annotations. It keeps latency under five seconds. It handled 2,000 simultaneous users in testing. These are real engineering achievements. Not one of them is a business outcome.
Where is the conversion rate data, independently audited? What happened to basket size? What is the repeat purchase rate among users who engaged with the tool versus those who did not? The report does not say. What it offers instead is “up to 4% conversion rate improvement.” That figure is a modelled estimate from the same Microsoft source, not a reported result from L’Oréal’s own investor communications.
Jo Malone’s AI Scent Advisor “recreates the brand’s in-store consultation experience digitally.” Unilever’s Market Insights agent “reduces strategic analysis from hours to seconds.” Nestlé “integrates AI-driven inventory prediction with Coupa.” Every one of these sentences describes what the tool does. None of them tells you what changed commercially as a result.
This is not a minor gap. It is the whole gap.
The pattern is consistent enough to be deliberate. Describe the inputs. Name the technology stack. Reference the scale of data processed. Then assert, rather than demonstrate, that commercial value follows. If you are a CMO or a supply chain director, that assertion is not enough to justify a budget.
The One Number That Should Set Your Standard
The most concrete case study in the current AI-in-FMCG literature involves AB InBev.
AB InBev reportedly runs 30 billion dollars in decisions through AI annually. The specific production deployment covered in detail is an order correction system. When a customer’s order contains an error, the system creates a ticket, contacts the customer to obtain the correction, and closes the ticket without human involvement. It handles roughly 30% of the 20 million annual tickets. The direct annual value created: $ 1.2 million.
AB InBev’s annual revenue is approximately 59 billion dollars. That value figure is roughly 0.002% of revenue. This is useful process automation. It is genuinely operational. It is not a transformation.
The second figure, 310 million dollars in orders delivered 1.5 days faster, sounds more significant. But the report provides no data on what that acceleration actually produced. Were those orders at risk of cancellation without it? Did the 1.5 days change purchasing behaviour downstream? We are not told.
The reason this matters is not to dismiss AB InBev’s work. It matters because if the strongest, most specific case in the literature produces 1.2 million dollars in direct value for a business of that scale, then the 124% to 282% three-year ROI projections require serious scrutiny before anyone acts on them.
The Wrong Question Is Already Everywhere
Every FMCG executive right now is being asked: “How is your organisation preparing for agentic AI?”
Wrong question. Ask instead: “Can you name a specific commercial outcome, a number tied to a quarter, verified by someone other than the vendor, that this technology produced?”
If the answer is no, you are looking at infrastructure investment and institutional anxiety dressed as strategy. That is the Fear vs. Strategy diagnostic playing out in real time. Companies are committing budgets not because they have identified a problem that agentic AI solves better than existing tools. They are committing budgets because they fear being the last one to act.
That is not inherently irrational. Being genuinely late to a technology shift carries real cost. But it is not the same as having evidence. And the two are routinely conflated in briefings designed to generate that exact anxiety.
The Hard Truth
The technology will matter. Possibly more than most operators currently believe. But the gap between what is being claimed today and what is being demonstrated today is substantial. And that gap is not an accident.
The current AI-in-FMCG narrative mirrors what happened with digital transformation in the mid-2010s, and with big data before that. Each cycle produced a wave of projected ROI figures, vendor-sourced case studies, and language about irreversible change. The companies that genuinely benefited did so quietly, years later, after the hype had dissipated and the actual use cases had been sorted from the theatre.
Who, besides the vendor, has independently measured what this produced? If the answer is 'nobody,' you are not reading the evidence. You are reading a forecast with a logo.
Act accordingly.