What the report found
A new report from Boston Consulting Group and The Consumer Goods Forum puts hard numbers on a pattern executives have suspected for two years: consumer packaged goods companies are experimenting with AI everywhere and scaling it almost nowhere. Surveying 39 senior executives across CPG and retail, the study found that 76 percent of CPG companies remain in pilot or exploration mode, while only 18 percent have scaled AI meaningfully across their businesses. The technology is spreading across forecasting, merchandising, marketing, and product innovation, yet most of that activity sits in controlled trials rather than production. The gap between adoption and scale is the report's central finding, and it maps closely to what BCG has documented in other sectors this year.
The financial stakes are large enough to make the stall costly. BCG estimates that fully scaled AI initiatives could generate between 220 and 350 basis points of cumulative EBIT improvement for CPG companies, and between 180 and 360 basis points for retailers. Future agentic systems, the report suggests, could expand that opportunity by roughly 1.7 times. Those are meaningful margin figures in an industry where a point of EBIT is fought over hard. The implication is blunt: the value concentrates in scaled deployment, and most companies have not reached it. Every quarter spent iterating on proofs of concept is a quarter of that margin opportunity left on the table, while better-prepared rivals compound their lead.
The pilot trap
The report's most quoted line explains why pilots so rarely graduate. AI projects, it notes, often look fantastic inside controlled environments with dedicated teams, clean data, and limited complexity. Then reality arrives carrying legacy systems, organizational politics, and 47 spreadsheets maintained by somebody who retired in 2018. The joke lands because every operator recognizes it. A demand-forecasting model that shines on a curated dataset collides with the actual data estate: fragmented ERP instances, inconsistent product hierarchies, and manual processes glued together over decades. The pilot succeeds precisely because it was insulated from that mess, which is exactly why it does not survive contact with production. Scaling is a data-and-process problem long before it is a model problem.
This reframes where CPG leaders should spend. The instinct after a promising pilot is to buy more models or hire more data scientists. The report implies the binding constraint is upstream: clean, governed, integrated data and the organizational will to change the workflows around it. That is slower, less exciting work than standing up another proof of concept, and it is the work that determines whether AI ever reaches the margin figures BCG cites. Companies that treat AI as a procurement exercise, buy a tool, run a pilot, and declare progress, will keep cycling through demos while the 18 percent who fixed their data foundations pull ahead. The differentiator is infrastructure and governance, not access to models everyone can license.
Why ROI stays invisible
A striking finding is how few companies can even tell whether their AI works. The report found that 53 percent of respondents do not formally measure the return on their AI investments, and only 11 percent claimed ROI greater than five times. When a majority of an industry cannot quantify the payback on a strategic priority, two things follow. First, investment decisions are being made on faith and fear of missing out rather than evidence. Second, the pilots that get killed may be dying for lack of measurement rather than lack of value, and the ones that survive may persist on executive enthusiasm rather than results. Absent measurement, the whole portfolio is flying blind, and good and bad bets look identical on the balance sheet.
The measurement gap compounds the pilot trap. If a company cannot attribute revenue lift or cost savings to a specific deployment, it cannot build the internal business case to fund the harder scaling work, and it cannot defend the budget when finance scrutinizes it. Measurement discipline is unglamorous and it is the precondition for everything else. The 11 percent reporting greater than fivefold returns are almost certainly the same cohort that instrumented their deployments from the start, because you cannot claim a return you never measured. For CxOs, the actionable lesson is to require a measurement plan before funding the pilot, so that the graduation decision rests on data rather than on the demo's polish.
Retailers ahead of the brands
The report draws a clear line between retailers and the brands that supply them. While only 18 percent of CPG companies have scaled AI, 45 percent of retailers report scaled deployment, more than double the rate. The gap makes sense structurally. Retailers sit on high-frequency transaction data, direct customer relationships, and concrete operational use cases, pricing, assortment, inventory, and labor, where AI produces measurable results quickly. Brands are a step removed from the shopper, often reliant on retailers and panel data for signal, and their highest-value AI applications in marketing and innovation are harder to measure. The asymmetry means retailers are compounding an advantage in exactly the data and AI capabilities that shape the trading relationship.
That should worry brand-side CIOs. Retail media networks already shifted leverage toward retailers by turning their first-party data into an ad business brands must buy into. A widening AI capability gap deepens the same dynamic: the retailer knows more about the shopper, optimizes the shelf and the price faster, and negotiates from a stronger analytical position. Brands that stay stuck in pilots while their retail partners scale risk becoming price-takers in a data-driven trading relationship they no longer fully understand. The defensive move is to build direct-to-consumer data and genuine AI capability rather than outsourcing the shopper relationship entirely. The report's numbers suggest most brands have not yet made that move at scale.
The agentic promise and its caveat
The report is bullish on what comes next while honest about where the industry stands. It found that 67 percent of companies use AI primarily for recommendations, with humans making the final decisions, and only 9 percent allow AI to execute decisions autonomously within guardrails. That distribution is the current reality of agentic AI in consumer goods: mostly a suggestion engine, rarely an actor. BCG's estimate that agentic systems could expand the EBIT opportunity by 1.7 times assumes a shift toward AI that executes, not merely advises. Getting there requires trust, governance, and data quality that most companies, by their own admission, have not yet built. The promise is real and the prerequisites are unglamorous.
We would treat the 9 percent figure as the honest measure of how early this is. Autonomous execution, an agent that reprices, reorders, or reallocates spend without a human approving each step, demands exactly the clean data, measurement, and organizational trust the rest of the report shows are missing. The companies letting AI act today are a small vanguard, and they got there by fixing foundations first. For everyone else, chasing agentic autonomy before the data and governance exist is a way to automate mistakes at scale. The sequencing the report implies is unforgiving: earn measurement, then earn scale, then earn autonomy. Skipping steps is how pilots become expensive cautionary tales.
Our read
The value of this report is that it replaces vibes with numbers, and the numbers are sobering. Two years into the generative AI wave, three quarters of CPG companies are still piloting, half cannot measure returns, and fewer than one in ten let AI act on its own. That is not a story of failure so much as a story of how hard the unglamorous middle of a technology transition actually is. The demos were always going to be easy. Turning them into scaled, measured, governed production systems inside companies running on decades of legacy infrastructure is the real work, and most of the industry is only now confronting it honestly.
For executives, the report is a useful corrective to two opposite errors. It rebuts the skeptics who call AI overhyped, because the margin opportunity is large and a minority is already capturing it. And it rebuts the maximalists who expect autonomous agents to run the enterprise next quarter, because the foundations for that do not exist in most companies. The grounded path sits between them: fix the data, instrument the pilots, scale what measurably works, and reach for autonomy only where trust and governance are earned. The 18 percent who have scaled and the 45 percent of retailers ahead of the brands show the payoff is attainable. The gating factor is discipline, and discipline is in shorter supply than technology.
