The 75/15 Problem: Why Most Enterprises Adopted AI Agents and Got Nothing Back
Three out of four enterprise leaders say they've adopted agentic AI. One in seven say it's producing returns.
That gap — 75 percent adoption against 15 percent value realization — is the central finding of Beyond chatbots: The agentic economy is here, a new report from Meta drawing on Forrester, Gartner, McKinsey, Deloitte and PwC research, plus executive commentary from Meta Chief Data Officer Alex Schultz. The report's argument is straightforward: the bottleneck isn't the models. It's everything underneath them.
From answering to acting
The distinction the report leans on is between a chatbot and an agent, and it's worth taking seriously because it's not just marketing vocabulary.
A chatbot answers a question and stops. An agent answers, recommends a product, closes the sale, processes the return, reorders the item, and follows up next week. One responds; the other executes. Meta frames this as a structural reconfiguration rather than an incremental upgrade — AI moving from the support function into the full customer lifecycle: discovery, lead generation, sales, service, re-engagement.
The market projections attached to this are large. McKinsey puts the global economic impact of agentic commerce at $3–5 trillion by 2030. Gartner forecasts that over 60 percent of enterprise customer service interactions will be handled end-to-end by agentic AI by 2028, up from roughly 20 percent in 2026. McKinsey also finds 55 percent of consumers already migrating product discovery from search engines to LLMs — which, if it holds, means the front door to commerce is being relocated.
Schultz offers the sharpest line in the report on timing: "Agents are the dumbest they are ever going to be. If you are building for where agents are today, you are building for the past six months from now."
Where the value leaks out
The report's diagnosis of the 75/15 gap is the most useful part of it, and it's less flattering to AI vendors than you'd expect from a vendor-published document.
Forrester's data has 40 percent of enterprises trapped in a pilot loop — proofs of concept that never scale, dashboards nobody acts on, agents disconnected from the work that matters. PwC's CEO survey finds only 12 percent of chief executives reporting that AI has delivered both cost and revenue benefits.
The common pattern among the stuck: agents that produce insights but still require a human to manually carry the recommendation across to execution. As the report puts it, "this distance between recommendation and execution is where ROI lives — or erodes." Forrester's framing is blunter still: the problem "stems from poor integration of agent models with business workflows."
That's a claim about plumbing, not intelligence. Buy a better model and the gap stays exactly where it was.
Five layers of infrastructure
The report proposes five capabilities an organization needs working together before agents produce returns:
Identity. When an agent represents a business, there has to be a verified identity behind it — because when it commits on your behalf, the counterparty needs to trust the commitment.
Relationships and discoverability. Agents need a directory to find other agents and businesses. And when a buyer's agent evaluates options, it queries structured data. Businesses without machine-readable catalogs, verified credentials and accessible pricing simply don't appear in the results. Discoverability to agents becomes a competitive surface in its own right.
Messaging. The report argues the best interface for agents is the one already in use, citing 1 billion-plus daily messages across Messenger, Instagram and WhatsApp. With a caveat worth noting: messaging only works as an agent channel if it stays two-way. Treat it as a push-notification pipe and you get initial conversions followed by declining engagement.
Commerce. Catalogs, checkout, marketplaces — the transaction layer has to keep pace with agents that compare options and negotiate at machine speed.
Models and protocols. A marketplace of specialist tools and agents to depend on. Schultz's analogy: "Where we built an app store for humans, we need an app store for agents. Where you have a payments infrastructure for consumers today, you are going to need that for agents. Where you have a legal structure today, you are going to need that for agents."
That last point gestures at the endgame the report calls agent-to-agent interaction — your agent sourcing quotes from a venue's agent, coordinating schedules, returning decision-ready options in seconds. If that arrives, a three-person operation gets the same always-on capability as a large enterprise, and infrastructure readiness becomes the differentiator rather than headcount.
What the early results show — and don't
The report includes three production case studies, all on WhatsApp via Meta's Business Agent Platform: Turkish retailer Trendyol reached 123,000 customers in a single week with a median response time under seven seconds; Brazilian tolling company Sem Parar saw a 70 percent payment completion rate for in-conversation payments; car rental firm Movida reported 85 percent of conversations resolved without human assistance and a conversion rate of 14.9 percent against a previous best of 9.7 percent.
These numbers are striking. They're also self-reported and flagged in the report's own footnotes as not identifiably repeatable. Read them as existence proofs — this can work — rather than as benchmarks.
The same caution applies to the report as a whole. It's a Meta publication, and its five infrastructure layers map neatly onto assets Meta already owns: verified business profiles in 180-plus countries, cross-session memory, in-thread checkout, first-party signals. The analysis is sound; the conclusion that Meta's stack is where those layers converge is a sales argument.
The part that survives the vendor framing
Strip out the product positioning and a durable point remains. Gartner projects that by 2030, 80 percent of sales and marketing leaders will treat agentic AI integration as critical to competitive advantage — up from under 50 percent in 2026.
If that's directionally right, the enterprises pulling ahead aren't the ones with better models. They're the ones that spent this window making their catalogs machine-readable, their identity verifiable, and their workflows executable end-to-end. That work compounds quietly, and it can't be bought in a hurry once competitors are two years into it.
Source: Meta, "Beyond chatbots: The agentic economy is here — An ROI blueprint for enterprises stuck between AI adoption and AI value." Statistics cited from Forrester, Gartner, McKinsey & Company, Deloitte, PwC and Meta internal data as attributed in the original report.