The 75/15 gap: why most AI agent pilots don't pay back, and what the 15% do differently
75% of enterprise leaders say they have adopted agentic AI. Only 15% are seeing returns. Both are footnoted to Forrester on the opening page of Meta's July 2026 report on the agentic economy. Its diagnosis: the constraint is the infrastructure underneath the model, and the return lives between an agent recommending something and the business doing it. This page explains the gap, shows what closing it looks like in the published cases, and sets out how Learnmind builds agents that act.
The short answer
Adoption is not value. 75% have adopted, 15% see returns, and the report places 40% in pilots that never scale.
The stuck pattern is an agent that advises while a person does the work. The report calls it the distance "between recommendation and execution".
The fix is connection, not intelligence. The agent books, takes payment and updates the record itself, within rules you set.
Measure results, not conversations. Cost per result, average order value and customer satisfaction, which is what Meta's own playbook recommends.
Smaller businesses can often close the gap faster. Fewer systems means fewer connections to build.
What "adopted" usually means
The report describes the 40% who "remain trapped". They are "running pilots that never scale, producing reports nobody acts on, and failing to connect agents to the work that matters." In a small or mid-sized business, that tends to look like one of three things.
- A widget that answers questions. It knows the opening hours and the FAQ. It cannot book, quote or check anything.
- An agent that collects and forwards. It takes the customer's details and promises that someone will be in touch. Someone often is not.
- A demo that never met real data. It impressed in a meeting, then stalled on the connection to the diary or the stock system.
Each of these is "adopted". None of them is a return. PwC's CEO survey, cited in the report, found only 12% of CEOs say AI has delivered both cost and revenue benefits.
The gap between recommending and doing
Meta's report is precise about where value leaks. Enterprises stuck at proof-of-concept share "a common pattern": agents that "deliver insights but still require humans to manually bridge workflows". The distance between recommendation and execution, it says, "is where ROI lives", or where it erodes.
| Step | Agent that advises | Agent that acts |
|---|---|---|
| Customer asks for Thursday | We have availability on Thursday | Checks the diary and offers two real slots |
| Customer picks a slot | A colleague will confirm shortly | Books it and confirms in writing |
| A deposit is required | We will send payment details | Sends a secure payment link in the thread |
| Afterwards | A person re-types the details into the system | The record is already in the CRM |
| Where the customer goes | Often to whoever confirmed first | Nowhere. It is done. |
The report quotes Forrester's version of the same point: "Too many AI agent implementations are struggling to get ROI. The problem stems from poor integration of agent models with business workflows."
What acting looks like in the published cases
The early movers in Meta's report share one trait. Their agents take the customer to the transaction inside the conversation: a booking, a payment or a one-tap checkout link.
| Business | What the agent completes | What was reported |
|---|---|---|
| Movida, car rental, Brazil | Booking, customer history and payment, connected to reservation, pricing, payment and biometric systems | 85% of conversations resolved without team assistance; 14.9% conversion in week four of the test, against a previous best of 9.7% |
| Sem Parar, toll and mobility payments, Latin America | Payment inside the conversation, plus a tag-subscription cross-sell | 70% payment completion for drivers who started a payment in the chat; 13% query-to-paid conversion; a cross-sell offer in 32% of sessions |
| Trendyol, e-commerce | Product recommendations as shoppable carousels, with a one-tap link to checkout | Recommendations in 13% of conversations; median response under seven seconds |
| Alcaz Media, performance agency, UAE (a Learnmind client) | Qualification, pricing and results shared in the chat, the strategy session booked, the deal created in the CRM | Replies within seconds at any hour; each qualified booking arrives with an email summary and the full transcript |
Meta marks its cases as self-reported and "not identifiably repeatable". The Alcaz Media row comes from our own published case study. Treat all four as evidence of a pattern, not as a forecast for your business.
The three things underneath a working agent
The report names what holds the 40% back. "Without mature orchestration, unified data, and clear governance, pilots stay circling without ever reaching production." Translated for a business that is not an enterprise:
- Orchestration: the agent can reach your systemsThe diary, the stock list, the payment link, the CRM. If the agent cannot touch them, it can only talk about them.
- Unified data: one version of each factOne price list, one customer record. Two sources of truth guarantee two answers.
- Governance: written rules on what it may do aloneWhat the agent can confirm, what needs a person, and who that person is. Without this, nobody trusts it with the booking.
Trendyol's CEO, Erdem Inan, made a related point at Meta's Conversations event in London. In Trendyol's failed AI projects, data quality was "one of the top issues", and "in most cases, the process was bad."
Why the gap widens the longer you wait
The report says the 15% "are pulling away", and elsewhere that the advantage compounds: "Every agent conversation generates intelligence that makes the next one better. Start late, and the gap widens with every interaction your competitor has that you don't."
It also cites Gartner's forecast that by 2028 more than 60% of enterprise customer service interactions will be handled end to end by agentic AI, up from 20% in 2026. That is a forecast, not a fact.
A 90-day plan that does not stall
Meta's own playbook has three stages: scope and connect, configure and integrate, launch and measure. Here is a 90-day version for a smaller business that wants a result rather than a pilot.
- Days 1 to 14: one journey, connected and livePick the highest-volume conversation, such as enquiries that should become bookings. Connect the diary and the CRM, write the rules and the handovers, and go live. Movida started with the reservation journey and is now extending the agent across the whole rental cycle.
- Days 15 to 45: read the conversations weeklyEvery decline, handover and hesitation shows a missing fact or a missing rule. Fix the source, not the prompt.
- Days 46 to 90: measure and extendTrack cost per result, average order value and customer satisfaction, the three measures Meta's playbook names. Only then add a second journey.
Checks before you approve another pilot
- It completes at least one transaction on its own. A booking, a payment or an order, not only an answer.
- It reads live data. Availability and prices come from your systems at the moment of asking.
- It hands over well. A named person receives the full conversation, not a request to call someone back.
- It writes the record. Nobody re-types what the agent already captured.
- It has a number attached. A cost per result you can compare with what the same result costs today.
What Learnmind builds
We build agents that act, on WhatsApp, Instagram and Facebook messages, and inbound calls. Done for you, with WhatsApp on the official WhatsApp Business Platform.
- Replies within seconds at any hour, in your voice, in Arabic and English.
- Qualifies the enquiry with the questions a good salesperson would ask.
- Books the appointment straight into the connected calendar.
- Hands hot leads to a named person with the conversation attached.
- Writes every enquiry, answer and booking into a built-in AI CRM.
Most Learnmind agents are live within two weeks. See how it works and the results from live deployments.
Want an agent that finishes the job?
We build agents that book, qualify and hand over, connected to your diary and CRM. Pricing logic, escalation rules and the booking calendar are settled with you before launch.
Terms worth being precise about
- Chatbot
- Answers a question and waits. In the report's words, an agent "answers, makes a personalized product recommendation, closes the sale, processes the return".
- AI agent
- Software that takes actions in your systems on a customer's behalf, within rules you set.
- Orchestration
- The connections that let an agent use your diary, stock, payments and CRM.
- Governance
- Written rules on what the agent may do alone and when a person takes over.
- Cost per result
- What you spend to produce one booking, sale or resolved request, all costs included.
Frequently asked questions
Why do most AI agent pilots fail to deliver a return?
Because the agent is not connected to the work. It answers or advises, and a person still has to book,
charge or update the record. Meta's report locates the lost return in that gap.
What is the difference between a chatbot and an AI agent?
A chatbot answers and waits. An agent completes the task: books the slot, sends the payment link,
updates the CRM.
How should I measure an AI agent's return?
By cost per result, average order value and customer satisfaction, against what the same results cost you now.
Not by the number of conversations it handled.
How long before an agent pays back?
We will not give a generic figure. Count your own unanswered or slow-answered enquiries for two weeks,
apply your own conversion rate and average value, and compare that with the cost.
Is this only a problem for large enterprises?
No, and smaller businesses can often close it faster. There are fewer systems to connect and fewer people
who need to agree the rules.
How we checked this, and what we could not settle
Checked: the 75%, 15%, 40% and 12% figures are as printed in Meta's report "Beyond chatbots: The agentic economy is here", published in July 2026. So are the Forrester and Gartner citations and the Movida, Sem Parar and Trendyol results. Trendyol's comments on failed projects come from the transcript of Meta's panel "Building the Customer Journey of the Future". The Alcaz Media row comes from our own results page.
Not settled: the report's statistics describe enterprises, and we have not found an equivalent survey of small and mid-sized businesses in the UAE. We have not read the underlying Forrester, Gartner or PwC reports, so we rely on the figures as Meta's report states them.
Sources
- PlatformMeta, Beyond chatbots: The agentic economy is here, July 2026, citing Forrester, Gartner, McKinsey and PwC
- PlatformMeta, AI Agents and the Future of Customer Engagement, including the panel with Movida and Trendyol
- LearnmindLearnmind results, the Alcaz Media case study
Written by Edmund Gay, Learnmind.ai, Dubai. Figures are as published by the sources named, on the dates given; forecasts are the forecasters' own.