Agentic economy: what the Meta report gets right and the layer it misses

The document arrived through the channel it describes

I received the report by email, triggered by WhatsApp Business. Twenty-nine pages, a cover signed by Alex Schultz, Data Director at Meta, and a title that does not hold back: beyond chatbots, the agentic economy has arrived.

It is worth reading. It is even more worth reading knowing what it is.

It is not an independent study. It is a commercial positioning piece built with legitimate third-party data, and built very well. The difference matters when taking these numbers to a board meeting.

What the agentic economy is

The report separates two things the Brazilian market still treats as synonyms.

A chatbot receives the question, answers, and stops. It closes the service at the point where the conversation stops being a doubt and starts being business.

An agent answers, recommends the right product, closes the sale, processes the return, reorders, and reaches out to the customer again the following week. It goes through the entire cycle. The agentic economy is what happens when this capability stops being the exception and becomes the standard way of operating.

The difference between a chatbot and an agent is not in the quality of the answer, it is in what happens after it.

The numbers the report puts on the table

The pair that sustains the entire report is the first one: 75 against 15. Three out of four companies say they have adopted it. One out of seven makes money from it.

The distance between adopting and generating revenue, distributed across one hundred companies.

The diagnosis is right

The most honest point in the document is the refusal to blame the model. According to the report, what traps 40% of companies in the proof of concept is not the AI quality. It is the infrastructure underneath it: immature orchestration, scattered data, undefined governance.

And the report goes further, in the excerpt I would highlight if I could keep just one sentence. Companies stuck in the pilot phase share a common pattern: the agent delivers the insight and a human needs to connect the workflow by hand. The distance between recommending and executing is exactly where the return is won or lost.

This is true. I see this every week. The company hires AI, turns on the agent, it answers beautifully, and the moment the conversation needs to become an order in the ERP, a real-time stock check, or a confirmed reservation, someone grabs their phone and solves it manually. The agent turned into an employee who only knows how to give opinions.

The five layers, and who they belong to

Based on the diagnosis, the report defines five infrastructure layers necessary for the agentic economy to work: verified identity, relationships and discovery, messaging, commerce, and finally, models and protocols.

Read the list again with a question in mind: who already has these five things ready, on a global scale?

Verified business profiles in more than 180 countries. A social discovery graph. WhatsApp, Messenger, and Instagram. Commerce Manager, catalog, and in-thread checkout. And now an agent platform.

The five layers were designed backwards, based on Meta's own asset inventory. The conclusion on page 24, that these layers already exist at scale on the company's platforms, is not a discovery of the study. It is its premise, dressed as a conclusion.

This does not invalidate the list. The five layers are real and Meta indeed has them. It only invalidates the idea that this list is complete.

The sixth layer

The layer that answers the question the report itself raised is missing.

If the bottleneck is the distance between recommending and executing, the layer that solves the bottleneck is orchestration. Designed process, unified data, handover to human with business rules, and measurement of what happened.

None of the five listed layers covers this. Identity does not route customer service. Catalog does not define when the agent stops and calls the sales consultant. Messaging does not decide if that lead enters the repurchase flow in seven or thirty days. No model knows on its own that your company's exchange policy changes for display products.

The five blue layers belong to Meta. The red one is what decides if the project generates a return.

This layer is where the operation lives. It is also, by chance or not, the layer that does not appear in the report.

Three absences that weigh heavily in Brazil

Price. Twenty-nine pages promising a practical roadmap for CXOs, COOs, and CEOs, and not a single line about unit cost. Who pays for the agent's inference, how it behaves in an operation with thousands of conversations a day, and how it aligns with the message pricing change already underway on WhatsApp. Omitting cost in a document about ROI is a choice, not an oversight.

LGPD. The report is in Brazilian Portuguese, features two Latin American cases, and lists compliance with ISO 27001, GDPR, and CCPA. LGPD does not appear. In a bank, professional council, public agency, or health insurance provider, this is the first line the legal department looks for before signing anything.

Attribution. The measurement recommendation fits in one line: cost per outcome, average order value, and customer satisfaction. That is too little for a thesis depending on proving value in a conversational journey, where the same conversation can contain discovery, sales, support, and repurchase.

Two numbers that slip up

If you are going to cite the report, cite carefully on two points.

On page 5, the document mentions more than one million businesses using the Meta Business Agent weekly. Two pages later, the same number appears described as businesses using AI chatbots on Messenger and WhatsApp. They are different things, and the second is much easier to achieve.

The other case is about units. Page 15 mentions more than one billion messages a day. Page 24 mentions more than one billion daily B2C conversations, based on internal Meta data from May 2025. A message and a conversation are not the same measure, and the difference between them is usually an order of magnitude.

The case studies also deserve reading the footnotes. Trendyol, Sem Parar, and Movida bring strong numbers, including 85% total resolution by the agent in one of them. The footnote informs that the results are self-reported and not reproducible in an identifiable way. They are sponsored case studies, not market references.

The window is short, and this part is true

All criticism aside, the report's sense of urgency is not made up.

Gartner's projections for customer service and competitive perception. Note that the 2030 metric measures something else.

Going from 20% to over 60% of customer service managed by agentic AI in two years is an operation reconfiguration, not a tool upgrade. And the accumulation argument is the strongest in the document: the knowledge you give to the agent, such as catalog, policy, and history, accumulates. Those who start earlier do not just get ahead, they get on a different curve.

There is an excerpt on page 13 that I endorse word for word and which, ironically, is the central MCI thesis written by Meta itself: session-based AI starts from scratch with every interaction, while companies maintaining persistent conversations accumulate intelligence and make each subsequent investment more efficient.

That is what it is all about. The channel is not WhatsApp. The channel is the relationship that WhatsApp allows you to keep open.

What to do in the next 90 days

The report ends with a three-step roadmap summarized as connect, configure, and measure. Things are missing. Here is the version I would give a client:

1. Map where your conversation dies today. Find the exact point where the service stops and a human needs to open another system. That point is your ROI bottleneck, not the bot's quality.
2. Fix the catalog before fixing the agent. An agent without structured and updated data does not recommend, it invents. If your stock changes daily, the catalog integration is the project, and the agent is the consequence.
3. Write the human handover rule before turning anything on. In what scenarios the agent stops. Who receives it. In how much time. Without this, you do not have automation, you have risk.
4. Define the metric before the pilot. Cost per resolved outcome, resolution rate without humans, and revenue attributed to the conversation. If you only measure message volume, you will conclude it worked even when it did not.
5. Decide the architecture with an eye on lock-in. Building natively on Meta's platform, integrating via a partner, or bringing your own agent are three decisions with very different exit costs. The third way is still under development.

A note from someone running this

I am not writing this from the outside. Since July 2026, I have been operating Meta's new AI agent in production, first in Hablla's customer service and support, then in an ecommerce with over five thousand SKUs, and now in a dealership network with six brands and a used car operation, all with a connected and daily updated catalog. The integrated version is in beta inside Hablla One.

What I have learned so far confirms the report's thesis and contradicts its implicit conclusion. The hard part was never turning the agent on. It was keeping five thousand SKUs coherent, deciding what happens when the customer asks about a car that just left the lot, and ensuring the conversation turns into an order in the right system.

In other words: layer six.

I will detail this operation in the next article, with the screens and numbers that can already be shown.

The full report is available in the WhatsApp Business resource library, and you can also read the 29 pages right here, in the reader above.