Gustavo Gadotti maps in five blocks what AI agents change in retail, from the cart to the sales team

On October 13, 2026, ACATE's Retail Vertical held the AI-Commerce Era immersion in Florianópolis, with a talk by Gustavo Gadotti on what AI agents change in retail and in the software behind it. Gadotti split the presentation into five blocks, which he called showing up, selling, what the company knows, executing and organizing.

This summary draws on an automatic transcript of the talk and on photos of the slides. Assessments, examples and cases are the speaker's. The only context fact checked against a public source is Google's announcement of its commerce protocol, which appears with its date.

Showing up

According to Gadotti, Google is bringing the purchase into search itself. He described a cart in which the shopper leaves an instruction for an agent, such as buying a shirt when the price drops 30%, with address and card already saved. In his assessment, most of these features are still rolling out in the United States.

To get into that flow, the speaker said, retailers need to adopt UCP, the Universal Commerce Protocol, with their catalog in a standardized format. Platforms such as Shopify and VTEX, he said, come in better prepared. Google announced the protocol on January 11, 2026 as an open standard for agent-driven commerce, from discovery to post-purchase. The first announced application was a checkout inside the conversational experiences of Search and the Gemini app, for eligible US retailers.

Gadotti also mentioned brand agents inside ads, where shoppers talk to the ad itself without having to visit the site, and loyalty and cashback programs entering the comparison of offers.

He also pointed to the shift from SEO to GEO, optimization for the answers of language models. In the speaker's reading, these models weigh a brand's credibility in the press, in reviews and on the Brazilian complaints site Reclame Aqui, along with the consistency of product data, and the brand's own site ranks behind other sources. According to him, visits coming from AI are still few, but they convert better and with a higher average ticket. On WhatsApp, he said, Meta started charging for service messages as well in October and wants to move to outcome-based pricing.

Selling

Gadotti described a sequence in how software gets paid for. First came the purchase, owning the software outright. Then the subscription, then usage, measured in tokens or credits, and now the outcome. As references, he cited Sequoia's article "Services: the new software" and HubSpot, which according to him charges per resolved conversation.

On one of the slides, the same climb shows up in what software does, in four steps. The system of record keeps what happened, like a printed map. The system of engagement is where team and customer talk, like Waze showing the route. The system of intelligence reads the record and the conversation and flags what matters before anyone asks, like Waze telling you when to leave. The system of action takes the next step on its own, within the rules, and records what it did, like Waymo driving the passenger. In his reading, pricing follows that climb.

What the company knows

In the third block, Gadotti talked about ontology, a term he associated with Palantir, and about the company brain, business knowledge organized in a way AI can read. The argument starts from a difference in nature. Language models are probabilistic, and business rules are deterministic. A maximum discount of 10% cannot turn into 18%. He cited an assessment by Salesforce's Marc Benioff at Dreamforce that models do not operate on their own inside companies.

To explain it, he used a restaurant analogy with four layers. The data is the recipe card and the pantry. The logic is the chef's judgment, knowing which ingredient can stand in for which. Action is the permission to order from the supplier. Security is the rule that a discount only goes out with the manager's sign-off. Without these layers, he said, the probabilistic chef swaps mushrooms for filet mignon.

As a first step, the speaker suggested asking an AI to interview you with ten questions to build your second brain, ideally answering by voice.

Executing

For the fourth block, Gadotti turned to Formula 1. In his example, the same Mercedes engine took McLaren to the title and left Williams in the middle of the grid. In his comparison, the language model plays the role of the engine, and what separates the results is the harness, the set of context, tools, memory and permissions around the agent. Evaluations matter too. In them, an evaluator agent blocks the error before the answer reaches a person.

To get started, he recommends picking a repetitive task with a clear rule and a cheap error, such as building presentations, qualifying leads and creating ad variations.

Organizing

The last block dealt with teams. On a slide about sales structure, Gadotti showed a team of 39 people, seven of them in management, shrinking to four people, one in management, with the focus on the people who sell. According to him, this smaller team works with the support of agents. The hard part, he summed up, is finding the right four.

He named two roles that, in his assessment, are gaining ground. The FDE, short for Forward Deployed Engineer, is half engineer and half consultant and only leaves once the AI is running. The other is the GTM Engineer, focused on go-to-market.

His final advice was "Eat your own food!", meaning learn by using. Gadotti said he built the entire talk with AI.

What it means for retail

As Future in Use reads it, the five blocks follow a single line. In the scenario Gadotti described, shoppers delegate part of the purchase to agents, software starts being priced on outcomes, and companies need to organize rules, data and people so that agents act within limits. The numbers and cases cited in the talk are the speaker's and have not been independently verified by Future in Use.

Sources and transparency

• Gustavo Gadotti's talk at the AI-Commerce Era immersion, ACATE Retail Vertical, Florianópolis, October 13, 2026, based on an automatic transcript and photos of the slides.
• Google, The Keyword, "New tech and tools for retailers to succeed in an agentic shopping era", January 11, 2026. https://blog.google/products/ads-commerce/agentic-commerce-ai-tools-protocol-retailers-platforms/

Future in Use is an editorial initiative linked to OperOS AI.