AI as the new consumption interface: what the data shows and what changes

Isabella Piratininga, Director of Technology & Innovation at iFood, opened her presentation at StartSe with a thought-provoking question:

"How much of your product was designed to change user behavior?"

From this provocation, we understand that behavioral change is inevitable, it is already happening, and it is up to companies only to adapt. The question here is to understand at what speed, with what friction, and for what segment of the market.

Each era of consumption has its main interface

The history of consumer behavior can be told as a sequence of dominant interfaces that have accumulated and have not been replaced.

In the physical store, the interface is the salesperson: power lies with those who know how to persuade and have the best commercial location, where the consumer enters, looks, and decides;
In e-commerce, the interface is search: power belongs to those who have the most relevance on Google, and the consumer gains autonomy; they research, decide, and buy themselves;
In apps and feeds, the interface is the algorithm: power belongs to those who have behavioral data and can predict what the consumer wants;
In the AI era, the interface is conversation: power goes to those who have the best-trained model on their user's behavior.

Each new interface adds a layer, expands the total market, and redistributes power within it. The conversational interface will follow the same pattern and will not replace what already exists; it will only become an additional channel of high relevance for specific contexts, leaving the market to decide where and for whom this new interface creates incremental value.

What changes when the interface is conversation?

Until the last era, the consumer arrived with an idea of what they wanted and browsed within a grid of products, display order, and highlights that the company decided to show, placing the intention of commerce ahead of the consumer's actual intention.

In a conversational interface, the consumer arrives with an intention, and the AI works to resolve it. A prompt is an instruction, and the AI that executes this instruction accurately and reliably wins the transaction, regardless of which brand is behind it. This inverts the logic that has dominated digital retail for decades because now the most relevant product is not the one with the most media budget, but the one that best responds to the command expressing the user's intention.

Shopify understood this logic that companies that only exist within their own app, waiting for the customer to come to them, have a real fragility. That's why they launched Storefront MCP, which allows any store to create AI agents capable of recommending products, assembling the cart, and guiding the consumer to checkout, all within a conversational interface.

And who has already taken the lead?

Nubank CEO David Vélez stated that they are not just adding AI to the bank; they are rebuilding the bank around AI. Nubank's AI Private Banker reached 15 million monthly active users. Ailo, iFood's AI assistant for orders via WhatsApp, registered a 48% higher chance of purchase completion and 33% faster orders in the first months.

According to CEO Andy Jassy, Amazon's shopping assistant Rufus generated nearly $12 billion in incremental sales with shoppers 60% more likely to complete a purchase. In May 2026, Amazon incorporated Rufus's capabilities into Alexa for Shopping, unifying the assistants under a single brand.

These numbers hide metrics on real user satisfaction and long-term impact, but they still generate positive expectations. McKinsey projects that agentic commerce, this model where AI agents become the new interface with the consumer, could move between $3 and $5 trillion globally by 2030, a margin that already says a lot about the level of uncertainty involved.

What does the enthusiasm tend to omit?

Execution errors have different consequences

LLMs make mistakes, and these mistakes are even more serious when they involve a transaction. When an agent executes a purchase, the error has direct financial consequences, and for these cases, confirmation, reversal, and liability mechanisms are still being designed by the technology industry.

Privacy and consent are open issues

To function well, an agent needs consent for access to financial data, behavioral history, and personal preferences. In Brazil, the LGPD and future specific regulations for AI are pushing in this direction, and in Europe, the AI Act already creates concrete obligations.

User experience is not uniform

There is a significant and persistent portion of consumers—by age group, digital profile, or personal preferences—who simply do not want an AI making decisions for them, especially when it comes to transactions with financial impact.

How can companies adapt?

iFood has 14 years of behavioral data, and Nubank built its model on over 100 terabytes of credit interactions. For a company without this history, entering the AI personalization game requires an infrastructure that takes years to build.

For small and medium-sized businesses, the window of opportunity lies in being found by agents, not in building their own agent. This means they should structure product data in formats that agents can read, such as open APIs and integration with WhatsApp Business, Shopify, or similar platforms; build a verifiable reputation that agents can analyze before recommending; and choose an initial problem to address with AI instead of transforming everything at once, such as customer service, lead qualification, or order triage.

For medium and large companies, where there is already a history, it is important to understand if the data is structured in a user-friendly way for a model to learn from it. Most of the time, it is necessary to resolve data fragmentation across systems, channels, and teams, addressing so-called organizational silos.

One route that has recently gained viability is the adoption of SLMs (Small Language Models), smaller, specialized language models. A compact model trained on the company's own data tends to outperform a generic LLM in the specific tasks of that context, with lower operational costs, lower latency, and without needing to transfer sensitive data to external servers. IBM with Granite, Microsoft with Phi, and Meta with Llama have bet in this direction.

For companies of any size, metrics need to change. For the conversational interface, what matters is the cost per task completed, intention resolution rate, and recommendation quality—cost per click, open rate, and session time are metrics that can lead the company to optimize for the wrong scenario.

The transition to conversational interfaces is happening, but the speed, actual adoption, and impact will vary enormously by sector, audience, and which problem AI can solve better than the previous channel.

The real risk is making expensive strategic decisions based on optimistic projections before clear sustainable patterns and business models are established. Observing, testing, and measuring what matters is what separates strategy that generates value from hype.