I recently joined Marcelo Oses, Head of Data & AI at Porto Bank, on stage at the Databricks Data + AI World Tour in São Paulo - one of the biggest Data & AI events in Latin America. We shared how Indicium AI worked with Porto Bank to turn a prime-time TV ad into a real-time, AI-powered sales engine, and what it really takes to get AI into production.
Why most AI never makes it into production
Research from the Massachusetts Institute of Technology found that 95% of corporate generative AI pilots show no measurable impact on profit and loss - numbers that should concern any leader investing in AI. But the problem isn't the technology. Most companies start at the wrong end: they begin with the product when they should begin with people.
AI is not just training, and it is not just technology. You build it with people. That means you measure maturity, define the right platforms, keep intellectual capital inside the business and turn new skills into real benefits for the company. AI is a means for employees, not an end in itself.
It was that thinking that shaped Assimov, the multi-agent architecture we built with Porto Bank.
Turning a TV moment into thousands of conversations
Porto Bank is one of Brazil's largest insurance and financial-services companies. It promotes a long-term financing product on a prime-time TV show that reaches 30 to 35 million viewers. A QR code on screen sends viewers to WhatsApp, so a single broadcast can generate tens of thousands of inquiries at once.
Advising on this product means considering a customer's goals, their ability to pay and the alternatives available. That work traditionally takes an advisor more than 30 minutes. Because of the nature of live TV, on these occasions it had to happen for thousands of people simultaneously, in under two seconds each. Financial data, access and traceability also had to stay under control at every step.
A multi-agent engine built for speed and scale
Assimov is an orchestration layer that receives the customer's intent, coordinates specialist agents and returns a personalized recommendation within a single WhatsApp conversation. A supervisor handles journey routing, short- and long-term memory and response validation. Beneath it, five specialist agents cover:
- Business rules and product knowledge
- Credit parameters
- Client profile
- Real-time propensity scoring
- Sales qualification
Speed came from architecture choices. Running agents in parallel rather than in sequence cut input/output latency by 40%. Today, 95% of recommendations arrive in under two seconds, even at peak demand.
Governance was built in from day one. Brazil's data protection rules set a high bar for consent, access control and auditability. Databricks' Unity Catalog gives Porto Bank central control over access to sensitive data, plus full data lineage.
Two seconds, 315% more leads, zero additional headcount
When Marcelo walked through the numbers on stage at the Data + AI World Tour, the room paid attention. Across seven campaigns, the solution has delivered 24% more customers and more than US$15 million in sales. Leads grew by 315%, from 20,573 to 85,311, with a 20x return on the channel's operating cost. Porto Bank achieved all of this without adding a single employee.
From campaign to competitive advantage
Our goal when we started working with Porto Bank was always bigger than one campaign. We set out to replace static models with intelligent interfaces and turn AI into a durable competitive advantage.
The same foundation now supports journeys in credit, collections, customer service, commercial performance and internal operations. Porto Bank has 160 agents in production, and new journeys reach the market in 40 days. A separate AI model acts as a judge, checking every response before it reaches a customer. Most telling of all, business teams create 58% of new agents themselves. That's what people-first AI looks like.
The biggest mistake I see enterprises make is speeding up processes faster than they can handle. Speed matters, but only when the foundations are ready. Successful AI comes from a disciplined sequence: map the process, run rapid proofs of concept, listen to customer feedback, optimize, and then execute.
Getting AI off the page and into production
At Indicium AI, our focus is transforming enterprises by taking AI off the page and deploying it into production at scale. We do that side-by-side with partners like Databricks and Anthropic. They bring enterprise-grade platforms, strong governance and frontier AI. Our specialist teams bring the strategy, engineering and operational expertise to make it work.
If you're ready for your enterprise to move AI from pilot to production, get in touch.
