Learn what changes when data stops feeding only analytics and starts sustaining AI applications and agents in production.
To scale AI, the challenge is not only the model or the application. The data architecture also needs to be prepared to sustain this new level of use. According to Gartner (Apr 2026), organizations with greater data maturity achieve 65% better results in revenue growth and cost optimization. Still, only 7% of companies have data that is fully ready for AI, according to Cloudera and Harvard Business Review (2026). The challenge is to recognize whether the current data foundation can sustain AI applications and agents in production, including agents that act on their own.
Data readiness is usually assessed based on analytical workloads, such as dashboards and reports. But when data starts feeding AI applications and agents that guide decisions or execute actions, new requirements appear: context, governance, and control. This gap is what can keep projects from moving past the pilot phase and into production.
In Brazil, perceived readiness is higher than actual readiness
44% of Brazilian companies report high data readiness for AI, above the global average of 32%. The figure comes from the Cisco AI Readiness Index: Brazil, 2024. At first glance, this is a positive result, and it reflects how much the topic has already become a priority for Brazilian companies. However, it is worth looking at the operational numbers from the same study side by side. 71% of companies in Brazil still report fragmented data to varying degrees, and 53% say there is room to improve how they trace the origin and lineage of their data. It is no coincidence that these are the two points that weigh most when the final consumer of the data is an artificial intelligence operating in real time.
Governance is key to scaling AI
The most common objection treats governance as what slows delivery. The same survey cited above lists six moves that data leaders need to make by 2030. One of them reverses that logic: treat context and semantics as critical infrastructure, because that is what allows AI to operate with confidence.
This is similar to the reasoning we already covered in the article AI in Production: 4 Questions Every CIO Needs to Answer. It gains an extra layer when a process moves from executing part of a workflow to an agent, a topic we explored in AI Agent Teams: What Actually Changes Day to Day. Without adequate governance, a project may advance through the pilot phase but tends to hit bigger barriers later on. In an architecture prepared for AI, governance needs to be built into the foundation from the start.
Lakehouse, data mesh, data warehouse: data architectures for AI
A fragmented data foundation holds AI back. This is what consistently shows up in the main studies on the subject, including a 2025 McKinsey survey. It cites data quality and architecture as central barriers to the gap between companies that adopt AI and companies that actually scale it. So the practical question is which architecture fixes this: data warehouse, lakehouse, or data mesh. It is worth understanding the differences before deciding which one fits your company.
What is a traditional data warehouse?
A data warehouse is an environment optimized for analytical queries on structured data and the classic foundation for reports and dashboards. Data arrives already transformed, in a fixed format defined in advance. This works well for structured analytical workloads, but it can require additional capabilities when a company needs to combine structured and unstructured data, machine learning, and AI applications. This often forces an AI environment to be built separately, with duplicated data and pipelines between the two.
What is a data lakehouse?
A lakehouse is an architectural category that unifies data warehouse and data lake functions on an open data format, with a transactional reliability layer on top (such as Delta Lake). In practice, this makes it possible to store raw and structured data in the same place, run analytical and AI workloads on a shared data foundation, reduce unnecessary duplication of data and pipelines, and apply the same governance layer to all of it.
Databricks is one of the leading platforms for implementing lakehouse architecture, and dti has a formal partnership with the company, with external recognition. In 2026, dti was classified as a Product Challenger in the ISG Provider Lens Databricks Ecosystem Services Partners report, in two evaluated quadrants. Further details, along with other cases from the partnership, are in the article Databricks: How to Align Data Architecture and Real AI Innovation.
What is a data mesh?
A data mesh is a more organizational than technical model, for when scale requires distributing responsibility for data. Instead of a single central team being the only one responsible for integrating all of a company's data, each business domain takes ownership of its own data as a product, with defined contracts, quality, documentation, and access. A common core keeps standards and governance consistent across the whole company. This kind of scale is what sustains Fazenda de Dados, which dti helped implement at dsm-firmenich.
At Fleury, a structured data foundation enabled an AI application
Only 13% of patients with a formal recommendation to return, noted in their lab report, actually came back for follow-up. The radiologist identified the finding and wrote the right recommendation. The challenge came afterward: tracking thousands of reports month after month and cross-checking them against each patient's history to know who needed to be contacted for a return visit.
Ricardo Camilo, Data Architect at dti, explains how long this challenge went on:
"Healthcare data sits in very separate silos [...] getting it into the same data environment, with governance and security, is a challenge, it's not simple [...] if you overcome this first barrier, you unlock enormous potential," he said in this episode of the Os Agilistas (The Agilists) podcast.
The laboratory system, the radiology report, and the doctor's clinical note lived in different environments. Only after this data was brought into a shared environment, with governance, did it make sense to apply a language model to read the report and identify who needed follow-up. The governance design was not a detail added later. It is what made the project safe enough to work in day-to-day clinical operations. The model reads the report and identifies the finding, but the decision of whether a patient should be contacted belongs to the prescribing physician, not the system.
This project grew out of the Data Hub data architecture, built on Databricks. We detail that architecture in the case study Data Modernization: Architecture That Underpins AI Adoption at Grupo Fleury. The case shows how a well-structured data architecture can enable AI in production securely. When the goal is to expand the autonomy of agents, that same foundation starts to require additional controls over decisions, actions, and human intervention.
What still raises questions about data and AI
What is data architecture, and why does it matter for scaling AI?
Data architecture is the structure that organizes how data is collected, stored, integrated, governed, and made available for different uses. To scale AI, it needs to provide trusted data, context, governance, and appropriate access to the different flows, so that projects can move beyond the pilot phase into production with security and consistency.
What changes in data governance when the final consumer becomes AI?
Beyond access controls and data usage rules, AI requires clear policies on which actions it can take, under which conditions, and when human intervention is necessary.
Where should you start preparing your data before putting AI into production?
Start by mapping before automating: where each piece of data lives, who owns it, and which decisions AI can make without review and which it cannot. Investment in generative AI is not the problem. What is missing, in most cases, is a data architecture capable of delivering quality, context, governance, and controls for AI, including agents that operate in production. If your company has not yet assessed its data foundation against this standard, the checklist in the article Data Foundation for AI: Checklist for Executives Before Investing is a good starting point.