When we talk about the adoption of enterprise Artificial Intelligence, the script in Brazilian companies tends to be very similar: the board demands an "internal ChatGPT" capable of answering complex questions about sales, inventory, industrial costs, or medical records. The immediate problem? The data needed to feed this AI is scattered across dozens of legacy systems, ERPs, CRMs, and departmental spreadsheets.
The traditional IT response has been to overload the data engineering team. They are asked to build endless pipelines to extract, transform, and load (ETL) all this information into a centralized repository or vector database before the AI can even read the first line. This process is slow, expensive, and creates a monumental bottleneck. However, a recent AWS publication on federated data access patterns for AI agents signals a paradigm shift that promises to bypass this technical hurdle.
The Promise of Federated Access for AI
The core concept presented by AWS involves the use of the Model Context Protocol (MCP) and data federation architectures. In practical terms, this means that AI agents no longer need to wait for data engineers to move every gigabyte of information into a central data lake. Instead, the AI receives the tools and permissions to query enterprise data where it already resides.
If a retail manager asks the AI about the profit margin of a specific product over the past week, the intelligent agent can translate that intent, connect directly to the transactional database (or a read replica), execute the necessary SQL query, interpret the result, and deliver the answer back to the user. This decentralized approach promises unprecedented agility, eliminating months of integration routine development.
The Harsh Reality: AI Does Not Fix Bad Data
At e-Setorial, with nearly 30 years of hands-on experience in data administration and modeling, we have seen many technological "silver bullets" come and go. Federated access for AI agents is undoubtedly a brilliant architectural evolution. However, it brutally exposes a much older and deeper problem: the quality and organization of data at its source.
The idea that an AI can simply plug into your legacy ERP and extract magical insights is a dangerous illusion. If your database has tables without declared foreign keys, columns with cryptic names like CLI_TIPO_2_OLD_BKP, and essential business rules that exist only in developers' heads or within application code, the AI agent will do exactly what an unsuspecting junior analyst would do: get the answer wrong.
The only difference is that, with federated access, Artificial Intelligence will generate hallucinations based on real data, delivering incorrect answers in milliseconds and with unwavering confidence. In critical sectors like healthcare, finance, or manufacturing, making strategic decisions based on these responses can be catastrophic.
The Non-Negotiable Foundation: Modeling and Data Dictionary
For data federation to work and for AI to deliver real business outcomes, data infrastructure must be treated with professional rigor. It is not just about granting network access or configuring APIs. It is about semantics and governance.
- Professional Data Dictionary: AI needs context. An up-to-date data dictionary that explains exactly what each table and column represents is the map that keeps the AI agent from getting lost. If the calculation rule for "Net Profit" is not documented in the database, the AI will inevitably confuse it with gross revenue.
- Dimensional Modeling (Kimball): Even with federated access, querying highly normalized transactional systems for analytical purposes is inefficient. Building clear dimensional models (facts and dimensions) remains the best way to present data for consumption, whether by BI tools like Power BI or by large language models (LLMs).
- Quality and Curation: Tools like Apache Hop remain vital. Even if you don't centralize all your data, you still need agile pipelines to cleanse, standardize, and validate information across silos before AI consumes it. Data quality is the absolute prerequisite for Artificial Intelligence.
- Database Performance: AI agents generate SQL queries autonomously and, often, unpredictably. If your database performance is not up to par — with proper SQL tuning, correct indexing, and smart partitioning — the AI's queries will bring down your transactional system. Optimizing the database is much better and far cheaper than simply paying for cloud hardware upgrades.
Conclusion
The shift toward federated access, as highlighted by AWS, is a fundamental step toward democratizing AI adoption across enterprises without bottlenecking data engineering. However, cutting-edge technology does not override the foundational rules of data science. Quite the contrary: it makes them even more critical.
Your company will only be ready to delegate analysis to autonomous AI agents once your house is in order. Well-managed, documented, and structured data is the only shortcut-free path to ensuring that Artificial Intelligence becomes a value-generating tool rather than an operational risk generator. The hype fades, but sound data engineering and administration keep delivering results.
Giving an AI direct access to data without organizing and documenting your house first is like handing a map with wrong street names to a rideshare driver: they will arrive incredibly fast, but at the completely wrong address.