Augmented analytics with generative AI | SUMāTO

For years, advanced analytics lived behind an invisible barrier: to ask your data a question, someone had to know how to write a query, build a dashboard, or wait on the BI team. In 2023, generative AI is tearing that barrier down. For the first time, a sales manager can type «show me why sales dropped in the South last month» and receive, in seconds, an answer with context and narrative. This is augmented analytics, and it changes the fundamental question: it is no longer about who knows how to use the tool, but about what you want to understand.

The short version: Generative AI lets you converse with your data in natural language and obtain not just figures, but explanations and actionable narratives. The leap is from static dashboards to a continuous dialogue. The challenge is not technical but one of governance: accuracy, traceability, and trust are the conditions for scaling responsibly.

From dashboards to conversations

The traditional dashboard answers the questions someone anticipated when designing it. If your question was not foreseen, you are stuck: you filter, export to a spreadsheet, or open a ticket with the data team. Generative AI inverts that model. Instead of navigating prebuilt views, you pose the question and the system builds the answer on the fly.

This transition has three practical consequences:

• Immediate iteration: every answer opens the door to the next question, with no friction and no intermediaries.
• Democratized access: people with no training in SQL or BI tools can explore data on their own.
• Fewer bottlenecks: the data team stops being a query desk and focuses on modeling, governing, and enabling.

How natural-language querying works

Behind the conversation there is a concrete mechanism. The language model does not «invent» the figures: it translates your question into a verifiable operation over the data. The most robust pattern right now is what is known as text-to-SQL or text-to-query, where the model generates a structured query that runs against the database and then drafts the answer from the real result.

For this to work well, several elements matter:

• A semantic layer: clear definitions of what each metric means («net revenue», «active customer») so the model does not interpret them as it pleases.
• Metadata and a data dictionary: table names, columns, and relationships the model needs as context.
• Context retrieval (RAG): the model consults schemas and documentation before generating, rather than relying solely on what it «remembers».

The difference between a fragile pilot and a reliable system almost always lies in this plumbing, not in the model itself.

From data to insight: narrative generation

The truly novel part is not obtaining a number, but obtaining an explanation. Generative AI can move from «sales fell 12%» to a paragraph that identifies which segments, which products, and which period concentrate that decline, and expresses it in business language.

This enables what we call narrative analytics:

• Automated summaries: an executive report drafted from the quarter's data, ready for human review.
• Anomaly detection and explanation: the system flags a figure that is out of range and proposes hypotheses about its cause.
• Comparative context: not just the what, but the relative to what (prior month, target, same period last year).

The value lies in turning data into a decision. An insight no one understands does not drive action; a clear narrative does.

The risks: accuracy and governance

It would be irresponsible to present this without its limits. The same generative capability that produces fluent narratives can produce incorrect statements with complete confidence. In analytics, an error written convincingly is more dangerous than an obvious one.

The main risks we see today:

• Hallucinations: the model can invent a figure, a trend, or a relationship that does not exist in the data.
• Mistranslated queries: an ambiguous question can generate a query that is syntactically valid but conceptually wrong.
• Inconsistent definitions: without a semantic layer, two people can obtain two different «truths» for the same metric.
• Governance and privacy: who is allowed to ask what? Natural language does not eliminate the need for access control over sensitive data.

How to mitigate: trust by design

Mitigating these risks is not optional; it is the condition for taking this technology beyond the pilot. The practices we recommend:

• Prioritize traceability: every answer must be able to show the query and the data that generated it. If you cannot audit the how, you should not trust the what.
• Define metrics once: centralize business logic in a governed semantic layer, not in each conversation.
• Keep a human in the loop: for high-impact decisions, the system proposes and the person validates.
• Inherit permissions: the conversational assistant must respect exactly the same access controls that already govern your data.
• Evaluate continuously: measure, against a benchmark set of questions, how many answers are correct, and monitor that figure over time.

The principle is simple: speed without trust is not an advantage, it is a liability. You can dig deeper into how we structure these layers in our analytics practice.

Starting without rebuilding everything

Good news for data teams in LATAM: adopting augmented analytics does not require rebuilding your platform. Most organizations can start with what they already have.

• Scope the domain: choose an area with clean, well-understood data (sales, for example) before opening the assistant to the entire company.
• Document first: the greatest return usually comes from tidying up definitions and metadata, something useful with or without AI.
• Design with an AI-first mindset: think of the conversational experience as part of the data product, not as a last-minute add-on.

Frequently asked questions

Does generative AI replace data analysts?

No. It changes their role. Repetitive extraction and reporting tasks are automated, but the need grows for those who design the semantic layer, govern quality, and validate results. The analyst moves from running queries to curating the system's trustworthiness.

How do I keep the system from inventing figures?

By connecting the model to your real data through verifiable queries, not by asking it to remember numbers. If every answer is built from an executed, auditable query, and you demand traceability, numeric hallucinations drop drastically.

Do I need my own model or one trained on my data?

In most cases, no. Today it is more effective to give context to an existing model through metadata and retrieval (RAG) than to train one from scratch. This reduces cost and time and keeps your data under your control.

Is it safe for sensitive data?

It can be, if the assistant inherits your existing access controls and it is clearly defined who can ask what. The conversation must not become a back door that bypasses the data governance you already have.

The first step

Augmented analytics with generative AI is not a distant promise: it is a capability available today, whose success depends less on the model and more on the discipline with which you organize your definitions, your governance, and your traceability. The time to experiment, with a scoped domain and the right controls, is now.

At SUMāTO we help data teams in LATAM design this transition from dashboards to conversations, without sacrificing trust. If you want to explore how to apply it in your organization, let's talk.