From days to minutes: ask your data in plain language

How much does it cost your company to answer a question like “what did this facility consume last month compared to the previous one”?

Not the cost of the data — the data is already in the lakehouse, paid for and governed. The cost of the answer: someone opens a ticket or messages an analyst, the analyst queues it, writes the query, exports to Excel, sends it over… and the answer arrives days later. By then, one of two things: either the decision was already made without the data, or the follow-up question (“and broken down by hour?”) restarts the whole cycle.

That cycle is invisible on the P&L because nobody invoices it. But it’s why most operational decisions are made on intuition, and why follow-up questions don’t get asked.

A Spanish energy company asked us to eliminate it. Here’s what changed:

The after: ask where you already work

Today, any employee at that company can type the question in plain language, exactly as they think it, from the tools they already use:

• In Teams, via M365 Copilot — for everyone, nothing to install. The answer arrives in minutes.
• In Claude — for analysts and data-savvy profiles. The answer arrives in seconds.

And the answer isn’t a dump of rows — as the capture at the top shows, the agent understands the question, queries the lakehouse, interprets the figures and replies with the magnitudes explained. The follow-up question — the one that used to restart the days-long cycle — is simply the next message.

Three doors, one identity

Seen from a distance, the system is simple: employees come in through the tool they already use, and every door leads to the same place.

The third door, VSCode, is for developers — same answers in seconds, inside their editor.

What makes this deployable to the whole organization is the arrow on the right: the agent queries the data with the identity of the person asking. The same Microsoft account an employee uses to open Teams is the one that reaches the data. No new passwords, no special access, no shared accounts.

The objection that matters: what about control?

It’s the first right question from any leadership team: if everyone can ask, who controls what they see?

The answer is that control wasn’t added to the agent — it’s inherited from the one that already existed:

• Everyone sees exactly what they could already see. Data permissions are still defined by the company where they always were: in its data platform. The agent opens no new doors; if someone asks about data they don’t have access to, the answer is that they don’t have access.
• Read-only. The agent queries; it cannot modify, delete or write anything. It’s built so it can’t, not instructed that it shouldn’t.
• Everything is logged. Every query leaves a trace: who asked, what and when.

Asking stopped being a privilege of those who know SQL — without ceasing to be governed.

What it takes to have it

Less than it seems. The two expensive ingredients are already in place at most mid-size and large companies:

• Organized data — a lakehouse or modern data warehouse (in this case, Microsoft Fabric) with its permissions defined.
• Corporate identity — Microsoft 365, the accounts people already work with.

The missing piece — the agent connecting both while respecting permissions — is a thin layer on top of what already exists, not a year-long transformation project. And because its business knowledge (what each table, code and unit means) lives in versioned documents rather than code, it improves week by week without redeploying anything.

---

The technical detail of how identity travels end to end — and why most “chat with your data” systems fail this test — is in A data agent with real identity.

How many questions go unasked in your organization because answering them takes days? Let’s talk.