Forward Pass #1

Autocomplete Was the Warm-Up

For three years, AI in analytics meant suggestion. A measure half written for you. A column named. A chart proposed. Easy to wave off as autocomplete with a marketing budget.

That framing expired this month. At Build 2026, Microsoft drew a hard line between AI that suggests and AI that executes, and put Fabric on the far side of it. The Fabric data agent is now generally available. The pitch is no longer “AI helps you write the query.” It is “describe the outcome, and the system builds it.”

Under the hype, agentic means something precise. An agent plans steps, acts against your data, and checks its own results, instead of handing you a snippet and waiting. Autocomplete finished your sentence. An agent runs the errand. The Fabric data agent takes a plain English question, picks the governed source, queries it, and answers, with no analyst in the loop to turn the request into code.

If that put a small knot in your stomach, you were reading carefully. Writing the query, building the model, shipping the report: for most of us that is not one task among many. It is the craft. Watching a system do it from a sentence is strange to sit with, and I would not trust anyone who told you otherwise.

So sit with it. Then look at what the machine did not do.

It did not decide the question was worth asking. It did not know that intercompany transactions come out, or that last April is unusable because of the ERP cutover. It did not notice the number was wrong. It cannot answer to the board on Thursday.

The work that survives is the work that was always the hard part. We used to bury it under the typing.

Where the Human Stays in the Loop

Four places, concretely:

1. You frame the question.

An agent answers what you ask. It has no view on whether that was the right question, and a precise answer to the wrong question is the most expensive artifact in analytics.

2. You define what correct means.

Grain, filters, exclusions, edge cases, the house standard for margin. None of that lives in the data. It lives in the business, and reaches the agent only if a person puts it there.

3. You own the meaning.

The semantic model is not plumbing. It is a written record of what your organization believes its numbers mean. Agents consume that record. They do not author it.

4. You are accountable.

The agent has no stake in the outcome. It will not lose the client or answer for the forecast. Someone signs their name, and that someone is a person.

Read those four together and a job description falls out. Less production, more direction. The analyst who pulls ahead is not the fastest at writing DAX. It is the one who is clearest about what correct looks like, and can prove it.

That is not a smaller role, but a better one.

Try This

Pick a measure you would normally write by hand. Open Copilot in the DAX query view of a model you trust and ask for it in plain English. As of mid 2026, Copilot generates DAX from a plain language description, and the Fabric warehouse turns a question into T-SQLthe same way.

Then do the step most people skip. Read what came back as if a new hire wrote it. Copilot is strong on simple aggregations and noticeably weakeron complex filter context and time intelligence, where it returns code that runs clean and answers the wrong question. The syntax is rarely the problem. The logic is.

Two minutes of this teaches more than any keynote. You feel the job change under your hands, and how much of it is still yours.

Prompt & Circumstance

One prompt worth stealing: hand over the spec, not the sentence.

Most people prompt the way they text. “Give me a profit margin measure.” You get something plausible and ungoverned. Give the agent the brief you would give a junior analyst:

Write a DAX measure for net profit margin. Grain is per product, per month. Use the Sales and Costs tables. Exclude intercompany transactions. If revenue is zero, return blank, not an error. Match the format of our existing margin measures.

You stated the grain, the filters, the edge case, and the house standard, so there is far less room for a confident wrong answer. Copilot now accepts up to 10,000 charactersper prompt for exactly this.

Notice what that prompt is. Domain knowledge, written down. The agent supplied none of it.

Signal

At Build, Microsoft showed Agent Skills for Power BI: describe a report in plain language, or hand over a screenshot, and an agent builds the model, generates the pages, and iterates on the visuals. Early, uneven, and aimed at the gap between a question and a finished dashboard. We will take it apart in a future issue.

Before You Let an Agent Loose

Here is the catch the demos skip. Copilot writes wrong DAX when relationships are ambiguous and cardinality is off. An agent answers with total confidence from a semantic layer nobody governs. These capabilities amplify what is already there. They do not repair what is not.

Which is good news, if you consider who does the repairing. The highest leverage AI work this year is not a clever prompt. It is a clean, governed, well named model, built by people who understand the business it describes.

If you want to know whether your environment is ready for agents or merely exposed to them, let’s talk. Our free Data Health Check looks at the foundations that decide whether agentic analytics becomes leverage or liability.

Until the next pass,

Isaac Truong | Founder, Allston Yale | Texas Power BI and Microsoft Fabric consulting

Sources: