Forward Pass #2

The Grunt Work Goes First

Every analyst has the task they dread. For me it was the export that arrived broken every single month. Merged cells where headers should be. Dates stored as text. Subtotals sitting in the data rows like landmines. Four spellings of the same region. You stop thinking about it after a while. You just accept that the first morning of the month belongs to the cleanup, not to you.

That task is the first thing the agents take. And here is the part worth sitting with, especially after last issue: they are good at it.

As of mid 2026, Copilot in a Fabric notebook will profile the file, sample it, and flag the quality problems before you write a line. Data Wrangler will suggest the operations, preview every change on a sample, generate reusable code, and never touch your original data until you say so. The dreaded morning becomes about ten minutes.

Notice where AI took the bite. Not from the top of the job, where the judgment lives. From the middle, where the reshaping lives. The mechanical work goes first, and the mechanical work was never why any of us got into this.

That is the reframe I want you to hold. The work that leaves first is the work you would have paid to never do again. Nobody became an analyst to unpivot columns and reconcile four spellings of Texas. When that disappears, you have not lost the job. You have lost the chore.

In the Loop

The agent supplies the labor. It cannot supply the standard, and the standard was always the actual job.

Three calls stay yours, every run.

1. You decide what clean means.

Which spelling of the region is canonical. Whether a blank cell is a zero or a missing value, because those are different numbers and the file does not know which one you meant. The agent will pick one. Only you know whether it picked right.

2. You approve the plan.

Dropping a column is a business decision wearing the costume of a formatting step. The agent proposes it in half a second. You are the one who knows that field is the one Finance needs at quarter close.

3. You defend the result.

When someone asks why the numbers moved, “the agent cleaned it” is not an answer. You own the explanation to whoever consumes the data, which means understanding every change well enough to stand behind it.

That is the shift in one line. Less scrubbing, more judgment. The cleanup leaves your plate. The authority over what clean means does not.

Try This

Take the messiest export you own. Drop it into a Fabric notebook and, before you ask for anything to be fixed, ask Copilot to tell you what is wrong. It will profile the table, list the columns and types, and flag the data quality issues it found. Then open Data Wrangler and watch it propose operations, preview each one, and write the code without overwriting a thing.

The tell is the explanation. Make it justify every change in plain language. A clean you can defend to whoever owns the source is a good clean. A clean you cannot follow is a liability wearing a nice coat. If you cannot say why a column was dropped, that is the signal to slow down, not to hit run.

Prompt & Circumstance

One prompt worth stealing: audit before you act.

The instinct is to type “clean this file.” Resist it. Tell the agent to look before it touches anything:

Profile this dataset first. List every column, its type, its missing-value count, and any values that look inconsistent or out of range. Propose a cleaning plan as a numbered list. Change nothing yet. I will approve the steps.

That flips the default from act-then-explain to explain-then-act. It is also how the tool wants to work. By default, Copilot asks for your approval before it runs a cell. Keep that setting on. The two minutes you spend reading the plan is the exact moment your judgment enters the pipeline, instead of arriving after the damage.

Signal

The cleaning is getting more autonomous. In an updated notebook Copilot, Microsoft moved from single-cell suggestions to an assistant that plans, generates, refines, and validates a multi-step transformation across a whole notebook. The direction is unmistakable: less “help me write this one step,” more “build the pipeline and show me your work.” The open question for all of us is how much of that plan you still author, and how much you simply approve.

The Part the Speed Can Hide

Handing off the grunt work is a real win, and I do not want to talk you out of it. But it comes with one condition the demos gloss over. An agent can make a file clean in minutes. It cannot tell you whether clean is the standard your business actually needs, or whether the source feeding it was worth trusting in the first place. Speed on a shaky foundation only gets you to a confident answer faster.

So before you hand the cleanup to an agent, it is worth knowing whether the foundation underneath it can carry the weight. That is what we check. If you want to see whether your data is ready to be automated safely, or only ready to be automated fast, let’s talk. Our free Data Health Check looks at exactly that.

Clean, though, is not the same as correct. That is the story of the next issue.

Until the next pass,

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

Sources:

LinkedIn: Connect with me

Forward Pass # 2: Originally Posted on LinkedIn, July 28, 2026

Scroll to Top