He Tried AI Without Me. Then He Checked the Numbers.

The surprise wasn’t just how quickly the analysis came back.

It was the relevant information he hadn’t asked for.

I had been working with a CFO on using AI. After that work, he was willing to try a new task without me sitting beside him.

He needed to analyze a property. Work like that can involve a lot of gathering information, checking numbers, and working through calculations before there’s anything useful to decide from.

He gave the agent the task, supplied the relevant context, and asked for specific information.

It returned what he requested. It also brought back additional numbers that were relevant to the decision.

That caught his attention.

The result came with something he could check

A confident answer wouldn’t have been enough.

The agent showed its work. He could look up the information it used and check the calculations instead of trying to reverse-engineer how it reached the result.

He did that review. According to his account, the information and arithmetic checked out on the first run.

He described it as days’ worth of analysis coming back in minutes. That’s his comparison with the work he expected, not an independently timed savings study or a promise about another task.

The useful result was more than speed. He had material he could inspect and use.

The financial decision still belonged to him.

A useful handoff leaves room for relevant questions

The agent didn’t stop at the exact numbers he listed. It recognized that other information would matter to the task and included it.

That doesn’t mean every extra number an agent provides is useful or correct. It means there was something worth reviewing.

The person who understands the business still needs to decide whether the information applies, whether its sources are appropriate, and whether anything important is missing.

In this case, the additional information checked out and made the decision-making work easier.

That’s a different experience from receiving an answer that sounds impressive but leaves you wondering what it’s based on.

One successful task made the next task easier to try

Since then, he’s been asking AI to work through more things.

We’ve also explored presenting reports as interactive HTML documents instead of only Word files or PDFs. When it fits the reporting need, charts and inputs can be adjustable rather than frozen on a page.

A static document can still be the right answer. The point is choosing a format that supports how someone will use the information.

If a reader needs to compare scenarios, an interactive view may be useful. If they need a fixed record, a PDF may make more sense.

The work should determine the format.

Coaching should carry into the next task

One of the encouraging parts of this experience is that he tried the analysis independently.

He had enough confidence to give the agent a real assignment and enough understanding to check what came back.

That’s what I want coaching to lead toward. Not someone depending on me to write every instruction, and not someone accepting every answer because the first attempt went well.

A successful first result is a reason to keep learning. It isn’t a reason to stop reviewing.

If you have a recurring analysis, check, or report you’d like to work through, bring a real example to a 90-Minute Workflow Optimization Session.

We’ll work toward a useful first version when the scope allows, or a clear blueprint you can continue from. The task, the evidence, and the decisions that remain yours come first.

The goal isn’t another impressive demonstration. It’s knowing how to give useful work to the tool and judge what comes back.