Analytics is as much about persuasion as it is truth-seeking

A few weeks ago, I had to put together a company update. I needed to look at a bunch of numbers, work out what story they told, and get everyone pointed in the same direction.

Here's what I actually did. I opened Claude Code, connected it to our Holistics MCP, and started asking questions. When something looked useful, I asked it to write the analysis into a markdown file, then copied the parts I needed into PowerPoint.

The slightly uncomfortable part is that I didn't rigorously verify every number. During the presentation, I spotted an error live in front of the team: the ending customer count for one period didn't reconcile with the starting count for the next.

Nothing particularly interesting happened. Nobody changed their mind, the conclusion didn't move, and we kept going.

Initially I thought the obvious lesson was that I'd been sloppy. But the more I thought about it, the less convinced I became. I think I was simply doing a different kind of analytics.

Two goals of analytics

In the last issue, I wrote about Vincent using an agent to reconcile our consolidated P&L against the warehouse. Looking back at that example next to what I was doing, the contrast is quite striking.

Vincent was in truth-seeking mode. Two numbers disagreed and he wanted to know why. If the answer was wrong, the task had failed. There wasn't really a version of that job where being approximately correct was good enough.

I was doing something different. I was trying to understand the broad picture, form a point of view, and put together an argument that could get a group of people aligned. I still needed the data to be broadly correct, of course, but if one number was 31% instead of 32% and correcting it wouldn't change the conclusion, I didn't care nearly as much.

I think of this as persuasion mode.

That doesn't mean manipulation or cherry-picking numbers. Inside organizations, people have different priorities, compete for attention and resources, and try to influence what happens next. Data is one of the things they use to test a view, sharpen it, explain it, and make a case.

That means internal analytics is often optimized for more than just accuracy. Time to value matters. Visibility matters. Presentation matters. And convincing power matters a lot.

These aren't two completely separate activities though. Most good analytics contains some of both. The difference is which goal dominates in a given situation. Sometimes the main job is to understand reality as faithfully as possible. Other times the main job is to make sense of a situation, form a view, and persuade other people to act on it.

Good enough can actually be good enough

Persuasion mode doesn't mean accuracy is irrelevant. I still need to believe the numbers are broadly right.

In my case, I have enough confidence in the AI that I expect it to get me into roughly the right neighborhood. I also have enough context about the business to notice when a number looks obviously wrong. And I need it to be accurate enough that a more precise answer wouldn't change the decision.

This is very different from something like finance, where rigor is often the job itself. If you're reconciling a P&L, being 1% wrong doesn't mean you've nearly finished the job; it means the job is still unfinished.

Obviously the boundary depends on the stakes. If I'm preparing something for the board, an investor, a customer invoice, finance, or anything where the number itself has consequences, I want rigor. But my guess is a surprisingly large amount of everyday internal analytics doesn't operate under that standard.

And this is where I think things become uncomfortable for people who build BI tools.

BI has mostly been built for truth-seeking

The implicit model behind business intelligence has traditionally been that there is a correct answer somewhere in the data, and our job is to help people get to it reliably. That's why we spend so much time thinking about governed metrics, semantic layers, lineage, certified dashboards, and single sources of truth.

I believe in all of those things. Holistics is very much built around them.

But increasingly I wonder whether we've been assuming that truth-seeking is the primary job people hire analytics to do.

Think about a normal week inside a company. A product manager is arguing for a roadmap change. A sales leader is explaining why a quarter went well or badly. An executive is trying to convince the company that one priority matters more than another. Someone is preparing a business case for hiring another five people.

All of these people are using data, but they aren't necessarily behaving like scientists trying to discover an unknown truth. Quite often they already have some intuition of the situation and are using analysis to test it, sharpen it, explain it, and eventually persuade other people that it is a sensible way to see the world.

Again, I don't think that's bad. Organizations need shared narratives in order to operate. Someone has to turn a pile of facts into a coherent explanation of what is happening and what everyone should do next.

The governed path is no longer automatic

In my case, Claude was connected to Holistics through MCP. I skipped the BI interface, but not the governed layer. The agent was still working with modeled data and shared definitions.

But recently I've been hearing more cases of business users connecting agents directly to the warehouse, or even straight to source systems like Shopify or HubSpot. Sometimes they're bypassing an existing governed layer. In other cases, there isn't really one to begin with, and the users may not feel the need to invest in building one.

Why would someone choose that path?

Part of it may simply be convenience. I think the deeper reason is that, for some kinds of internal analytics, governance is not the thing the user is optimizing for. If an agent alone can already get business users to a useful answer, will they still see enough value in having a governed data layer underneath?

If you're in truth-seeking mode, governance is central to the job. In persuasion mode, speed, clarity, presentation and getting to something useful enough to support a decision may matter more.

For a long time, users didn't have much of a choice. If you wanted to analyze company data, you usually went through infrastructure built by the data team, and governance came bundled. Going around it generally meant knowing SQL and doing the work yourself.

Agents essentially flip that upside down. A business user can now choose between asking the governed layer and asking the raw systems directly — or, in some cases, never building the governed layer in the first place.

Parting question

As data leaders, we all understand why governance matters, and why having a shared semantic layer matters.

But here's the question I have for you, my fellow data leaders:

If you're a data leader working with an executive who thinks like the one I've described above: they trust the AI enough, they trust their own judgment, and they're comfortable with "good enough" accuracy if it doesn't change the decision-making.

What would you do?

Hit reply and let me know.