An ESG score that meant nothing to the fund reading it.

Preqin was opening a new product line — ESG data sold into private markets — and presenting it as a static snapshot. Two users could look at the same figure and neither could say what it meant for them, which is a hard thing to charge for. Limited partners were worried about greenwashing; general partners wanted to know what their peers disclosed.

That made the product problem and the commercial problem the same problem. An uninterpretable number cannot be demoed either, so a score nobody could read was as much a sales obstacle as a usability one, and the fix had to work in both conversations at once. The product had to put the number in context: this type of fund, this geography, these industries.

It took three versions of the module to get there, and each failed the same question in a different way — where do I look first. Version three shipped, alongside 7 contracts and 50+ pre-sales calls, roughly $560k of contract value the modules contributed to at the account values in the pilot group.

Then I tried twice to finish the job those modules opened. Both attempts failed, and why they failed is the most useful thing here — it turned out to be a limit on what this product was allowed to be, not a limit on how well I built it.

The two user types are asking opposite questions — one is trying to catch a fund overstating itself, the other is trying to see how it compares — and it would have been easy to read that as two products. It isn’t. Both questions are unanswerable for the same reason and become answerable through the same mechanism: put the number against a relevant peer set and show which way it has been moving. Suspicion and benchmarking are the same operation pointed in different directions.

So the work wasn’t about surfacing more data. There was no missing data. It was about moving from a static snapshot to momentum over time and specific gaps against comparable funds.

That was the project’s most useful finding, and it doesn’t show up in the results. A data product can credibly identify a gap and lose that credibility when it tells users what to ask about it. Identifying a gap is evidence. Recommending questions is advice. These customers were buying something to check their reasoning against, not advice.

The commercial result is the visible outcome and the boundary is the durable one. The company came out of this knowing something about its own product line that it did not know going in: these customers will buy evidence and will not buy advice, and the wall between the two sits somewhere before a prepared list of questions. That is a positioning constraint on everything the data business might build next, and it cost two prototypes rather than a launch to find.

The three-version sequence produced a smaller reusable result. Tables, then a multi-dimensional risk view, then heatmaps over the tables — that is one question asked three times, and the version that won kept the losers underneath rather than replacing them. Exact figures stayed available beneath the shading because diligence still needs them.

Components from the modules went back into the design system, which is the least interesting thing here and the one most likely to still be in use.

Neither the drawer nor the wizard failed on usability. People could operate both. What they rejected was the proposition underneath — a tool that had been checking their reasoning was now doing the reasoning, and the part it took over was the part they believed they were paying for.

That class of risk is invisible to the testing I was doing. A usability session asks whether someone can complete the task, and both prototypes passed; nothing in that method asks whether they want the task done for them. I ran the right method against the wrong question, twice.

What I do differently since: when a feature extends what a product is for rather than how well it does it, I test the positioning claim before building the interface — describe the changed product and ask what they would stop doing themselves. That is a cheaper experiment than either of the two I ran, and it is the one that would have found this. The note below is about the thing I still don’t know, which is where exactly the line falls.