Nobody would staff it, so I made the ask smaller.

Leadership wanted Preqin’s insight problem fixed and had separately concluded that a project couldn’t do it, so nothing was staffed. The PM agreed. Meanwhile investment professionals were spending weeks assembling findings the platform already held.

That is a specific kind of problem to have. The company already owned the expensive half: models capable of generating hundreds of findings a day. What it didn’t own was any way to deliver them, so subscribers were paying for a data platform and then supplying the labour that turned it into an answer. Value the customer has to assemble themselves is value the renewal conversation never gets credit for.

There was no evidence I could put in front of anyone, because producing it required the staffing nobody would grant. So I stopped trying to win the argument and changed what was being asked for: not a commitment, a question. Inside that framing I designed the insights experience — a templated feed of machine-generated findings ranked by actionability.

One constraint sat behind every design decision. The backend couldn’t deliver in real time and would not before we shipped, which is what made the findings arrive pre-composed rather than computed. Time-to-insight fell 90%, weekly actives 20%, upsell 8%.

Investment professionals worked through scattered dashboards to assemble a picture the platform could in principle have assembled for them. Decisions came slowly and adoption suffered.

Interviews with 12 asset managers found three things. Volume and noise made insights impossible to prioritize. Contextual metrics sat across separate dashboards, so patterns were only visible to someone already looking for them. And peripheral signals (alerts, comparisons, portfolio anomalies) were ignored almost entirely.

On the other side, the models could generate hundreds of findings a day, and the backend couldn’t deliver any of them in real time.

This mattered because it changed what kind of project it was. Read as an insight problem, it needs better models, more data science, and a research programme, all of which are expensive and none of which anyone was going to fund. Read as a delivery problem, the models are a given and the work is getting a few hundred findings a day in front of someone in a form they will read.

The second version is much smaller than the first, and it is also the accurate one — the models were already producing more than anyone was consuming. That gap is the whole opportunity, and it is why the eventual ask could be made small enough to say yes to.

Worth being exact about what moved here: it wasn’t persuasion. No evidence changed anyone’s position, because there was none. What changed was the size of the ask.

We shipped the Insights Feed: templated findings from the generative models, ranked by actionability, each expanding from a headline metric into the narrative behind it. It fit existing workflows and required no training.

The template is the piece with the longest reach. Because findings arrive pre-composed into slots, a new finding type is a data-side change rather than an interface project — the team can ship a kind of insight nobody had thought of without design or front-end work, and without anyone learning a new tool. That converts the roadmap from a series of features into a supply question.

Ranking with its reasoning on the card does something similar for trust: a rank a user can check is a rank the team can change. Tuning the model afterwards doesn’t require re-earning the user’s confidence, because the confidence was never in the ranking, it was in being able to see why.

The organisational unlock is the one I’d point at first, though. A thing leadership had concluded was not viable shipped and produced numbers, which changes what can be asked for next in a way no individual feature does.

Leadership had already reached a conclusion, the PM agreed with it, and the evidence that would have changed either mind could only be produced by the staffing being refused. Every additional hour spent building a case would have been spent on a case with nothing in it.

What I do differently since: when I can’t change what someone believes, I look at what I’m asking them to agree to. A commitment needs belief. A bounded question only needs to be worth answering, and that is a much lower bar for someone to clear — which is why the ask got taken up on terms nobody had to defend.

It is not a general-purpose move and it has a specific failure mode I was lucky to avoid. An experiment nobody committed to can be under-resourced into a false negative, and a negative result on an idea like this can bury it for years. It only works where the question is genuinely cheap to answer either way. If the cheapest honest test is still expensive, shrinking the ask just relocates the problem.