The machine that asked better questions

I was researching different ways to structure a database that would improve on how well it generated answers to query’s when I stumbled into something that resonated strongly with the what I do all the time: train cross-functional pharma teams on how to get access to and engage healthcare stakeholders.

In one of the paragraphs, Claude (the AI) gave me a summary of STORM, a research system out of Stanford. You give it a topic, it goes away and comes back with a long, sourced article written from nothing. I thought that sounded very interesting and as I went further down the rabbit hole I struck gold.

What STORM actually does

Most of us would research a topic the obvious way. Sit down, think of what we want to know, google or ask your preferred AI and document the answer. STORM was built because that obvious way produces shallow research. The first questions you think of are the ones everybody thinks of, so you get the article everybody could already write.

So STORM does two things before it writes a word.

First it finds perspectives. Before asking anything, it surveys what already exists around the topic and works out the different points of view the topic can be examined from. The regulator’s view. The critic’s. The practitioner’s. Each one is a different person to ask as.

Then it holds conversations. For each perspective, a curious writer questions an expert who is only allowed to answer from real, retrieved sources, not from memory. The writer reads each answer and asks a sharper follow-up. Then a sharper one. Only after all of that does it outline and write.

Stanford put two plain ideas underneath the whole thing. Diverse perspectives produce varied questions. And depth comes from iteration, the good question is almost never the first one, it is the third, built on the first two answers. Read those two sentences again and tell me they are about software.

The first idea is the homework

Before STORM asks a single question, it uses the selected topic to work out who to be when it asks questions. I teach and train cross-functional teams a version of this: the most useful preparation before meeting an account stakeholder is not collecting more facts about your product, it is generating a set of hypotheses about where this account’s problem actually sits, built from more than one point of view.

What does this clinic look like to the budget holder. To the ward nurse who meets all the patients. To the patient who waited eight months for a decision. Each lens hands you a different question worth asking, and a different opportunity worth testing during the meeting.

Skip that and you walk in with one point of view, your own. You ask what you came to ask. You hear what confirms it. You leave with the answer you already had. The limit on what that meeting could teach you was set before you entered the room, by how few people you had engaged when doing the homework.

The second idea is the meeting itself

STORM does not fire its questions in a list and leave. It asks, it listens, it lets the answer decide the next question. I have written before about this, talking about adaptive selling. The insight worth having in an account conversation almost never arrives in answer to the opening question. It arrives three questions later, once you have followed the thread instead of returning to your agenda.

This is where most discovery quietly dies. We prepare a good first question and no second one. The stakeholder gives us something real, we tick it off and move to the next item on the list, when the next meeting was sitting one “why” further down. STORM is an interesting thing to watch, because it is a machine built to prove that the success of the meeting rests on a well-crafted and thought-out follow-up question.

A note to you, before your next meeting

I recommend two changes, and neither needs a managers approval.

Before you go in, prepare perspectives, not a pitch. Write down three people who see this account differently, and the one question each of them would ask. Walk in carrying those instead of a pre-prepared sequence of questions designed to challenge the stakeholders perspective.

Then, during the meeting, earn the third question. When you get a real answer, do not move on. Ask why. Ask what it costs them. Ask what they have already tried. The perpetual access you keep looking for is waiting behind the question you did not think you had time to ask.

A research machine reminded me of that. It never pitched me once.