Christopher Parsons interviews a Snøhetta practitioner (Speaker B) about using synthesis AI search in the firm’s knowledge management. Speaker B reports that AI search quickly surfaced project facts hidden in uploaded unstructured documents, which changed their plan from immediately backfilling structured fields to a combined workflow: upload more unstructured documents, let AI surface data points, then have qualified staff vet and populate structured fields. Benefits include parallel progress (others find data before full backfill), time savings, and more reliable answers when data is present in structured fields; risks include inconsistent values across unstructured sources (e.g., five different recorded square footages). The clip centers on using unstructured data as a bridge to higher-quality structured data and on the operational trade-offs of that approach.
Chapters
You've been in the synthesis AI search beta since the very beginning. Tell me about what, what you're doing with it. Like, how is it, what are you, what are you excited about the capabilities for it to do for Snohetta? And how, how is that impacting the way you're managing knowledge to take advantage what AI can do?
It's a lot of interesting things. Before we started experimenting with it, I was very focused on the idea that I was going to be filling in all of these data blanks right away. You know, as soon as I got the. The basic framework of synthesis set up, that my very next priority was going to be going around and filling in all the project data blanks that we were talking about earlier, that it was going to be a very focused process of truly filling in the blanks. And then we started using the AI search and I was starting to find little bits of information hidden away in corners that I didn't know was there. And as soon, you know, it was like documents that I didn't expect to have, any project information had been uploaded and they revealed little data points. And the AI search started to bring those to my attention almost immediately. And so then I realized that there was an opportunity to use that as a tool to make it a less onerous process for me and others around the studio. If we upload more of the documents that already contain those data points and then use the AI search to get at them and then fill in the blanks, it really can be a time and effort saver to save someone from having to answer the question they already answered 50 times before. Right. So it's interesting for that. It also means that people, other people than me, are already stumbling on that information. Even though my project of backfilling the project data isn't done, people are still finding that project data even without me doing it, which is interesting and unexpected. So it means that we can be now progressing in parallel more, rather than them waiting for me to finish something for them to experience it.
I'm inferring what you're saying, something in what you're saying, but I don't want to make sure. So I'm kind of imagining you're talking about things like town halls or design crits that talk about projects, or maybe a past proposal or something like that that talks about a project that's what we would call unstructured knowledge. You're going to go behind and kind of populate a structured database that's a little more tight and tidy and accurate and audited. So that's kind of. The unstructured data is kind of a bridge until you have better structured data to replace it. Is that my understanding?
Right? That's correct. And some people have, of course, wondered, is it even necessary to go fill in the structured data if you have all that unstructured data? That's already doing that. And the answer is yes, because we've found this out through testing. The AI is going to more reliably pull the right information from the structured data fields. That gives it all the context and that it needs to understand what that data point is. And it will more reliably serve up the right answer if the data exists in the structured field. But yes, I can already get at some of it through the unstructured data. And as long as I or somebody else who's qualified to sort of vet that data and make sure that it is the correct data gets into the structured field, then we can use it that way. If I left it completely unstructured, every firm, I'm sure, has this, where the square footage of a building has been written down five different ways on five different documents. And so if I left it unstructured, you would get a different answer each time you queried it.
We call it pulling the slot, the AI slot machine lever. And, like, we'll just see what answer we get. Yeah. So five stars for your answer. Unstructured data. I am aligned very much in everything you had to say.
