A few thoughts as I wind down an AI governance group

It's not that we've decided not to govern AI, it's that we spent a solid six or eight months asking people to please run their requests for AI-powered applications or features in applications we already had in the portfolio past us, and it's time to treat "AI" (by which we mean things backed by an LLM) as an everyday consideration for our security, compliance, legal, and IT teams.

It'll be a small relief. Small, pre-IPO businesses are not fun places to have a review committee. Nobody is  really there temperamentally. In this case, I kind of liked this group because nobody was a process czar or policy grind, so we settled into a pragmatic groove and got relatively quickly to "let's work this into the normal review processes we all have going anyhow." So if you do customer trust and you're trying to do business with us, blame me and that governance group for the dozen questions we added to our procurement questionnaire.

Anyhow, this is not to talk about being on a governance committee, but rather what I've walked away from the experience thinking. It crystallized after a briefing I took from an enterprise search vendor that is coming at the matter of how to integrate LLMs into their product a little backward of the Anthropics and OpenAIs. Their reasoning is that they're already good at insinuating themselves into all our data sources, so their "last mile" problem isn't getting that stuff into the corpus, it's getting the LLM to provide useful structure without turning everybody into a prompt engineer.

They had a compelling demo compared to one I took from a business that prefers to come at things from a "keep the chat interface, figure out the ingestion later" approach. I don't know what they're doing under the hood, but the output was consistently structured and formatted reporting that did a good job of attributing and linking to the context it was drawing from. With chat-first interfaces I've found myself chasing the output and dealing with the whole "just do that again except change this to that" thing that makes me feel like I'm playing an excruciatingly user hostile Infocom game. This tool has worked out how to guide that interaction out of sight of the end user, so you end up with a pretty good reporting tool vs. a super irritating idiot oracle. I continue to think the "chat" part of these tools is an awful distraction. Like, it's a cool demo for the credulous, but it's not where the power is. 

So, if you're keeping up the Mike AI Skepticism score card, you can score me a qualified "no, there's value there." To put it more subjectively and socially, there's a certain kind of "it's all snake oil" skeptic out there responding to the extravagant claims vs. the every day reality, and I tend to stop reading them after the first paragraph even as I grant them the usual point, which is that management desperately wants to believe the extravagant claims because it thinks effectively managing humans is a disgusting task for little people.

On the other hand, it isn't lost on me that a few vendors are incredibly reticent to extend trials past a few weeks ... maybe a month at most. They're completely allergic to "well, we're trying to find the value for our business and this is very new, so we'd like a little more time than two weeks with 10 hand-picked users guided into use cases you're already good with" requests, because, one of them said in a moment of candor, it's just too expensive for them to run a substantial trial or POC if you're going to walk away. But the commitment is eye-watering. If I went all in with licenses for every user for the big two chat tools, I'd exceed my spend on Salesforce, triple what I pay annually for Slack and similar for Zoom, and triple what I pay for GSuite, all of which are integrating LLMs into their products at the far more useful end of the spectrum. 

So I am paying a lot more attention to the people skeptical of the economics. They do not tend to lose me after the first paragraph.