A Successful AI Application
Doing customer research. Reminiscent of my early Web business.
Interviewed by Joe Lonsdale, Aaron Cannon says,
it's a hard thing selling to the giant enterprises. Like there's basically there's both like headwinds and tailwinds into approaching them. So I think there's a like right now obviously the technology is moving extremely fast. Humans are moving reasonably fast right there is a huge amount of social pressure and people want to keep their jobs and there's a lot of that and then enterprises move slowly and so you're dealing with all three speeds at the same time.
Cannon’s idea was to use AI to do customer research. That is, if a business is offering a product or service, it wants to know what customers are happy about and unhappy about. Surveys are crude and annoying. The most effective way to do research is to actually interview customers, but that is very labor intensive. AI solves the labor intensity problem.
But selling anything to big enterprises is hard. You are asked to meet with a dozen people in the organization, spending hours and hours trying to explain and persuade. Half the people you meet with are skeptics who are going to argue that your service won’t work. The other half are going to argue internally that the service is a good idea, but the enterprise should build it itself instead of buying it from you.
Internal politics always favors building rather than buying from you, because building enables some middle managers to expand their empires, while buying threatens someone with losing an empire. And even if the enterprise decides to buy rather than build, it will still think that buying from a small startup is too risky. I remember at Freddie Mac when we wanted to build a system to use credit scores automatically in our loan underwriting process, there was a small startup that already was doing that. After endless meetings with the small startup, Freddie Mac gave the contract to a big IT consulting firm instead, even though that firm had to figure out from scratch how to do the work.
Later in the interview, Cannon says,
we started we were it was entirely educational, right? Every call was like what is this thing, right? And I was educating what’s possible.
There were there were calls I joined where they did not know what LLMs were, right?
…the beginning was tough because it we were in category creation mode, right? Where we have to will this thing into existence.
It was like this for me in 1994, when I started my web site, which was for home buyers, and I wanted big companies to be sponsors. There were no banner ads or popups on the Web in those days—and our site never adopted them. My idea of sponsorship was that, in the context of content, we would provide links to contact forms for service providers.
In 1994, I was so early that only universities, some government agencies, and tech companies knew that the Internet existed. The Web was all but inaccessible to ordinary households, and I even had another Internet entrepreneur advise me to not bother with the Web for that reason.
I was trying to sell to non-tech firms that had zero experience with the Web. The first sales call I made was to a large mortgage lender. They wrote up a contract that began something like, “whereas there is a service, called the Internet, owned by Arnold Kling…” That wording was very flattering, but I could not let it stand.
For the first year and a half, I was dealing with prospective customers that did not yet know what the Web was. And yet it occurred to me, using exponential growth estimates, that within a year or two, they would all have web sites of their own. That was a strange realization.
I eventually acquired a business partner who noticed that a lot of the value we were providing to customers was educating them about the Web. He had the insight that, as he put it, “We should charge tuition!” Once he started doing that, we became profitable.
Cannon goes on,
I think in the last 12 to 18 months the market tipped where it went from this like you know a default skepticism and default kind of need of education to oh this is the the killer application for insights and research with AI period.
I experienced a similar switch, from ignorance about the Internet being the biggest obstacle to sales in 1995 to companies being so desperate to catch up on the Web in 1997 that my partner could charge almost whatever he wanted for “tuition.”
Between 1994 and 2000, the capabilities of the Web grew enormously. Almost none of the user interfaces that we take for granted today were even possible in 1994. I suspect that there are business opportunities today in building interfaces that expose the potential of AI to more people. If AI “agents,” for example, are just an esoteric tool for advanced software engineers, that is as adverse for diffusion of AI as the lack of a TCP-IP stack and the limits of 14.4 modems were for the Web in 1994.
substacks referenced above: @
I had experience with two prototype web sites around 1997 on my 486DX PC running Linux. The first would have really benefited from AI. A retired guy had an idea for a complaint website; memory says it was a four step process: choose company, product, and model, then write up your complaint. Monetization would be from selling messaging access to the manufacturers, a semi-extortionate process, but still honest, I think. I did the backend and database, a friend handled the graphics and layout. We made it really really clear to him, or at least tried to, that we were only interested in the prototype proof of concept that he could shop around to "real" developers, that we knew nothing about scaling up for production. He claimed he had a million bucks in backers. Turned out to be a lie; he was raiding his retirement fund, and once his wife found out, she put a stop to it. He then decided he wanted to change the entire model from that four step process, that we should do the work free as a bug fix. It was working and running on my 486DX, and after a couple of months of no further accesses, I pulled the plug.
The second one was much more straightforward. An experienced businessman wanted a web site for coupons to replace those newspaper clippings I still get in the mail. The idea was that office mates are going to lunch; where to go? Visit the coupon site! See who has specials, print the coupons, and they include the kind of detailed info which you can't get from mass market newspaper coupons: weather, time of day, location (zip code? I forget). He knew exactly what he wanted, he had the graphics and layout, and I wrote the prototype backend, again running on my 486DX. He actually shopped it around for six months or so, and I personally thought he had a good project. But in spite of his successful business background (he owned a $2.7 million house at Tahoe and had sold his previous startup), the Internet was too new and he couldn't interest anyone in turning it into a real public website. I don't think AI would have helped at all, since the prototype was simple and working, he just couldn't sell it in those early days.
Building rather than buying isn’t just about empire building. Third party systems are often built around data models and interfaces that aren’t compatible with existing internal systems. This forces the enterprise to either migrate their existing systems to the new model or build a complex integration layer. Almost all enterprise software has to be customized, sometimes heavily. If the Enterprise already has software engineers, it’s often easier and faster to build than buy. Buying has long term costs that are never reflected in the initial proposal. These systems often become unsupported, legacy stacks and a major source of technical debt.