What’s hard about AI? Operations!

What’s hard about #AI? It’s not creating the models or doing the data science—it’s actually making it part of an operational process.

You don’t have to take my word for it, here’s a quote from Gartner AI Expert Erick Brethenoux talking on a recent (and free!) web seminar “Beyond the Hype: Enterprise Impact of ChatGPT and Generative AI

“As an AI expert myself, I didn’t have enough appreciation for my colleagues that are downstream, doing all the hard work and getting the AI systems to be operational.

Operationalizing AI has been the biggest issue since I’ve been at Gartner—actually trying to get the system to deliver business value for the enterprise…we can design upstream, we can get all the content we want, we can get all the models, but at the end of the day, it’s going to have to be part of a system somewhere.

So we see a shift—not a disappearance—of the skills, that are going from AI experts in general a lot more into the software engineering side of things, to get those systems to have a direct and practical impact on the value for the enterprise”

And that’s exactly what I’ve seen with our most innovative customers. Data scientists typically extract data from operational systems and move it to a hyperscalar data lake, then use open source algorithms to create and test their mdoels. The problem is that the vast majority of those models never go into production, for two reasons:

First, a lot of the business context of the data is lost during the extraction process (special formats used in the production system, hierarchies, etc.), which can end up in models that “work” on the test data, but not in the operational system.

Second, there has to be an end-to-end process that takes into the realities of operational systems: it has to respect security protocols, compliance frameworks such as GDPR; it has to be maintainable over time (how and when do the models get refreshed, etc); and it has to scale to the needs of the operational activity.

Fixing these means that the data science teams have to work closely with the core applications teams from the beginning, with the business goals in mind, rather than treating the operational systems simply as a “dumb source of data”.


Posted

in

by

Tags: