Labeling, Experts, and RLHF Challenges

Welcome to 2026, a year I coin “The Year of Enterprise AI.” As you’ll read about (and hear about) in our 2026 Imperatives launch, the coming year is all about AI moving from “assistants” to “agents” to “solutions.” And there are three big considerations to ponder.

First, the cost of AI is skyrocketing, so we’re going to have to focus on high-value use-cases and business-specific solutions. That’s not to say AI assistants and meeting summaries are not valuable, but once you start paying by the token you’re going to want to go deeper. As we discuss in our new Systemic HR AI Framework, we’re sitting on billions of dollars of real business opportunities now, and they go far beyond individual assistants. (We call these Superagents.) And the cost of AI will accelerate this focus.

Second, the data center buildout, energy costs, and political issues with data centers will matter. For corporate users this means understanding the underlying “costs” of AI usage (creating a single high powered image uses as much as 25% of the battery in your phone). I point this out to make you aware that these AI chatbots are not “free” – there are acres of computing campuses being built behind the scenes. And that means your “software providers” are turning into capital intensive companies. (And a new industry of data center companies may take over.)

(For those of you in the energy industry, it’s a wild time – almost as exciting as I’ve seen since my early days as an energy engineer during the OPEC Arab Oil Embargo in the late 1970s.)

Third is the fast-changing issue of AI’s accuracy, trust, and voracious appetite for data. As I discuss, the real opportunity for corporate AI is to take this problem head-on, and focus on your company’s data quality, governance, human feedback, and data labeling. The big AI labs are struggling to reduce the “Jaggedness” of AI (it’s strange ability to be really good at some things and totally dumb about others), and that encourages us to focus on narrow, domain-specific AI applications. And we all need to learn about RLHF (reinforcement learning with human feedback).

Our experience with Galileo proves that an AI solution that focuses on a vertical domain can be infinitely more reliable and intelligent than a general purpose AI. But don’t let me argue with Sam Altman, you’ll have to figure this out yourself :-).

We are launching our 2026 Imperatives research the third week of January, and there will be a special release of Galileo to accompany all the study. Our goal is not to give you a bunch of pithy predictions, but rather to give you a dozen hard-hitting “Must Do’s” for the year ahead.

I look forward to talking with many of your this coming year as we travel around the world, and I hope this podcast gives you some useful insights for an exciting year ahead!

Like this podcast? Rate us on Spotify or Apple or YouTube.

Additional Information

Empire of AI: Dreams and Nightmares in Sam Altman’s OpenAI (NYT bestseller)

From Assistants to Agents to SuperAgents: Where AI Is Going Next (podcast)

Galileo: The World’s Expert AI for HR and Leaders