Enterprise AGI is changing the challenge from capability to control

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 1:08

   You get to see it firsthand. Tell me a little bit of the insights and trends that are unique to now, 2026, that weren't true a few years ago. Absolutely. Yeah. So like, well, for the context, for us who don't know, we at Stack AI, we have a platform that enables enterprises to build custom AGI harnesses for automation of the internal back office processes. We work with the largest law firm in the country, the largest head firm in the country, with large companies like BAE Systems for retirement.

2. 1:33

   We help them come and take AI agents and tailor them for the specific office functions that they care about, whether it's commerce, whether it's compliance, legal, and do so in heavily regulated environments, whether it's on-premise on AirGap networks or doing so in IBM mainframe machines. We bring AI agents to these environments. Some of the things that we have seen, of course, progress in the past year has been a few things.

3. 1:55

   I think that February 2026 was kind of like a crossing point at which, at least from our own internal BIbe benchmarks, AGI was achieved for the purpose of enterprise automation. Really like the five things that make an AI agent relevant for the enterprise, which is understanding unstructured data, querying unstructured data, being able to orchestrate tools and actions in a non-deterministic way, the ability to interact with visual interfaces and the ability to synthesize unstructured information, you know, documents or tools, what not, presentations or not.

4. 2:26

   Those five things became really good to a level of reliability that matches human quality by the beginning of this year. And you could see, I noticed a bit of an inflection point as to where our customers are going to take problems and as to where the challenge is coming away from. I would remember like perhaps two years ago, the biggest challenge was like prompting in the end. Last year, a lot of it was like the API and NCP server design.

5. 2:50

   Now, a lot of it is kind of control around governance and how they're like, now they have these systems that work very reliably towards the task and are quite comprehensive towards it. How do you control what they do? How do you govern what they do? How do you ensure that only the right people get access to it? So we have to like make a big effort towards designing a platform that could ensure that and build an FDE team that had that first in mind.

6. 3:11

   Interesting. So AGI was a fully achieved in February, 2026. Enterprise AGI. Enterprise AGI. What metric do you have to make that claim? Vibes. But the truth is that we internally, we have a set of tasks that we care about that we say, look, and it ties back to these five skills I was talking to, where we say, like, if an LLM can come and receive a task oriented towards this and achieve it reliably enough with at least the level of quality of a human, given the correct problem and the correct set of tools, it is, for all intents and purposes of that task, AGI.

7. 3:44

   So, when we pick that as a benchmark, we say, how can we achieve this matter? All the benchmarks out there, whether it's a computer bench or enterprise bench or whatnot, all of them do kind of like point towards the improvement of this to a level that exceeds human reliability. I think that when you look at it, it's like a computer bench or enterprise bench or whatnot. Beyond the enterprise AGI, I say personal AGI or video AGI, we're not, okay, there's still more room to grow, but for this purpose, it's there.

8. 4:07

   And that has achieved a few major changes in the way companies think about it. Companies are not as stressed about ensuring a deep partnership with the big frontier labs or with large model providers. I think that there's been an explosion of open weight earlier this year that I don't think we saw coming, and we've become a lot more prevalent. especially in companies that really care about security and that really care about owning and governance and residency.

9. 4:32

   And we have to adapt to that and make it easily accessible and easily adaptable for them. Whereas, of course, people want to continue using the latest cloud, the LGBT model, but they'll have less of a pressure to do so and they'll feel less of a lock-in to each of these individual providers to achieve that.