What Enterprises Actually Ask About Private AI

Captions as YouTube publishes them, usually machine-generated, so product and people names can be misheard.

1. 0:09

   I think it'd be awesome to just start with brief context on each of your backgrounds. So maybe SUJ can kick it off. We'll go down the line. Everybody can get to know you and then we'll jump in. Sure. Uh can everyone hear me? All right. So uh my name is Sujay Jaladi. I am the VP of security at Falcon X. I've been in um blockchain crypto industry since 2017 professionally was uh the

2. 0:33

   head of security for Ripple um for a few years. Then I joined a startup which got acquired by Bitco um and I've been with Faknix for the last uh four years. Very excited to be here and uh to be um part of this panel. Thanks for having me.

3. 0:50

   My name is Kelly Combmes. I'm a managing director at KPMG. Um, I helped found our trusted AI offering about eight years ago around how to operationalize AI governance and I work with a lot of Fortune 100 companies on what does it really mean to implement AI governance at scale now increasingly so thinking about generative AI and I focus primarily within the tech, fintech and financial services industry. So excited to bring some practical application to you all today.

4. 1:22

   My name is Lucas. I'm co-founder and CEO of Tatio. My background is a cryptography researcher so from academia but since 4 years I'm building tio and we're building encrypted compute infrastructure. So basically the idea of having encryption also during the computation not only in transit and in rest but yeah like the not the operator the vendor or any counterparty can see the data whilst it is processed.

5. 1:51

   Super cool. So we've got an awesome setup in terms of varying points of view. SUJ as a practitioner inside an operating business, Kelly advising Fortune 100 companies and Lucas actually building the core technology that hopefully unlocks some of these more advanced use cases in the enterprise context. So with that said, Kelly, I think it'd be helpful just to start by kind of bridging the gap from some of the frontier conversations that we've

6. 2:15

   been having today in terms of, you know, model performance, the edge of AI, agentic commerce, um, and kind of ground our discussion in context of where your clients and enterprise actually is in the adoption curve of generative AI.

7. 2:29

   Yeah, I was laughing to myself a little bit earlier on the conversation about which model is the best performing model and reflecting on a lot of the organizations I work with and the question is not really what's the best performing model. It's how do we get started and build AI muscle memory to build the foundation to become competitive in the future. And so I think there's a you know there's the theory uh and there's the cutting edge

8. 2:54

   on what the capabilities are and then there's the practical application of that and there's definitely a big gap. Uh one of the other things we see is you know even just to go through a client's risk assessment process to implement a new AI system or model can currently take 8 to 10 months. And so at that point you know do you even have the best-in-class model that you're trying to get approved for the enterprise. That

9. 3:16

   being said, we're seeing a lot of progression um you know on some of these concepts and it's really exciting to hear that some of the biggest concerns our clients have around data integrity, intellectual property and specifically security that there's a lot of new techniques and capabilities out there uh that will change their risk appetite to move a little bit faster with AI adoption.

10. 3:38

    Yeah, it's super helpful and I wonder how does that compare and contrast? Obviously, this is Fortune00 largely public companies that Kelly's interacting with. SUJ, you're kind of on the, you know, 10 plus billion dollar private company side on the edge of the security apparatus. Um, how does this compare and contrast towards the risk appetite and the pace that you're seeing adoption inside of Falcon X?

11. 4:01

    It's not very different, especially when you're talking about the regulated uh industry. It's [clears throat] uh it doesn't matter where you sit, you're still within those lines, right? Um it's amazing to see how quickly individuals are uh trying to adopt uh AI and working through it. But when we try to bring that into the enterprise, how tough it is because you know you have no clear guidance. There is no proper way of assessing the risk and you want to do something that is private. You can only

12. 4:30

    rely on the promises that they make through the terms and conditions and the legal uh contracts that they put together. So it's it's still a fight. it's still something that is extremely um that that's that the enterprises are struggling with extremely to bring them in. So it continues to be that there are certain things that that were done but again it goes back to how much data and

13. 4:53

    how far is the company willing to actually share becomes the key. Most of the data that can be shared and that can be used to process is still the data that is not proprietary right you still want to uh manage and maintain your uh crown jewel so there is still a lot of struggle irrespective of whether you're talking about the fortune uh 100 or whether you're talking about uh you know

14. 5:16

    tens of billions or midsize smaller companies if you're in the regulated space you still continue to see the same friction in adopting uh AI yeah I mean it's consistent with my experience both prior to running sovereign as well as now in the conversations we're having and I think some of these kind of impediments are technical in nature in terms of the building blocks that um that are being kind of delivered on the cryptography side and so Lucas how do you see kind of

15. 5:41

    the pace of innovation and adoption that Kelly and SUJ are framing out being impacted by what you're building at too yes so we started like four years ago very broad so sector agnostic and we were talking to medical universities to large industry players. But eventually it all centered around blockchain again because the risk appetite for using these emerging technologies was not that great in like the enterprise use cases and blockchain for better or worse like

16. 6:11

    they have a very high pain tolerance and like they want to use the newest paper that is uploaded on some scientific server and just use it no matter what if it's peer-reviewed or not. And this is like where all this uh privacy enhancing technologies really got developed further uh because yeah you could just do it in practice and there was money running over it and this was battle

17. 6:36

    testing it uh live now that we're now seeing that AI plays a bigger role and it in enterprises and the guarantees that we have from let's say the cloud era database access controls um policies and so on they seem to not really fit AI use case like you train probably this model and where is the data now who is responsible for it what is like the decision that is made who is accountable for it they look for new tools in my opinion that can handle this yeah no that makes complete sense I

18. 7:09

    think it plays on as well some of the points that h was making earlier today in terms of like the actual adoption curve and incremental progress than all at once um beyond the technical components in terms of you know privacy preserving cryptography that gives you disclosure mechanis mechisms like what are the other kind of impediments from a risk compliance governance perspective that you see that contribute to these long adoption deployment cycles to

19. 7:34

    production maybe you can start yeah most of the large enterprises I'm at there's a couple issues at play um one is who's accountable for AI given the cross functional nature of AI we're seeing the rise of the chief AI officer being a role at some AI first organizations but many organizations ations are still parsing that responsibility across the syso, the CIO, co and so just even from an accountability um and who owns this space is one of the challenges that we're seeing. Um, I think the second

20. 8:06

    piece is how do we move away from organizational constructs, committee-based decisions on AI within some of these enterprises to more of uh digital governance and and quantitative decision-m using data in the process of both how we test our AI systems before deployment and then on the back end, do we have the right data to truly do robust AI performance and operations uh monitoring? And I think when we can get over the hurdle of um you know how do we

21. 8:36

    not just talk about that we have risk management practices but how do we evidence it at scale using KPIs and KIS is really going to be a differentiator for large enterprises like a question like for the privacy space like it emerged and like you had this data protection officer and I was very much involved in that space and this was basically this hot potato nobody wanted like the the it said like

22. 9:02

    that's That's legal and legal said like that's too technical for us like we don't want this with the AI. Is it similar or does everybody want to have the role of the chief AI officer to have also the resources like do you see people running away from it or fighting for it? I I think it's more fighting for it because you know we're a couple things right we're bringing data science closer to business stakeholders and there's an implication in that and that touches many risk stripes from having

23. 9:27

    humans in the loop in the function to legal wanting to understand contractual implications to third party having to understand the vendor relationship. So it's both um you know motivated by I'm involved in AI but I also um want to ensure we have the right mitigation in place. Um, but I think the rise of the chief AI officer uh and and hopefully drawing the lines around those responsibilities will help. I also think that we should start expecting that anywhere where technology and humans

24. 9:58

    interact, AI is going to be there. And that's that's a fundamental shift in enterprises which is how do we think about you know AI features on a platform to no I need to change my mindset that AI is one of those foundational technologies integral to my architecture and how do I think about the governance model that needs to wrap around that in the future.

25. 10:20

    Yeah, you made a point earlier Suj in terms of throw to choke and you're playing on GDPR with the data privacy officer which it felt like in kind of the mid210s like the same incarnation of what you're describing there as you're interfacing with regulators from the Falcon X perspective like what are the top things in terms of their mind is it privacy and confidentiality is it you know post- deployment testing and monitoring what are the top of- mind

26. 10:45

    considerations from those discussions so every regulator that I have contacted or uh that I have been in touch with has asked questions about AI specifically, right? They're asking the the the questions that we get are a full breath. They they actually don't know what exactly they should be knowing. So, they keep asking everything that they can that they find in any checklist that they can come up with. Um I think it comes from a good place but it's just that there is no

27. 11:10

    clarity for anyone either the regulators or for the enterprises in terms of how deep they can go. But the fears of losing the crown jewels or the proprietary data or something bad happening exists. So [snorts] when we talk about cryptographic uh attestation of you know confidentiality or privacy that continues to start making more sense for both parties. Until then there will be a lot of question and answers.

28. 11:36

    There will be a lot of ambiguity but it's all going to be you know unclear for both parties. Right. Um we've had international uh regulators ask specific questions about do you use AI? What do you do with that? Like how far do you process the data? What kind of data do you process? So every regulator is asking these questions both within the US as well as international um but there

29. 12:00

    is still clarity that needs to be uh brought in to understand what exactly um are they asking for right uh or what they want to do with that data that they get. Yeah. For for anyone who's grown up in crypto, the regulatory factf finding journey um and dialogues and learning process is extremely familiar, but you're lucky in some regards for those in the AI side that we spent the past six years articulating cryptography and how that operates and now it's being

30. 12:28

    applied in the context of data which hopefully helps facilitate the learning curve to some extent. Um, I think that's a good segue, Lucas, maybe to talk about too in terms of some of the building blocks. And you were describing your technology in a way that was aligned to some of the principles that I've been thinking about around, you know, how do we get to provable conditions around confidentiality and privacy, but also

31. 12:52

    allow for harmony with future regulatory expectations. And maybe you can just kind of draw out how Too is solving those problems. Sure. So, I think the core core issue that blockchains and especially cryptography tries to solve is to give us guarantees. So the internet was mostly championed by pioneers or companies that run on the motto like uh don't be evil but and then like it wasn't really good at that time

32. 13:24

    but now with AI I think it's completely broken. So we need something that can't be evil and that's why we rely on blockchain and cryptography to have like really properties that are closest to mathematical proofs as we know and with the really good thing that like even a powerful AI cannot break it. So we have this guarantees that the AI once we figured out how accountability looks like we can enforce it. When we want to see like how was the decision made we can verify that it is made under the assumptions we had in mind.

33. 13:58

    What we're doing at Tatsio specifically is working on two core technologies called SE knowledge proofs and multi-party computation. They were both invented in the 80s and for like 30 years mostly theoretical but through blockchain and the need of like computation that you do not trust your counterparty. There was a lot of funding going on in the last 10 years. And even like Dan Bonet from Stanford mentioned that like nobody really has seen that often a technology like serial knowledge

34. 14:30

    proofs go so fast from a theoretical concept that was developed for complexity theory like a very very theoretical concept to something that we can deploy on scale today. Yeah, I completely relate to that. It's kind of like this this story in terms of MPC and ZKP being accelerated by crypto now setting the building blocks and foundations technically that can be applied in the context of generative AI you know it's a perfect example of how

35. 14:57

    convergence is accelerating the velocity of adoption um so in the context of this like obviously every organization Falcon X and every enterprise has to embrace generative AI to some extent how are you seeing the risk appetite and risk decisioning model shift in terms of fear of being left behind in competition if you don't embrace the technology versus fear of regulatory or customer backlash if you

36. 15:21

    do and there's a cyber security or data event. Maybe Kelly you can kick it off and kind of start this track. Yeah. So for many large enterprises 2025 was really the year of PC's and just experimenting building the foundation um giving LLMs for employees for broad-based productivity and then going after functional use cases that solve you know some quick win ROI benefits.

37. 15:46

    Um, in 2026, we're looking at that paradigm shift, which is I'm expecting to get value out of this investment in AI capabilities, and I need to broaden my risk aperture in order to get that value and growth at the speed that I think that the market's moving. So, a lot of the financial services institutions are very concerned about what their peers are doing. Are they moving fast enough? Again, 10 months to

38. 16:09

    approve an AI system is not fast enough in the life of an LLM model version change. Um and so what we're doing is concepts like how do we get comfortable with patterning? So if we have an approved platform for a certain data classification for a certain division's use uh with certain features on it, are there cases that can be fasttracked or preapproved because they systematically

39. 16:32

    meet a set of pre-approved patterns and then allowing things that are deemed as moderate or higher risk to go to more of a um formal risk assessment process. And again I mentioned it before it's super basic but organizations are meeting in a committee-based structure and saying what's the risk associated with using this system go you know what are all the reasons that can go wrong and we're

40. 16:54

    saying no you need to use data you need to expand your AI security testing framework and use that data to drive is there really risk there and how do we mitigate that risk before deployment instead of making qualitative discussions in a committee and so it's again it's it's um thinking about industrializing AI uh you know versus saying we're going to do it [snorts] you know through consortium and forums.

41. 17:22

    Yeah, I mean that's a huge progression even of itself to move past these kind of enterprise committee based governance. I feel like the majority of folks in this room probably haven't had to deal with those painful technology operation risk committees or security councils. Um, but the impediment even just in basic things like scheduling those groups to convene to discuss these topics ends up being a huge contributor to that eight-month lag time. How have

42. 17:45

    you seen that applied to Jay in the context of what you guys are doing at Falcon X where Kelly's describing data classification, model classification frameworks and uh thinking about kind of your security review and approval process for new models, tooling that's being introduced, use cases, etc.

43. 18:01

    Yeah. So, um, going back to what Kelly was saying about 2025, right? Everybody wants to be in it, wants to try it, but they don't have an idea on how they can actually integrate it into their own ecosystem of, you know, the infrastructure and the data. Um, there was a lot of experimentation that has happened. 2026 we're continuing to see solid integration into the productivity

44. 18:23

    and the corporate side of the world where you know that is taking off really well but integrating it into the actual applications and using your own data. We're still going back to running those models locally locally and then trying to train your own uh models with with your data not letting the data go out.

45. 18:42

    Um, a lot of these are very sensitive, right? Some of these are secret sources that a lot of companies use and you really don't want that information to be out on a model that is trained that can go ahead and and you know train your competitor or start creating your competition uh companies that would pop up right so I there is there is a lot of effort that is going into it. FalconX

46. 19:04

    has uh uh over a year year and a half ago developed uh an AI product called Focal which is a financial analyst that was provided to all of uh our clients but the um the the lift that it did and the data processing that it did was all public data that was available. So we prepare the the um um financial analyst dossier for any single token or stock or you know whatever you want to trade and provide them with all of the information and then the analyst would take it and they would still run their own models

47. 19:33

    locally right. Um I that is there is a lot of push for these to be adopted more widely but those fears need to be you know we have to get over those fears and the only way that we will get over those fears is going to be with you know cryptographic at astation and and proof that you know what you're saying is what you are doing and what you are what you're being told that you are getting is what you're getting. It needs to be in proof rather than just uh someone

48. 20:03

    making a statement or putting it in the legal contracts. Yeah, that makes complete sense. I think like the old adage from an auditor perspective of like trust but verify and now we're introducing the same cryptographic verification that we expect in a zero knowledge context um to be applied in this world. It's interesting to see to some extent like we talked earlier today I think Arjun

49. 20:24

    made a comment that he'd put 100% of his assets in you know an open claw implementation in 12 months from now and I feel like what I'm hearing from you guys in terms of the progression towards enterprise adoption maybe that's consumerf facing for you know high net wealth or retail investors but um even in the context of what you're describing on Falcon X is very much public data oriented versus private portfolio level analytics risk modeling etc. Um so as we

50. 20:50

    look out in like the inertia and the competitive pressure that continues to build around these dynamics, do you see a change in the velocity and adoption curve and timeline to deployment from these you know eliminations of technology oversight committees or how do you see it changing over the next 12 to 24 months?

51. 21:08

    Yeah, I can imagine that like it really changes this time like around like December last year like a lot of extremely good coders felt like something has changed about the capabilities of AI. Mhm. And I've seen also with other like founders that even the non-technical ones are using co-work or whatever and are now really hyped on having FOMO and like this social aspect like what you're saying like we have to look what our peers are doing and so on. This can get

52. 21:38

    a dynamic where people might like FOMO is like an really the stronger force than the risk stuff. Uh so this could be one dynamic I can see in the next 12 months which really accelerates adoption. Cali I mean I think we're seeing processes start to move faster in the enterprise sense from months to weeks to to move on some of these PC's and ideas.

53. 22:05

    I think what enterprises are struggling with is I want AI fast, I want it cheap, and I want it at high quality and secure. And that doesn't exist in today's world right now, right? You can have AI fast, but it may be expensive.

54. 22:19

    And if you in order to secure it, again, the there there's some requirements around how to get there. Um, and we still see a lot of human in the loop. Uh, I think the most common term is, well, at the end of the day, we always have human in the loop. And it's like, well, how long is that the backs stop using AI? Agents are here. Uh, but agents for the enterprise are not widely adopted yet. So I think we're at the

55. 22:43

    shift of how do we build trust and confidence in the output the precision again tuning these uh models using an organization's data to make them more accurate while security protocols improve and then we're going to see those risk cycles move a lot faster in the future. But I think we're there. I think we're we're seeing the pressure of I feel like I'm not moving fast enough.

56. 23:04

    Um but again we're still talking probably weeks, not days, right? Yeah. Yeah. I I'll actually use the analogy of uh how you know cryptocurrency bitcoin got adopted by enterprises. Everyone wants wants in it. They will continue to dip their toes in it but they're not going to be making public statements or talking about how they are adopting and what they are doing with it because they are still trying to figure out how to build those guard rails, how to protect their data.

57. 23:30

    And until they get to that point, it's still going to be a very uh protected and um carved out implementation of what it is that they want to do. Very niche things that they that they know they can do really efficiently and effectively and they will be doing those until you get regulatory clarity like what we have seen come through in the recent uh year or two for Bitcoin and cryptocurrency

58. 23:54

    when you know everybody started publicly getting it on their balance sheet, right? So we'll we'll we'll continue to see the same uh uh approach on AI as well within enterprises. I I to I totally agree and Kelly you were drawing on kind of like these competing attributes or primitives and I would encourage you as well as the audience to check out ironclaw because I think as we advance that technology there will be competitors in that category that

59. 24:17

    embrace these principles unlocking kind of the abundant opportunities that generative AI presents and wrapping it inside of technical guardrails that protect data provide confidentiality and cryptographic guarantees around them.

60. 24:30

    Um, so I'm super super excited to see how this unfolds. Maybe 12 months we'll do a look back and have a conversation amongst the group. But we're out of time. So I wanted to say huge thanks to Lucas, Kelly, and SUJ for joining us and sharing your insights today. Appreciate it.

61. 24:45

    Thanks for having us. Of course.