Coffee With Digital Trailblazers
The Hidden Cost of Confident AI: Semantics, Guardrails, and the Context Layer
Isaac opened with a 9-11 remembrance and shared data points on AI failures and governance gaps. Felix Van de Maele, CEO of Collibra, explained the “hallucination tax” as the cost of AI agents getting things wrong and the need for context, feedback loops, and evaluation frameworks. The group discussed context engineering, semantic layers, ontologies, and the importance of governing both inputs and outputs. Liz, Derek, Joe, Martin, and Joanne contributed perspectives on governance, trust, accountability, and the challenge of moving AI from POC to production. The conversation covered AI debt, risk, policies, and the role of platforms like Calibra as an enterprise AI control plane. After the stream ended, the group debriefed, shared audience numbers, and discussed potential future topics like AI existential risk and the economics of AI.
[00:00:00] Speaker A: Greetings, everyone. Welcome to the 187th episode of the Coffee with Digital Trailblazers, our weekly LinkedIn Live event for digital data and AI leaders around transforming their organizations. Welcome to this very special session today on the hidden costs of confident AI, semantic guardrails and the context layer.
It’s going to be a fascinating conversation. My special guest today is Felix Vandermal. He’s the entrepreneur, founder and CEO of Collibra.
Very excited to be here today with the group and just waiting a few seconds for everybody to join this conversation where we’re really going to get deep and rich around how we really govern our AI and deliver transformation results with our data and AI capabilities.
Folks, I think most of you know today is September 11th. 25 years ago today, New York City and our country changed forever. It’s a day none of us who lived through it will ever forget. And it shaped how so many of us think about resilience, leadership, and what actually matters in life. I was in Manhattan for part of that morning and I wrote about my experience in my book, Digital Trailblazer. This morning I published that expert on my blog and if you’d like, you can go to drive starcio.com to read it. Our episode today is brought to you by Collibra, the enterprise AI control plane, governing context and control across any data, any model and any agent. Thank you for being with us here today and I want to get you started off with some research and just give me a second here while I move our banner and get to our panel.
Got some really good data points. Oops.
Got some really good data points from Calibra around this earlier today and hopefully things are moving slow to here today. Hopefully we can get that back up. There we go.
[00:02:14] Speaker B: Wow.
[00:02:16] Speaker A: Here we go.
Just some data points for all of you. Enterprise AI isn’t failing because the models aren’t good enough. It’s failing because the models don’t understand the business they’re operating in. This comes from Felix, CEO of Collibra, who’s here with us today.
Gartner predicts that 40% of agentic AI projects will be canceled by the end of next year. And analytics agents operating on raw ungoverned data answered correctly less than 21% of the time.
We all know data is really important around AI. There’s a real metric for you.
Calibra did a poll with Harris. 55%. Decision makers report they sometimes encounter an AI issue requiring them to personally push back or correct the output. I’m sure we have all seen this.
It’s pretty common and but here’s a really interesting thing. 90% are developing Agentic AI, 86% are confident it will return financial returns, 48% have governance to oversee it. So we definitely have a gap here that we need to address. By 2030, semantic leaders will be viewed as critical infrastructure.
2027 Gartner says organizations that practice prioritize semantics in AI data will increase their agentic AI accuracy by up to 80% and reduce cost by 60%.
I am sure Felix has something to say around that. And the semantic layer isn’t a new idea. What has changed is the cost of getting it wrong. I have five links for you here that go to the Calibra blog blog.
I told Felix earlier this week they have one of the best data blogs, AI blogs that I’ve seen.
Lots of just good knowledge, de jargoning things explaining how things work, connecting strategy with governance. And I just selected a bunch of these. Here is one of the themes that we’ll talk about today. How do we go from just talking about AI governance to having policies around governance to actually implementing governance? And I think that’s the real heart of what we’re talking about here and what you’re seeing at the bottom.
Collibra is doing a couple of sessions on context governance. I will be on the next one that’s airing on September 23rd and you can find out about these sessions@calibra.com context governance. That link will be up on the whiteboard in case you missed it and Want to welcome Felix to the floor. Felix and I have known each other for quite a few quite some time.
I’ve been following Calibra for a good number of years and been to their conferences a few times. Felix, welcome to the floor. I want to just talk about the hidden cost of confident AI. It’s really interesting topic because we need confidence to get value in the AI. And yet you’re suggesting there’s an underlying cost around it.
We used words like, you know, AI debt and AI data debt, which are terms that I tend to use. Others talk about slop, model drift and even context rot. But you have an interesting phrase and I want you to be able to explain it for everyone. You call it the hallucination tax. What is it and how much is it costing enterprises? Hello Felix, welcome to the program.
[00:05:56] Speaker C: Yeah, thank you Isaac. Thanks for having me. Fantastic to be here.
Great discussion on a very important topic. I think we all agree and to your point, there’s so much going on with AI, there’s so much terminology Being used everywhere.
We didn’t help the situation by having our own. But to your point, what we call that hallucination text is something that we think every organization, frankly every AI user has, has faced and it’s frankly facing today, right? And it’s only getting worse in some sense because the more independent and autonomous AI becomes. We started with ChatGPT and just kind of interaction to now this promise of autonomous agents. We gotta be able to trust what those agents do, right? Whatever decisions they make, whatever tasks they, they complete for us, ideally in an autonomous way.
They’re only valuable if we can trust that what they’re going to do, the outcomes that they’re going to drive, are going to be the ones that we want them to do. And the reality is today that’s often not the case. We often have to babysit our agents.
I think many of us feel this extra burden of how do we actually manage our agents, manage our models, manage our AI use cases. That actually requires quite a bit of, of time and efforts. And when the agents get it wrong, the cost is real, right? And this is what we call this hallucination tax. The cost of getting for the agents, getting the task, the decision, the outcome wrong. And if that’s the case, if we don’t trust what they’re doing, we’ll have to, like I said, babysitter. We have to continuously validate, review, double check. And in some ways agents have only made our lives harder because we have to kind of spend more time. And I think we all know the feeling like it’s just faster if I’m going to do it myself. And obviously that’s not what we want agents, that’s not how we want agents to behave. So I think it’s really important that as an organization, as we kind of deploy AI across the organization, we really deal with that hallucination index. I think a lot of organizations have woken up this year with the realization that look, AI isn’t free, right? It’s actually costing money. So ROI becomes incredibly important. Is it worth doing? What do we get from it? I have, dealing with hallucination tax is a, is a, it’s a huge, huge topic.
[00:08:28] Speaker A: You know, there’s a paradox here, Felix, because we want to trust AI on one hand and the other hand we want our employees in particular to be able to validate results, right? We don’t want to just say go ahead and turn the key and start the ignition and start driving somewhere in the case of a self driving car. So we want people to be involved in the recommendations that an AI is giving, we want to be able to trust it, and then the underlying cost is when it gets it wrong. So close the gap for me between participation and the collaboration between person and AI and reducing the underlying risks. That it’s not just observing and learning, it’s. I actually got to go back and tweak and do a lot of things to make sure that the AI is actually right.
[00:09:23] Speaker C: Absolutely. And I think a simple analogy, and I know analogies always fail at some point, but imagine your agents to be a newly hired employee in your organization, right? They might be really smart, have actually have a lot of experience.
They might have actually industry experience. And that’s the analogy with the models that actually have become really good. But if you don’t train that employee, if you don’t explain how the organization works, who does what, how you work, what processes we have, what we name things, how we get things done, the smartest employee is still going to be ineffective.
[00:09:56] Speaker D: Right.
[00:09:56] Speaker C: And so think of an agent that you haven’t trained, that you haven’t given that context too, is still going to be ineffective. They’re going to do things that you don’t want them to do that are not what you have intended to. And that requires a lot of time. And so just as we have to enable and train new employees, we have to train, enable new agents. And that requires time and effort and investments. And until you feel comfortable and confident that they have enough context, the understanding of the organization, how you work, what to do, what not to do, you’re going to have to verify.
[00:10:32] Speaker E: Right.
[00:10:32] Speaker C: You cannot take the risk for them to do something that you don’t know, because ultimately accountability still relies with, I would argue, the humans behind the agents, not with the agent themselves. And so this is really important.
How can we effectively train, educate, enable our agents to understand our organization, how while we do that, how do we make sure we stay in control, be verified? Because accountability still lies with the people behind it.
[00:10:59] Speaker A: Felix, a lot of questions I have around accountability that maybe we’ll talk about a little bit later.
But I want to leave this notion of hallucination tax without asking you your ideas about how to actually measure it. Like how does an organization either sense or directly measure that? You know what, we’re spending more time fixing the AI than getting results from it.
[00:11:29] Speaker C: Absolutely. And this is, we’ve learned, very important kind of running AI in production, deploying it is that feedback loop, right? That feedback loop, the evaluation of the agents, the use cases, what the models say, because they often get it wrong. How do we make sure we feed that back into the inputs that we give we give the agents and so we can, we can adjust. Typically we typically organizational really train or fine tune their models. Typically we use open weights or commercial models. But it’s really the feedback loop that helps us provide better context so that the agents appropriately acts, makes the right decisions, gives the right context. And so starting to measure that is really, really important. The best way we can measure that is to have an evaluation framework to say, okay, if we ask X, is it going to answer me why? Because that’s what we believe is the right answer.
And that allows you to also then start evaluating different models. As models come out that are stronger but also typically more expensive, is it worth me using the faster, better, more expensive model or is it okay if I use the existing model because my evaluations give me enough confidence that it’s going to act appropriately? Again, the feedback loop then as we learn and how do we train the agent to act appropriately? Feeding that back into the context and the context layer, maybe we can talk about that as well, becomes really important to what we call hill climb and keep making better and better and better over time.
[00:13:01] Speaker A: Thank you. Felix, I want to hit up with my speaker board. Liz has got her hand raised first. She heard governance and has to have something to say about that.
[00:13:13] Speaker D: You.
[00:13:14] Speaker F: Yes, governance is my trigger point.
So as I’ve said many times, governance is not a four letter word, it’s actually a seven letter word which is support.
So when I’m thinking about this problem of the hallucination tax, it seems that it’s all about losing context.
And in the sense that if the model itself is losing context and hasn’t been trained to have the appropriate context, is there a way that we could be actively managing, getting ahead of that to provide a consistent reminder of what the context is and in a sense create an agent to continually create context in the background and maybe even, you know, possibly reduce token usage.
So it’s more of a question than an answer because I’m not really sure how much of this is really handled by training and how much of it is, you know, getting ahead of it and managing, actively managing the context.
[00:14:26] Speaker A: Felix, you want to answer that? I mean
[00:14:30] Speaker C: it’s, it’s an amazing, fantastic question because it’s, it’s super, super important. Right. The best way to improve the results, the accuracy and to your point, also the cost, right. The token, the limited token exemption hopefully of our agents is to provide it the best possible context and we have to Treat it very seriously. We have to really engineer and build that context.
Actually this whole hype, if you want to call it that, of this four deployed engineer that now everybody has, what are forward engineers doing?
They’re building context. Ultimately they’re building context so that the agents, the models that everybody, everybody’s using the same model can operate effectively. And I would argue today we are brute forcing that context. We built a context up to a place so that the agents can solve our problem as effective.
But then that’s kind of step one. Right. What we’re starting to see very, very quickly is that we have to maintain that context. Right. That context decays over time. The organization changes, the model changes, things change. You have to continuously maintain that context. And so what I see the most mature organizations doing, they really think of that context as a living artifact that has to be managed and has to be governed independently manage for the first time is you have to do your point, build the context. And you said about terminology is a typical simple example, but it’s still a complex one. Like if we want to, we want to ask our agent, how many customers did we add over the last quarter? Well, as I’m sure lots of different ways to count how a customer, what makes a customer, if we don’t tell the agent that the number is going to come back with is going to be completely meaningless. Right. That’s going to be a hallucination tax that we’ll have to pay. Right. And so terminology is often step one. And this is why there’s a lot of conversation nowadays around ontology and semantics, because that’s really the core foundation. Because if the agent doesn’t agree on the terminology and the business context of your organization, there’s no way they’re going to be effective. So that’s the context, management and engineering components. You also mentioned the opportunity to use agents to actually build the context and improve the context continuously. But what I call at compile time was like pre inference. So that inference, we don’t have to do it at inference time because that becomes incredibly expensive. And I’ll give you a typical example. I’m sure organization have sharepoints, we have tens if not hundreds of thousands of documents in SharePoint. An agent can very well parse all those documents. If I ask a question, it might read all those documents and give me the answer. But imagine how expensive that question would be. That’s not the best way to get the most cost effective and accurate answer. A much more cost effective way is to actually use a different agent to parse and manage all of These documents in SharePoint, add structure, filter them out, understand which one are the most recent. What’s the quality? If I ask a question, the agent is not going to look at 100,000 documents. Going to look at 100 documents and obviously it’s going to be much more cost effective and much more accurate. Right. And this is an example of engineering the context to be much more effective for consumption.
[00:17:59] Speaker A: Thank Felix. Let’s keep going around the room here. Derek, you know Governance is a seven letter words with 70 dimensions, right?
[00:18:10] Speaker D: Absolutely, absolutely, yeah. I mean Felix, you got a lot of great comments and insights here. I greatly appreciate that. But we’ll go back to the hallucination tax. I mean I look at this as I look at this first of all, and you’re looking at the responses from a confidence, trust and an operational risk point of view. I see this as where you’re looking at the dangers of organizations now. They’re having confidence in what AI is telling them, but not realize the output is not something that they can actually go with. So the confidence actually creates this false sense of certainty that they’re trying to go with and make it wrong. The danger isn’t the AI getting the things wrong and humans get things wrong too. The danger is really getting things wrong with absolute confidence which creates the problem. And when I look at this, I think about, you know, we talk about tech debt that we had before. I now see this being as creating as AI debt. I look at the AI debt, data debt that’s now being created because everything that we AI generates has to be verified. And now we’re taking time and resources that we didn’t plan for because we didn’t know to. Because when we first rolled out these AI models and these, these tools, we thought they were going to solve some of the bigger problems that we have. But now they’re gener documents have been verified, the recommendations, analysis of decisions which now are creating future risk and somebody has to pay that. And I think when you’re looking at this from perspective of, you know, what are we really getting to, how much is costing? That exponential cost was still to be determined because people are still trying to figure it out as we go along. And I really see this as something that we need to kind of get our hands around sooner rather than later. But they got so many models that are being created with so many risks and issues. How do you get around now this wave of new technologies and services, all the information coming from like the anthropics and the, the other Companies that are now talking about some of their risk and some of the things that they’re finding with some of the things that are taking place associated with this solutionation tax.
[00:19:58] Speaker A: Felix, thoughts on that? I mean, just a lot of technology, a lot of models out there and a lot of risks materializing.
[00:20:05] Speaker C: Yeah, exactly. The fragmentation is increasing, the rate of innovation is increasing, the level of control and certainty we have on, on everything that’s being released, I think it’s getting lower and lower. So just the velocity is really hard to deal with.
So I think what we’re starting to see again as governing not just the context, but also the agents, the models and what they’re allowed to do. And a lot of organizations start to really think through the AI policies. Like what do we as an organization feel comfortable for agents to be able to do? Right. I just recently heard of an organization said, look, our agents are able to read data from systems, but we as a policy today do not allow them to write data back into systems. I’ve heard of agents that have done horrible things with HR applications like really, really impactful. So again, as an organization, you have to think through what is my policy, how much risk do I want to take, what do I feel comfortable with? And again, you have to do that in the context of your industry. Right. If you’re a healthcare organization, that probably looks different than if you’re a technology organization. So one is defining the policies. And then, and this is why we think a lot around the AI use case. It’s not really just about the model and the data. You have to look at it in the context of the use case. What, what is that model, that agent actually going to do? What data is it going to consume? What’s the intent? That becomes really important. An internal agent, I’m probably going to be able to accommodate more risk than an agent that is going to interact with my, my customer base.
So how do you, how do you think about that AI use case? How do you make sure you’re intentional about the policies that you as an organization want to, want to take on? And then really importantly, how do you enforce that in production?
[00:21:50] Speaker E: Right.
[00:21:50] Speaker C: When you deploy that agent, how do I actually make sure that that agent is going to act according to the policy that I have defined? And this is kind of very much still, I think work in progress. We’re all starting to do a lot of work. Obviously the AI labs are doing a lot of work around alignment. Like how do we kind of align the models to be appropriate. But again, as we deploy it for our particular use case, we have to almost align the use case to our internal policies. And this is where we are starting to work on the concept what we call guardian agents. Like how do you define a contract? Almost the agent contract like that specifies what the agent is allowed and not allowed to do and then how do you enforce that contract in production to make sure that you are the agents are acting in accordance to your organization’s policies.
[00:22:42] Speaker A: Felix, we’re all going to talk, become regulators it sounds like.
And I think that’s important. I mean, look, I did a paper or an article a few weeks ago because I’d been hearing about enough, just, you know, a lot of terminology, semantic layers, knowledge graphs, guardrails, context layers.
It almost gives this false illusion that organizations are going to just define their context layer and agents are going to have all the knowledge it needs to do things appropriately. And context has context.
It’s not just a universal one size information base that we’re giving all of our agent. I need to think about who’s accessing this information, what is the type of problem they’re solving for and what information is relevant to that. Can you go deeper into how you define the context layer and how you govern it?
[00:23:42] Speaker C: Yes. And it’s again, we ask people how do they find context? Probably 10 different answers. Right. Ultimately we think as context, as what do we provide the model so that it again answers our question or execute on our tasks in the most accurate, cost effective way.
What we believe ultimately is that context is going to be created in lots of different places. There’s not going to be one place where we create context. We have ChatGPT, we have Copilot, we have Cloud, we have custom harnesses, we have AI platforms, data platforms, systems of records. All of that contains context. Right? Context though is going to be consumed in lots of different places. Again, we have lots of different ways to interact with agents. We all make our own agents.
Organization is going to make agents that we’re going to reuse. Every business application comes down with agents. So there’s going to be tremendous fragmentation in context. So thinking about how to actually manage, maintain and drive a level of consistency and reusability of that context becomes really, really important. Like I said today, it feels like we’re just brute forcing.
Everybody hires a bunch of four deployed engineers to build the context, get the use case in production and then like mission successful and we can move on. Of course that’s not the reality. Right. And this is why it’s also way too expensive the way we’re Tackling the problem. Today, we need a much more systematized way to deal with that context, handle that context, govern that context. And we obviously believe that you need a platform to do so.
[00:25:19] Speaker A: Why?
[00:25:19] Speaker C: Because, again, you want to have a level of independence that gives you the ability and the flexibility to choose where and how you’re going to consume that context, right? What models you want to use, what harnesses you want to use, what platforms you want to use. Innovation just changing so rapidly. You don’t want to get kind of fixed or stuck with kind of one option. But also you’re going to have lots of different sources of that context and you want to have a level of independence that you can use any type of context wherever it’s created, whether it’s in what I call a system of knowledge, a sharepoint, a confluence, code, policy, documents, what I call a system of records, SAP, Salesforce workday, or a system of data, a data breaks, the data warehouse. All of that is going to contain context. And that context platform needs to be independent from kind of those different kind of both sources and consumption of context.
Now, within that context, again, there’s different layers, if you will. We touched upon, upon it briefly. There’s a lot of conversations around semantics, ontologies as a core way to kind of ground the agents and the models in your business vocabulary, your business meaning.
And that’s becoming an important part of the whole context layer, context platform as well.
[00:26:40] Speaker A: Thank you, Felix. I’m going to go to my CIOs next. Joe has been waiting patiently. Joe, you mentioned you wanted people to verify in the comment stream, so I want you to comment a little bit all around that, but also go into, you know, a lot of this language or semantics, guardrails, context layers, knowledge graphs, they’re all new concepts at the C level and at the board level. So how are you simplifying it for them?
[00:27:12] Speaker E: Well, take, take the second part first. You know, my belly is communication. And I think it’s essential that from the board on down, what we are doing in terms that people can understand have to be frequently and regularly communicated to the entire employee base. So the education aspect of this is of course something I would hang my hat on.
But Isaac, I want to go back to something Felix said early on.
I love the analogy of training a new employee. I think we can draw a very simplistic illustration of all the problems we’ve been discussing if we put it in the context of a new employee. Because the new employee has to be trained. No one would argue the point. Someone comes in they need to be told what they can access, what they cannot access.
We have forever had mechanisms to prevent people from accessing data they shouldn’t access. We certainly don’t let new HR employees update records that are sensitive, you know, until they’ve been trained properly. The second aspect that I think Felix brought up, which was spot on, was auditing that, you know, once you train and you check frequently as you’re training, it doesn’t stop there.
You continue to poke in and make sure the new trainee, the new employee is up to snow and not, you know, not veering off left or right of center.
But the part that we didn’t talk about is instilling in the new employee curiosity and doubt.
Because it’s important that we make sure a new employee understands if they encounter a situation and they really don’t know what to do, they should raise their hand and ask the question.
Be guided, right? And in the AI world, I think we need an analogy to that. We need to be able to instill in the agent that sense of I’m not really sure. So let me go back to a human and get some additional guidance. What do you think about that, Felix?
[00:29:27] Speaker C: It’s a great comment and I think it’s not an. I think, I agree it’s really important and I don’t think we’ve necessarily solved it already because we all know, I think models can often be very confidently wrong. Right. And that’s a big problem. So how do we almost train guide guardrail, the models, the agents, to, if they’re not sure to your point, raise their hand and ask a question.
I think the way we’re starting to solve some of that is through a good harness.
[00:30:01] Speaker D: Right?
[00:30:01] Speaker C: Because a harness in a way gives literally, it says in the word, the harness in which the agent can operate. And so if the agent being asked a question to do a task and the agent doesn’t feel like it knows, it’s confident enough, it doesn’t have the right context or data, or it’s not sure. How do you kind of build the harness in a way that they do raise their hand or ask the question or ask more context or ask more guidance.
But then also we’ve seen that agents have a tendency, this kind of how the model’s been trained to find a way. And so it almost goes against a bit of the model training that we’ve done because they are goal seeking. If you ask them to do something, they’re going to try to achieve their goal. And we’ve all heard the, the stories of how far they sometimes go to try to achieve their goal. So I think that’s a big problem to be solved. Like how do you, how do you find the right balance between. I’m not sure I’m going to raise my hand if you, if you do that too often, then I’ve going to have an hallucination tax. And, and you want a level of independence. If, if you’re on the other hand of the, of the equation is if they don’t do it enough, enough and, and they, they try to find a way and this is where they can sometimes act inappropriately. But it’s a really good, really good question and I think having a strong analogy and then also implementation to make that real is going to be very important.
[00:31:26] Speaker A: Folks, you have joined this week’s coffee with digital trailblazers on the hidden cost of confident AI.
We’re going through our de jargoning of semantic guardrails and the context layer and we’re going to have our closing session talk about implementing AI governance and getting a better understanding of how to go beyond just having policies in place. How do we develop our context layer and how do we get feedback so that we reduce our organization’s hallucination tax?
Folks, this week’s episode is brought to you by Collibra. Felix, CEO of Calibra, is here. Everyone knows AI needs context. The hard part is governing it. Your agents are pulling from documents, tickets and systems that nobody has owned in years. And when the context goes stale or wrong, the agent acts on it anyway. That’s the gap Calibra closes as the enterprise AI control plane. Collibra gives context, an owner, a system of record, an audit trail so you can see, trace and govern every agent you put into production. Felix, thank you for being here. I want to go a little bit deeper before I go to Martin. Joanne, talk to me about semantics a little bit.
Felix, you know, enterprises are complex and you mentioned they have a lot of data around here.
How should organizations be thinking about developing their semantic layers?
[00:33:01] Speaker C: Yes, and semantic layer is another word that’s not necessarily new, but also get to use there lots of different ways.
So I think semantically is really. It’s not a new world, it’s not a new concept. It started from the business intelligence world, frankly, 20, 30 years ago with business objects where you have to define how to calculate certain metrics, Right. If you say revenue, how do I calculate revenue? If I say gross margin, ebitda, like what’s the actual calculation of that metric of that dimension? And then we’ve kind of advanced in the new kind of modern cloud data platforms and BI platforms, analytics platforms. And we kind of separated the semantic layer. And now with agents, I think that becomes even more important because without that semantic layer, without having a very clear definition of how to calculate a certain metric, the answer that the agent is going to give is often going to be wrong, or at least not trustworthy.
[00:33:58] Speaker G: Right.
[00:33:58] Speaker C: And so that semantic layer comes really, really important. The question is how broad do you consider semantics? Do you look at it as a purely metrics layer that I think have a really important place, but that’s really constrained to kind of analytical metrics. Typically it just looks at a SQL query, right? What’s the query that I run on my database to calculate this particular number? This is how analytics platforms have built it. Now in the agentic world, those metrics layers are still important, but often they get broader because we don’t just want to answer SQL like questions, we also want to ask more why questions.
[00:34:40] Speaker G: Right.
[00:34:40] Speaker C: And this is why the semantics layer are expanding, I would call into the world of the ontology.
[00:34:46] Speaker G: Right.
[00:34:46] Speaker C: Which is also a form of semantics.
[00:34:48] Speaker B: Right.
[00:34:48] Speaker C: But it’s much broader around how does my organization work, what’s the vocabulary, what’s the map of my organization, if you will, beyond just a SQL like question, a metrics question. And again, that becomes incredibly important to govern the context because if we don’t agree on that terminology, the agent, there’s no way we can trust the agent with the answer.
And so that’s an important part of everything about AI governance. I think AI governance, I think there’s two sides of it. There’s governing the inputs. A lot of that is around context semantics and then it’s around governing the output. Right? Like what does the agent actually do and is allowed to do? And I think that’s the other side of AI governance that maybe we can talk about as well.
[00:35:37] Speaker A: No, I think I love this definition. How does our organization work? It goes above the data catalogs and data dictionaries that are the starting point.
And I think about, you know, diving into anybody’s ERP or CRM and trying to get an answer around what is a dollar or a date actually mean?
Or looking into their BI implementation, saying what are all these groupings and formulas going on? And if people don’t understand the organization semantics, you can’t expect the AI to read between the lines and know which data actually use in their analysis. Let’ to Martin, we’re trying to make the gap, the translation from using AI to avoiding debt and tax to creating the context layers that our organizations need to be effective in AI thought your thoughts and questions for Felix.
[00:36:36] Speaker G: Thanks for having me, Isaac, as always.
I just want to throw something in as well. Only kind of the whole thing around hallucination. And we talked a little bit about.
Yeah, kind of a new employee or something like that. I’m always going to take it further back and say, think about the industrial revolution. You had horses plowing fields which gave way to tractors, which gave way to GPS controlled tractors, which gave way to robots. And you think about each stage of those and even from the days of horses. Yeah. Felix mentioned harnesses. Yeah. You’ve got a horse in a harness, a physical harness, and you’re controlling what it does. And I’m just going to think paralleling the kind of industrial revolutions and how things changed with this whole AI piece. And it just kind of struck me as kind of quite interesting how that physical plowing of field really parallels some of the issues we’re having with AI. And so that was kind of the first thing. And then I, I was going to pick up on something else that Felix kind of. Yeah, we talk about guardrails and you just mentioned briefly in that Felix, I think there’s also a strong place there for guided rails. Yeah. As in we tell the agents and we tell people what they can’t do and we put blocks in place but we don’t spend as much time on the guide rails of kind of encouraging the actual approach. So maybe Felix, you want to kind of chat about that a little bit.
[00:38:07] Speaker C: Yeah, it’s a great thought. And actually I love the word guide rails, but it is actually something that we’ve learned to do and how important it is to get an agent to behave. And a lot of what we call I think skills and tools is really guide rails. Right. It’s a form of context. It’s really guiding the agent first to this, then do that. If that happens, don’t consider that it’s basically guiding the agent and how to solve a certain problem without being too prescriptive.
[00:38:40] Speaker D: Right.
[00:38:40] Speaker C: So it’s not a prescriptive like xyz, but it’s really the guidance of this is kind of direction where you have to look through. And I think it’s a great way of describing what these skills and the tools are that we use to literally guide agents to be successful. And we know it’s just text ultimately.
[00:39:00] Speaker E: Right.
[00:39:00] Speaker C: And we’ve probably all written these skills and tools. But it’s actually. I love the analogy or the contrast with the boat and the importance of both the guardrails, like these are two here and not further.
And the guide rails of. Okay, how do I best and most effectively guide the agents so that it comes back with a strong answer? So I love that. I love both of these terms.
[00:39:26] Speaker A: Thank you, Felix and Martin. Joanne, welcome to the floor. Can’t believe it’s 40 minutes in and we’re only getting to hear from you. So your wisdom and your question for Felix.
[00:39:38] Speaker B: Okay, well, my wisdom I kind of put down in your notes.
It’s a piece that I’ll release next week, but basically I’m looking at it from the point of view of governance and, and you know, we’ve built this very heavy semantic spine around our governance layer. But that being said, I, I’m hearing this more and more often.
I don’t want to be sued. That’s what CEOs are saying.
Help me figure out how I determine what these, what governance should be for my organization. And I’ve kind of put it into seven A’s, call it the agentic A list. You know, about agency. What are you telling AI to do for you? And really to view it in the frame of I am giving artificial intelligence, a proxy or a delegation for decision making in my organization. And the reason I start with something so provocative is because if you do it at the individual use case, what gets omitted? And this is why the question of I don’t want to get sued keeps coming up, is because a single process touches many, many parts of an organization and people are building their use cases around a single use case without necessarily taking into consideration the broader picture.
So in order to really overcome and kick it up to that higher level, whether it’s a board level or a C suite level of how do you put these things in place?
Look at agency, look at the authority. Who’s taking authority? Who’s owning the authority? Who’s taking accountability? What is activation really mean for agency? What does adaptation mean? In other words, you can have some very prescriptive rules that are corporate policy or regulatory, but you also have to make them adaptable. What is the lingua franca of my organization? How do we really operate and what comes out of governance? If you do this at the outset, or even if you’re already in POC hell and nothing is moving into production, is that governance also becomes something of a way to reshape the organization and how it works, but also its business processes, which all overlap. It can be in manufacturing, it can be in any other industry. You’re never Going to have a single process as an island.
So we started applying a framework to it which then gets supported in the tools. Whether it’s a semantic layer, whether it’s a knowledge graph, it doesn’t matter matter what those tools are. Stop thinking about tools, start thinking about overall what you’re trying to achieve and how you’re delegating the or giving a proxy to the AI because that’s really where it can come back and bite you. So the holistic perspective is the one that seems to prevent the I don’t want to get sued and the holistic perspective is also how you save point meaning in token cost in development time, etc. Etc. Felix, do you agree with this?
[00:42:55] Speaker C: Yes, it’s great framework. I couldn’t agree more. And we see it a lot. If you’re not very intentional about this framework and really thinking through this agency, authority, accountability, activation, adaptation. I love all these concepts. It often happens. It’s a very tactical point, but it happens where you’re starting to do a lot of work and you mentioned that you’re in POC hell. Nothing actually gets in production.
So this is often. Sometimes governance gets a bad rep because it feels like it slows things down and it’s bureaucratic. I think governance done well can actually be a force function to accelerate, to create clarity, to create predictability.
That creates a great framework and a great framework that gives very clear guardrails for everyone to know how to operate. So I think it’s the organization that are most serious governance really think of this beyond the technology. 100% agree that it’s not about technology, it’s about the framework and the intentionality about what are our AI policies, how do we want to operate? How do I create a really easy and understandable framework that people understand how to operate within actually drives so much more efficiency, so much more real AI use cases in production. And you avoid the, I think the fallacy that a lot of organizations are having today where you build a lot of really good prototypes and POCs, but you always get stuck on things not getting in production.
[00:44:27] Speaker A: You know, there’s some really good comments here, everyone on the comment stream and a lot of good questions that we’re not going to get to. There’s a few comments here around.
Around autonomy. I think autonomy has to come later. I love that Joanne, that you have this at the beginning. At the end.
I think the two that really focus on is around authority.
I always ask people whose whose decision is this? And in some cases it’s obvious around a budget decision, but a lot of other cases it’s not. And I think what we’re really trying to get to is that is making sure that our agents are accountable. And so, Felix, help us out with this and then we’ll go back to Joanne.
She had a second thought here, but I want to talk about this shift from just talking about governance and writing up our policies, educating our employees. That’s usually where we’re starting from when we talk about governance and when I talk about, you can’t start hitting the gas pedal on strategy and experimentation and getting things into production without policy and employee education.
But then what does the actual implementation look like when we start saying, okay, or how are we going to have an execution layer around governance so that we’re actually knowing we’re practicing what our policy states?
Break this down for us. How do we implement AI governance?
[00:46:01] Speaker C: Yes. And so I think one to our earlier discussion, not having a policy, I think is definitely a fallacy and something to avoid, but only having a policy in paper and not being able to effectively implement it as an equally big problem. And this is really why we are expanded what we do. And historically we’ve done for a long time on data governance for regulatory use cases around data which we believe are very applicable to also AI governance. It’s really starting with the policy and there’s great frameworks out there. And then how do we make that a reality?
And to us, it’s really about workflow and process and lifecycle. And we believe from an AI governance perspective, the core artifact is what we call an AI use case. As the combination of the model that I’m going to use, the data that I’m going to use in that model, and what do I want to actually solve for the use case? The intent plays a huge role in what risk does this AI use case bring? How does it apply to our policies? And then what we’re helping organizations with is how do I move this use case through a lifecycle so that all the different stakeholders we talked about, authority, for example, can make sure that when the use case goes into production, we have checked the box, or we are aligned to all the policies that we have. And so in practice, how that typically works, there’s often a prototyping discovery stage. I would love to develop this type of use case using our customer data to do some analysis. Great. Maybe we can do a psc, but you can start ahead. But then before it goes into production, there’s a lot of stakeholders that need to say, look, about authority, about accountability, about adaptation, all these frameworks that we discussed, the way that typically happens is through attestations and assessments.
There’s a lot of regulations out there. How do we make sure that we are complying to the Privacy regulation, the European AI act, the Data Act? There’s a lot of regulation out there which we’ve got to make sure that everything we’re doing is compliant to.
So pushing that to a lifecycle where the different stakeholders, whether it’s risk or legal, engineering, products, finance, what’s the cost? All these stakeholders have a say. You have to bring those together in a collaborative kind of platform to push that through a life cycle and a process. And again, that process, that life cycle might be different for every organization, depending on your policy framework.
And then you define, okay, we can put that kind of agent, that AI use case in production. That’s a really important step that we’re doing for a lot of organizations. Then the second step, which I think is starting to become a focus now, is then, okay, once we’ve defined, let’s say, the guardrails, the policies that are applicable to these agents, how do we now then enforce that in production?
[00:48:53] Speaker D: Right.
[00:48:53] Speaker C: We deploy the agent in some runtime. How do we make sure that the agent actually behaves in the way that we’ve defined that it can behave? And this is then, this is the policy enforcement, the runtime enforcement which we’re, which we’re developing right now.
[00:49:10] Speaker A: Very, very interesting around this. I love how you’ve taken some complex concepts around, you know, data, model, intent and just capturing that as an AI use case. And then talking about governance not being a policy, it’s really about, you know, are we deploying agents that continuously behave across a life cycle? Before I go to the group, Felix, I do have a question about this because literally every platform out there is advertising some form of AI governance, including platforms that Calibra partners with. Right. You can’t sell an AI capability without having some form of governance into this. Otherwise enterprises aren’t going to buy this. So giving me a clarifying statement about what Collibra is doing with AI governance.
[00:50:05] Speaker C: Yes. So I think AI governance is important on multiple levels. Any platform where you’re able to build an agent will have some form of governance. Often governance in that context is used, means access.
[00:50:21] Speaker B: Right.
[00:50:21] Speaker C: What can my agent access? And that’s really important, the context of that platform. If I’m configuring an agent in Salesforce, obviously I have some governance capabilities within Salesforce to make sure what data my agent can access, what they can do now you’re going to have a lot of different platforms where you’re able to build agents, right? You’re your systems of records, your application platforms, your data platforms, you have custom harnesses, you have cloud core work, you have ChatGPT. So you have lots of different places, each independently offer a form of governance. But as an organization, you also need to have a overview across all those platforms 1 and 2. You have to be able to define your organizational policies and apply those to those platforms. And this is really where Collibria is focused on, is the end to end control play across all those different platforms in which you configure and run agents. So it’s on a higher level where we’re able to create almost a registry of all your agents, wherever they might be running, right?
Then we create a lifecycle and the process and the policy management around what do we want to do before that agent can actually run in production? Who needs to approve it, who needs to review it, who needs to have assessment and attestations?
And then three, how do we make sure we enforce those policies? And then enforcement can happen natively in that platform because that platform offers those capabilities and then that’s great. Or if, for example, I’m building a custom harness, I don’t have a way to natively enforce it and this is where then our policy enforcement capabilities can be applied.
[00:51:59] Speaker A: Yeah, Felix, I think of this in layers. I mean, we could do Access control across SaaS and clouds, we can do some level of entitlements there. But as we start getting into more complexity around policies, not only thinking about a push to production, but also monitoring AI and then thinking through that life cycle, acknowledging that our data has an input around this, we start putting all this together and saying all my data is all over the place, all my agents are going to be all over the place. How am I going to manage this efficiently?
I think that’s what you’re getting at when you talk about Collibra and AI’s end to end control plane. I want to bring Joanne back. I cut you off earlier, Joanne. So where do you want to go? We have a few minutes left.
[00:52:50] Speaker B: I just wanted to make the comment, you know this framework that I meant, I mentioned before, part of the reason to do it in, in this fashion from the kind of a big blue ocean down. And it’s not to create a morass of tasks and it’s certainly not to complicate matters, it’s actually to simplify because to your point, I put autonomy at the end because the goal is autonomy. But in order to have autonomy, you need to build trust. And trust is not only within the organization, it’s outside of the organization. Can this organization that is running a lot of AI now be trusted given hallucination, given the taxes that we talked about before?
Building that trust comes from defining agency, defining authority, accountability, and all the A’s on that list.
So autonomy. You can build agents to be autonomous, as we have, but we release them and in a trust framework, and there’s a path, a path to get there, and that’s how much feedback in each loop. Who has accountability, how well do the agents run? And it all goes to own your AI and trust your providence and lineage, which are two words that you did not name earlier as part of what you’re putting out there in your governance framework. Regardless of whether it’s this or any other, governance has to start with trust. And trust has to be determined by evidence.
Evidence comes from provenance and lineage. And to Felix’s point, you can build agents in any kind of software you want or, or talk to it through an API or an MCP server. It doesn’t matter. But that auditability in the. In the provenance and lineage needs to be the foundation of trust. So many organizations are coming to the realization that part of our AI needs to be prescriptive and part of it needs to be probabilistic. And it’s the hybrid that gets you where you want to go.
[00:55:00] Speaker A: Thank you, Joanne. Let’s just do a speed round. Liz. Derek. And then we’re going to end with Felix. Liz, go ahead.
[00:55:06] Speaker F: I’m really, really interested. I first of all agree with everything that you guys were just talking about around governance, because it’s my heart.
But what about business case? Have you. I would love to hear from you about what kind of business case you’re making for supporting your clients.
[00:55:24] Speaker C: Of course.
[00:55:26] Speaker A: Go ahead, Felix.
[00:55:27] Speaker C: Yes. So I think it depends on what we focused on. This AI governance. One, it’s around, obviously, compliance, right? There’s just regulation that organizations need to comply with, and we help them do that much more effectively than if they would do that without quadibria, reducing risk, things like that. But ultimately, the bigger business case is giving organizations the confidence to actually put AI use cases in production and doing that more effectively. And this is, I think, the biggest business case. There’s a lot of promise with AI. When I see a lot of organizations, they really struggle getting their AI use cases in production. And a good governance actually helps them do that more quickly.
[00:56:09] Speaker A: I’m going to call that Calibra’s value proposition. Go ahead Derek.
[00:56:13] Speaker D: Yeah, it’s just one of the things in looking at this in the way doing the governance, especially when you’re talking about the output of this thing. I’m really looking at the AI governance implementation and the layered effect. You know when you look at layer one and be AI data governance type of services, AI the second layer would be AI guardrails making sure you have the controls in place to make sure make it work. But also layer three got to have a human in loop validation. Until we get more trusted capabilities and services out there where we can rely on AI, we have to have human intervention and looking at and validating the things that’s real. And the fourth is looking at AI threat intelligence, continuous monitoring. Those are let us know where our holes are, where the new holes are developing and the service taking place. I see AI as a future winner. It’s not going to be the organizations with the most powerful AI engines, but ones with the most trusted AI services. That’s where it’s going to be and I think people need to understand it’s, it’s a, it’s a journey, it’s not going to happen overnight.
[00:57:08] Speaker A: Thank you Derek. Felix, last word here.
A lot of enterprises have that gap between what they’re experimenting and what’s going on in production metric I got that only about a third of companies are seeing their earnings impacted by their AI investments.
What is your, what are you advising enterprises in order to start closing that gap?
[00:57:35] Speaker C: Yeah, that’s what we’re seeing the AI promise and the AI reality there’s still a gap. So how do you to your point close that gap? And I would say as well what we talked about, governance matter and governance of both the context. So there’s more confidence and accuracy in what the agents actually do and then governing the agents and what they allow it to do. So there’s confidence in the use case more broadly so the organization knows that they can put these things actually in production and they feel like they stay in control because if that’s not the case, it’s not going to happen.
[00:58:09] Speaker A: Felix, really great to have you here. Thank you for joining. I agree with what you’re saying. What you’re saying it is actually easy to build an AI agent. It’s really hard to trust them. And when you start getting to scale and having dozens and hundreds of agents, it’s important that you have that governance layer in place. So stop guessing and start governing. Let’s register for Calibra’s context governance series@calibra.com context-governance the link is here on the whiteboard. It’s also in the common stream and I will speak be speaking. I think it’s at the September 23rd episode around unstructured data governance. Felix, thank you for having us joining us here and thank you for being a sponsor.
[00:58:53] Speaker C: Thanks for having me. It was a pleasure folks.
[00:58:55] Speaker A: We’ll be back next week, November, September 18th about talking about tech budget 2027, what AI bets to fund and what IT and security projects to protect. On the 25th we’ll be talking about getting hired in the AI era, standing out, signaling judgment and landing the role. Thanks for being here. Thanks for all the comments. I’m sorry we didn’t get to everybody’s questions. I will data mine these and see if I want to help out with an upcoming episode or in an upcoming blog post.
Folks, 911 here, never forget and thank you for joining here this week. We will see you here on the 18th. And again thank you Felix for joining us and thank you for Calibra being our sponsor. Everybody have a great weekend.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。