Episode 12 — Your Employees Are Creating AI IP. Is Your Company Capturing It?

Introducing Werner Goertz

Sunil Baliga: Token Drop, Episode 12 — and today we have a very special guest, Werner Goertz. Werner and I have known each other for a number of years, as has Sajjad — Sajjad and Werner worked together, and Werner, I believe I was your customer at a previous company as well.

A bit of background on Werner: he's the founder of Anicca AR, an analyst relations firm. Previously, he was an analyst at Gartner, and before that held analyst relations leadership roles at AWS and IBM — he knows the AI space extremely well. Werner, anything else you'd like to add about your company before we dive in?

Werner Goertz: I think you summarized it perfectly. We're living through a really exciting inflection point in the AI space right now, and I'm looking forward to this conversation.

Sunil: You mentioned "Anicca" — Sam actually asked me about that beforehand. Can you tell us what it means? I thought it was lovely.

Werner: It's the Buddhist term for impermanence — the idea that everything in life changes. That applies to life circumstances generally, but I thought it applied especially well to the technology space Anicca AR operates in. What we've observed is that impermanence used to play out on a monthly, quarterly, maybe annual cadence. But the rate of change keeps compressing — in the AI sector specifically, that impermanence now shows up almost daily.

Setting up today's topic: governance beyond compliance

Sunil: The reason we invited Werner on is that we were talking recently about some areas that connect directly to a previous episode of ours on AI agent governance — where we approached governance mainly from a security, compliance, and data-protection angle. Werner's point was: those things matter, but there are other dimensions too. What knowledge are we actually creating? What workflows are producing real value? Which prompts should be standardized across the organization?

The other thing Werner raised: in most enterprises today, a lot of AI work happens at the individual level. A salesperson discovers a prompt that makes them, say, 25–50% more effective at writing proposals. A marketer builds a workflow that cuts campaign creation from hours to minutes. That knowledge lives entirely in that employee's head, or on their laptop. Organizations have figured out how to scale and share that kind of institutional knowledge for other categories of work — how do they do it for AI-generated knowledge specifically?

Standard disclaimer before we start: all opinions expressed are those of the individual speaking, not necessarily of the company they work for or represent. Werner, what do you think?

The "hit by a bus" problem

Werner: Well said — and in my view, 2026 is a year defined by several inflection points at once. The examples you mentioned are all manifestations of the same underlying inflection in enterprise AI: we're at a genuine crossroads. Historically, individuals created prompts on their own ChatGPT or Claude accounts and generated real value that way.

But here's the catch: the way these AI tools are set up today, an individual account and an individual chat history essentially get brought into the enterprise informally. What happens if that employee gets hit by a bus — or simply goes and joins the company across the street? All of that prompt history, all of that contextual knowledge, and all of the outcomes of that AI-assisted work leave the company with them, without ever having become real institutional knowledge.

So the inflection point is this: what governance criteria, what AI management practices, does the enterprise need to put in place to ensure that the value individuals create through AI actually stays with the company — and genuinely contributes to the company's intellectual property? That raises an immediate follow-up question: who's responsible for that? What function or role within the enterprise owns institutional IP creation from AI? That's exactly what customers come to me for — fractional AI leadership guidance: building guardrails that aren't just about regulatory compliance, but genuine governance guardrails that ensure AI is used responsibly, effectively, and in a way that contributes to the enterprise's intellectual property as a whole — so that the creativity and productivity AI enables actually stays with the company.

Sunil: This is interesting, because at every company I've worked for, there's always been a push to modularize software into reusable IP — build a library that other teams can draw from instead of reinventing the wheel. Great idea in principle. But the real challenge has always been getting people to actually contribute to that library, and getting other people to look there first instead of just building their own version. It sounds like the same dynamic applies to AI.

From individual silos to a knowledge fabric

Werner: Great example — because enterprise software development, over the preceding decades, has largely been about enabling teams. Some software was literally named "Teams" for exactly that reason: it was built as the foundation for leveraging individual creativity for the benefit of team collaboration and enterprise IP.

Then these AI tools arrived, and suddenly we're back to individual user accounts, with chat history sitting at the individual level. Even genuinely beneficial, revolutionary tools like Claude Cowork ask for data sources that live not just at the individual level, but at the individual device level — folders on someone's personal laptop. Get hit by a bus, or go work for a competitor, and that laptop gets wiped, along with everything you contributed. So after decades of enterprise productivity software being built team-first and enterprise-first, we now find ourselves with powerful tools that are, by default, designed for personal productivity. That's the core challenge.

Sunil: What do you two think — Sajjad, Sam?

Sajjad Khazipura: I recognize exactly what Werner's describing. I think we're beginning to shift enterprises away from centralized, standardized workflows and processes that dictated what employees did and how they did it. Historically, everything else an employee came up with — new selling techniques, new ways of building something — stayed locked inside that individual. Despite every effort to get people to write it up, blog about it, or push it into a central repository, you're fighting an uphill battle. It's genuinely hard to get people to do extra work just to aggregate knowledge.

What the arrival of LLMs, and especially agentic architectures, gives us is a real opportunity to stitch enterprise knowledge together into a fabric — one that spans both the core knowledge sitting in centralized databases and the knowledge held by departments, work groups, and individuals, integrating and aggregating all of it. That fabric becomes the enterprise's key differentiator. It's always been understood that a company's most important competitive asset is its knowledge — the hard part has always been actually materializing that knowledge into something usable. I think we're now at the inflection point where that's finally possible.

This connects directly to what we've been building, Sunil: knowledge aggregation and knowledge processing — constructing knowledge graphs that let us stitch together islands of knowledge through appropriate linkages. Those graphs learn from user interactions and also absorb the enterprise knowledge already sitting in centralized repositories. In effect, you're getting enterprises to double down on building their own knowledge assets even as they pay outside providers for the underlying LLMs — using those tools as the mechanism for aggregating and building your knowledge, not as the end product itself.

Institutionalizing the role of the AI leader

Werner: I'm glad you picked up on my "inflection" framing, because this is another one we'll keep seeing through 2026: the centralization and institutionalization of the AI platform — and of the AI leadership role — across every stakeholder in the enterprise. That's step one.

In the past, individuals brought their own AI workloads into the company; maybe a business unit aggregated usage at the team level. Now, enterprises are moving up the value chain and formally institutionalizing the AI leadership role — some call it a Chief AI Officer (CAIO), others call it a Chief Data and Analytics Officer. Whatever the title, this centralization needs to happen close to, or at, the C-level, so that role has a genuine bird's-eye view of everywhere AI touches the organization — and honestly, I have a hard time finding any part of a modern organization that AI doesn't touch.

So the first step is appointing that structural, organizational role to aggregate everything. The second step is exactly what Sajjad described — using technologies capable of aggregating, storing, and leveraging the enterprise's entire knowledge history, whether through graphs, ontologies, or platform tooling. Those are implementation details, but institutionalization at the leadership level, followed by tool aggregation at the operational level, is what we'll keep seeing through the rest of 2026.

From data to knowledge: a discipline that doesn't fully exist yet

Sajjad: That's a huge shift, really. Up to now, enterprises have mostly dealt in data — data processing, data assets. Moving from data to knowledge is a genuinely big shift: building knowledge tools, curating knowledge, knowledge management, knowledge engineering, and fusing knowledge across domains according to a business's specific core specialty. A medical device company might have medical device ontologies, but it also has sales, finance, operations, and manufacturing functions. How do you integrate all of that knowledge through a common layer that lets you fuse and work across it — and use that as the foundation for orchestrating integrated workflows that can operate autonomously?

Werner: I like that you say "integrated workflows," because historically, workflows ran in silos, and in a fully deterministic way. R&D built a product based on algorithmic functions with deterministic outcomes; legal worked the same way; so did finance. Now we have the opportunity to aggregate all of that data into a holistic, probabilistic view of the entire corporation — and build applications that transcend those original silos and transcend individual deterministic outcomes, letting you look at the business strategically, as a whole.

Sajjad: Exactly — and that lets you translate a CEO's objectives directly into rippled-through work actions and workflows for everyone across the enterprise.

Three stages of AI maturity

Werner: I want to sketch out three maturity stages I typically see in how organizations adopt AI. It usually starts with customers asking me how to institutionalize AI comprehensively across the organization — the goal there is creating efficiencies across existing workflows and existing company paradigms. That's stage one: institutionalization.

Stage two is the productization of AI — infusing AI workflows directly into product or service development. This transcends the silos we were just discussing: product, finance, and legal (for compliance) all get involved. Productizing AI is the next step in the maturity curve.

Once you have that holistic, enterprise-wide view — accounting for the company's strategic assets at the CEO level — you can start identifying genuinely new markets and opportunities nobody has penetrated yet: white spaces, where your company's assets and ontology point toward disruptive new business opportunities. This is where I like to use a term I've been criticized for — "weaponizing AI" — using it to identify white spaces, penetrate adjacent markets, and defend those positions against new entrants. Those are the three maturity stages I see in AI adoption.

Enterprise AI vs. institutional AI: a replaceable "lobe" vs. a persistent asset

Sam Pooni: I want to build on this. Knowledge graphs — and more broadly, ontologies and retrieval-augmented approaches — are really the substrate for what you'd call institutional AI specifically. They're model-agnostic, provenance-bearing, governed, and genuinely queryable. In that kind of architecture, the LLM itself becomes a replaceable component — almost like a "lobe" you can swap out — while the institutional knowledge graph is the actual persistent asset. Is that a fair way to frame it: enterprise AI sells the lobe, while institutional AI builds the underlying asset?

Sajjad: I think what Werner's really describing is aggregating and concentrating enterprise knowledge as an abstract entity — and how you manifest that technically (an LLM, some graphs, an ontology to guide it all) can and will change; that's a fast-moving layer. What matters conceptually is building a flywheel that continuously adds to enterprise knowledge across every function — and that becomes what you curate, govern, and use as your core differentiator: for defense, for offense, and as a vehicle for innovation across everything the enterprise does.

Sam: Right — so what's captured, structured, governed, versioned, and made queryable survives changes in people, products, or even the underlying model, and compounds over time. I see institutional AI and enterprise AI as genuinely two different things. What you need structurally is a formal model of the business — that's the ontology piece — a governance or contract layer defining who can do what and with what authority, and a record of agent actions. Ultimately, the real question is how you institutionalize that overall body of knowledge.

Sajjad: It's honestly not a discipline that fully exists yet. We talk a lot about data management, data lifecycles, protecting and reporting from data — but not many people talk about knowledge curation specifically: what is knowledge as a unit, what's its provenance, how do you store and process it? That discipline is much less mature.

A self-improving, ever-changing system

Werner: The concept I'd add is the organic, ever-changing nature of this — back to the "Anicca" idea. What you have today isn't what you'll have tomorrow, and it won't live in one single monolithic LLM. Quite the opposite — over time, you'll build out a continuum of multiple foundational contributors, some of which will be generatively adversarial to each other. Imagine a workflow where you build something using specialist LLMs trained on specific tasks, and then, to keep that original workflow honest, you introduce another foundational entity — another LLM, or some other mechanism — that, in a GAN-like way, tests and probes the first system's output for compliance, correctness, and hallucination. That's just one example, but the broader idea is exactly what Sajjad described: an organic, ever-changing portfolio of technologies, including but not limited to LLMs, that comprehensively governs the enterprise — and governs itself.

Sunil: Isn't that adversarial piece essentially what we now call a verifier?

Sajjad: It's one form of verifier, yes. To extend Werner's point — think of how the human brain works during sleep: short-term memories get processed and either stored as long-term memory, if they matter, or cleared out if they don't. All of that integration and aggregation happens biologically while you rest — which is part of why rest is necessary. These enterprise systems will effectively do that same kind of consolidation continuously, day and night, without ever needing to "sleep" the way a human does. An enterprise never rests — these systems will be constantly organizing knowledge, verifying its accuracy, establishing lineage (where did this come from, is this the right source, are there conflicts), managing access controls (who should and shouldn't see this knowledge), and checking whether knowledge fusion across manufacturing, R&D, and legal has actually been done correctly. It's effectively a living entity that continuously refines itself and strengthens the company's competitive position.

Werner: That self-improving quality is worth highlighting. Generative adversarial networks were originally seen as a possible precursor on the road toward AGI — I won't get into whether that's accurate — but the GAN concept implies more than one entity exercising editorial control over another; it implies a genuine back-and-forth, ideas being bounced between systems and continuously improved through that exchange.

I met someone at Duke's Fuqua School of Business recently who'd built an LLM specifically designed to be contrarian — a fairly simple setup. Its sole purpose was to push back: you'd test a thesis against it, and it would deliberately generate output designed to challenge your original prompt. I found that incredibly useful — not in a "who's right" sense, but because that ping-pong of going back and forth continuously refines your original thesis, so that with each round it becomes more substantiated, better founded, and more defensible.

Sajjad: That connects to Judea Pearl, the well-known UCLA computer science professor — he wrote a book about six or seven years ago called The Book of Why. One of his central arguments is that neural networks are very good at correlation: you see a ball drop and correlate that with gravity. But establishing actual causation is different — and reasoning about counterfactuals (what would have happened if this event hadn't occurred?) requires capability well beyond what today's LLMs can do on their own. That's exactly where reasoning systems that operate directly on structured knowledge can help — it speaks to what we've been describing: continuously refining and battle-testing knowledge through adversarial scenarios, distilling it down to be more precise and more relevant to the actual business need.

Is institutional AI generative or extractive?

Sam: In a way, would you say institutional AI is generative — because it builds and prunes on top of itself over time — while enterprise AI, by comparison, is more extractive? It holistically looks at what was done and how, and improves on it.

Sajjad: Not exactly — "generative" can also imply predictive, in the sense of an LLM generating something without any guarantee it's correct, and that's not what we mean here. What we're describing is synthesizing knowledge: assembling verified pieces of knowledge like a jigsaw puzzle, so the complete picture is logically integrated, fits together cohesively, and leaves no room for the kind of predictive or probabilistic error you'd get from a generative model. This is meant to be the defining, dependable knowledge base you use to make real business decisions and run operations across the enterprise.

Werner: No question this institutional knowledge base will have real generative impact and generative efficiencies. But the interesting part starts once you apply it in a genuinely agentic context — because that's where it can fact-check itself and harness its own capabilities. Take model drift, for example: the system will be able to observe its own outputs drifting away from the guardrails originally set for it, and self-correct. It will also be able to control the "blast radius" of agentic functions. Right now, we can let agentic AI loose to, say, generate millions of emails — but if a serious error occurs, what actually stops it? What's the blast radius when something goes wrong? Honestly, nobody fully knows yet. The real value of this kind of institutional knowledge and workflow base kicks in when it can start addressing exactly these problems — governance, model drift, blast radius, AI security, runaway agentic behavior — issues even we as humans struggle with.

Sunil: That contrarian Duke LLM is one I'd personally never want to use — I've worked with a couple of people like that in real life, and it drove me nuts. Whatever I said, they'd take the opposite side; say "up," they'd say "down." I imagine whoever built that LLM trained it on people just like that.

Werner: It's actually a fun exercise, because it forces you into a kind of Socratic dialogue with yourself.

Sunil: Sure, but I can't reach through the screen and strangle it the way I sometimes wanted to with those people.

Turning enterprise knowledge into licensable IP

Sunil: This whole discussion about knowledge becoming IP reminds me of something I've seen in the Valley — I have a friend who works in IP licensing at a large local company. They build their own chips, and part of his job is licensing out chip IP to other customers, since licensing itself isn't their core competency. Couldn't the same logic apply to AI — taking a prompt or a workflow, turning it into IP, and reselling it the same way, monetizing something that's extremely valuable even if it's not your core business?

Werner: Let me give you a concrete example — I had an interesting conversation recently with one of the major travel information platforms. Decades ago, Amadeus built a travel information platform that became the backbone source of flight and hotel data industry-wide. It's actually a great case study for the three maturity stages I described earlier.

You could treat that flight-data database purely as an efficiency play — AI creating efficiencies just by letting people query that dataset directly. That's stage one. Second, as you said, Sunil, a company could productize that data — build applications on top of it. I'm a United flyer, so think United.com infused with AI: that's productization. But here's where your original point comes in — a company could go further and use the entire platform to create an entirely new application, even a new industry. Imagine how powerful it would be if a company like Amadeus or Sabre built its own ontology on top of its existing data and used it to disintermediate the Travelocitys and Expedias of the world. That's the kind of "weaponizing AI" I mentioned earlier — using it to define entirely new markets and disrupt industries that would otherwise become legacy.

Vertical disintermediation and what it means for the global economy

Sajjad: Extending that further — there are roughly 8 billion people on the planet, and everyone brings something unique to the table. You specialize in your own uniqueness. A knowledge fabric like this effectively synthesizes everyone's specialized knowledge into a single, unified fabric. Someone who specializes in manufacturing at scale — say, Apple — doesn't leave any obvious reason why an Android phone maker needs to maintain its own manufacturing facilities; it could use Apple's manufacturing capacity and instead specialize in its own differentiation. You'd end up with companies that purely manufacture, and separate companies that purely design. It opens up the economy so that anyone with genuine talent or specialized skill can contribute and compete on the basis of their own unique capability.

Werner: We're so used to thinking in terms of vertical integration — what you're describing is essentially vertical disintermediation.

Sam: Exactly.

Werner: I want to reaffirm Sajjad's point about vertical disintermediation, but also flag the real tension it creates — think about the efficiencies it generates, but also the problems, given that we're operating in a world of interconnected global supply chains, with plenty of geopolitical friction already. Do we really want to introduce more disruption to global supply chains with a concept like this? If so, the consequences are on you, Sajjad — we'll blame you.

Sajjad: I'd argue the counterfactual: maybe those global conflicts wouldn't exist in the first place, precisely because everyone would be so interdependent that genuine collaboration becomes the only workable path.

Werner: Macroeconomic and geopolitical trends generally point the other way, but let's not go down that road today.

Sunil: That's a topic for next week's podcast.

Wrap-up

Sunil: Well, a pleasure as always, Werner — thank you so much. If you can stick around for a moment, we really appreciate your time.

Werner: This has been a lot of fun. Thank you, gentlemen.

Sunil: Thank you.