Beth Rudden, CEO & Chair of Bast.ai
Beth Rudden, CEO and Chair of Bast.ai, offers a candid, technically grounded tour of what it really takes to make enterprise AI trustworthy, auditable, and economically defensible at scale. Throughout the interview, she argues that controls belong within the surrounding structure—ontologies, data scopes, and reasoning traces—rather than in ever more aspirational prompts, and she shows how ontology-driven, agentic architectures can both reduce compute and accelerate deployment while satisfying emerging regulatory and legal scrutiny.
Along the way, Beth illustrates these ideas with concrete stories: an internal auditor demanding a real explanation instead of a feature-importance chart, a nursing program that cut dropout by restructuring knowledge around retrieval needs, and enterprises learning the hard way that owning their ontology library—not their favorite model—is what ultimately determines leverage.
Top 3 Takeaways
- Explainability is structural, not moral. Telling models to “be accurate” or “never hallucinate” is wishful prompting; real control comes from governing what data is in scope, how meaning is defined via ontologies, and how every answer can be traced back to specific source content for auditors, regulators, and courts.
- Ontology-driven, agentic systems beat monoliths and raw RAG for enterprises. By decomposing work into small, auditable agents operating over domain-specific ontologies, organizations can cut compute (by sending less noise), reach production in weeks once problems are well-structured, and swap underlying models as commodities while preserving their core knowledge assets.
- The main failure mode is epistemic, not technical. Most organizations falter because they do not know, in a structured way, what problem they are solving or how their knowledge should be modeled; the winners will be those that invest in building and owning a library of ontologies tuned to real retrieval problems, closing the gap between “we have data” and “we can justify this specific AI-driven decision.”
The Beth Rudden Interview
What broke in earlier enterprise AI that convinced you explainability had to sit at the core?
I can tell you the exact scene because it keeps replaying. An internal auditor has a rejected vendor application on her desk and a letter from the vendor’s lawyer asking why. The AI did the rejecting. The closest thing anyone can produce is a confidence score and a feature importance plot. She reads it twice and puts it down. “That’s not an answer,” she says. “That’s a description of an answer.” She’s right. The industry spent years aiming controls at the wrong layer — writing increasingly elaborate instructions to the model (“don’t hallucinate,” “be accurate,” “only use verified sources”) as if moral language were an engineering constraint.
Those are wishes, not controls. What actually governs output quality is the proxy structure around the model: which data is in scope, which framework defines terms, which ontology anchors meaning. Earlier approaches tried to control the model directly. We built Bast to control the structure within which models operate.
What does “system of record for AI” mean operationally?
Strip the marketing: it means every source document, every version of your ontology, every retrieval context, and every reasoning step gets versioned and snapshotted — the way financial systems version transactions. When you trace a payment in a well-run organization, you walk a trail of evidence. Documents, approvals, system records, timestamps and names. AI systems, for the most part, were not built to leave that trail. That’s not a feature someone forgot — it’s what most of these systems actually are.
An enterprise buyer should ask one question: “When your model gives me an answer, can I walk it back to a specific piece of source content that I gave you?” The vendors who can answer yes are operating on a different architecture. The difference will matter the first time a regulator, a plaintiff’s lawyer, or a court asks the same question — and as of this year, they’re asking.
Where does ontology-driven AI outperform prompt-plus-retrieval?
Anywhere a word means something specific in your domain, it doesn’t mean on the internet. When someone asks whether a person is “nostalgic,” whether a situation is “complex,” or whether a relationship is “mature,” those words are already saturated with theory. RAG retrieves chunks that are semantically close — but semantic closeness isn’t logical relevance. Without a named framework making the lens explicit, the model defaults to the statistical center of internet usage.
Here’s where the industry conversation has moved, though. Evaluation frameworks like RAGAS — which measure faithfulness, context precision, and answer relevance — should run continuously on every system. They’re the standard, or should be. But the real shift is past evaluation-as-checkpoint entirely. What we’re building now are agentic frameworks in which every task is a discrete, verified operation with controlled input and output. A greeting agent. A knowledge-base agent. An end-conversation agent. Each one scoped, each one auditable.
When something breaks or drifts, the system identifies the gap and surfaces it — not as an error, but as a signal to the knowledge manager. That’s the microservices promise finally fulfilled in AI: not bigger models, but smaller, composable, self-healing agents operating against a structured knowledge base. The ontology makes this possible because it provides every agent with a map of what things are and how they connect in this specific domain, so retrieval maps to the actual business problem rather than relying on guesswork.
70% compute reduction, eight weeks to production — what assumptions sit behind those numbers?
The 70% comes from a structural move: when your ontology is explicit, the model doesn’t need to infer your entire conceptual model from scattered files. You send it less noise and more signal — smaller payloads, fewer tokens, framework-grounded queries instead of open-ended prompts, hoping the model figures out what you mean.
The eight weeks are the human time, not the technical time. We can get someone onboarded and running in minutes. The technology is not the bottleneck. The bottleneck is getting the humans to understand their own problem well enough to structure it. Eight weeks is what it takes for subject-matter experts to work through ten questions that aren’t technical at all:
- What business problem are you solving first?
- What outcome do you want?
- Who is the user, and what are they trying to get done?
- What information does that user need?
- Where does it live today?
- Which parts are repeatable and which need human judgment?
- What would make you trust the output?
- What could go wrong?
- Which sources should it use and which should it avoid?
- How will you know this is working after launch?
Einstein’s line applies—if you gave him an hour to save the world, he’d spend 59 minutes understanding the problem and 1 minute on the solution. The eight weeks are 59 minutes. The solution — the actual deployment — takes minutes once the structure is right. And the vision is that this work compounds. We envision a world where everyone has their own set of AI assistants they choose to grow and feed — curating knowledge, refining ontologies, adding sources — so they get the right information at the right time for the problems they actually face. The assistants get better because the humans did the work of understanding what they needed.
Where organizations underestimate: that gap between “we have the data” and “we know what problem we’re solving with it and how the knowledge needs to be structured to solve it.” Gartner is predicting 60% of agentic AI projects will fail this year due to a lack of AI-ready data. Fivetran’s readiness index shows only 15% fully prepared — even as 60% invest millions. The readiness problem isn’t technical. It’s epistemic. Organizations don’t yet know what they know.
Does Karpathy’s LLM Knowledge Base validate your thinking, or miss something essential?
It validates the instinct. Structured, human-readable knowledge layers beat shoving everything into a vector database. His markdown-wiki approach is great for collection, condensation, and recall — it gathers and compacts everything the system knows. But a wiki happily preserves different internal ontologies side by side; it doesn’t reconcile them. It can’t enforce access controls, handle concurrent updates across hundreds of users, or tell you why two pages disagree.
What we build is the layer underneath: a formal domain model where shared meaning and provenance live, where you can ask “where in our ontology does this concept sit, and what data path supports the statement?” The productive move is to layer both — use the wiki/memory layer to gather, then require that canonical decisions be anchored into the ontology. Mapped to classes, relations, and sources. Or explicitly flagged as unmapped, which is itself a design signal — it tells you where to extend the ontology or where your own internal language hasn’t converged yet. That flag is one of the most valuable outputs the system produces.
How much of an enterprise AI strategy should be built around optionality?
There’s a useful symmetry here. On one side: a powerful model inside a powerful harness — thin, evolving, increasingly commoditized. On the other hand, powerful data inside powerful ontologies — structured, linked, specific to how your organization actually understands its world. And I want to stress the plural. An ontology is a map, and you need many maps. Each knowledge base gets its own. A clinical pharmacology team and a billing compliance team don’t share a map — their concepts overlap, but their structures serve different retrieval problems.
Ontologies are also lenses. You can look at the same knowledge base through a patient safety lens, a regulatory lens, a cost lens — as many filters as you can define. The idea that an organization has “an ontology” the way it has “a database” is the wrong mental model. You have a library of them, and the ability to compose and swap them is where optionality actually lives.
Models are rented. The durable asset is that library of structured knowledge and the ontologies that make it navigable. Any enterprise building its strategy around allegiance to a single model provider is investing in the commodity layer and neglecting the layer that’s actually theirs. MCP just got donated to the Linux Foundation’s Agentic AI Foundation — adopted by OpenAI, Google, Microsoft and Anthropic.
Over 10,000 active public MCP servers. The plumbing is commoditizing fast. Your infrastructure should let you swap models the way you swap cloud regions — without rewriting your understanding of your own domain. The organizations with leverage in 18 months are the ones that own their ontologies, because those are what every model reads and every agent operates against. Everything else is interchangeable.
Where does symbolic or rules-based logic still change business outcomes?
Anywhere the domain has actual rules, and the cost of being wrong carries real liability. But I’d frame it more precisely: symbolic reasoning is what makes the model’s output legible and the system’s operations composable.
In practice, what we build is discrete agent pools — a greeting agent, a knowledge-base answer agent, an end-conversation agent, a redirect agent — each with a defined scope, controlled input and controlled output. That’s not a monolithic model deciding what to do. That’s an orchestrated pipeline where each step is verifiable. When the knowledge-base agent answers, you can see the exact chunks it pulled, verbatim, from the knowledge base. When it can’t answer, it routes to the next agent in the hierarchy — web handler, out-of-scope response — and the whole flow is documented and auditable.
The same problem appears in the documentation and the code. We tell AI coding assistants to “generate documentation from the codebase” as if the repository contained the full system understanding. It doesn’t. The AI summarizes per file, but cannot conjure the architecture you never made explicit. Locally accurate, globally incoherent. That’s the vibe-coding governance problem Dan has been writing about, and it’s the same structural gap. The symbolic layer — the ontology, the reasoning flow, the agent hierarchy — is what holds the global coherence that pure neural approaches can’t produce on their own.
What still needs to evolve for enterprises to trust autonomous agents?
Our designer, Adam Cutler, pushed hard early on against AI-washing language in our product — no claims that the system is “thinking” or “reasoning like a human.” Even in our trace logs, you’ll never see the word “reasoning” or “thinking.” We describe concrete operations: retrieving, collating, comparing, checking and ranking. I was on a call with a team of public librarians last week and told them — if you catch anything that sounds like a human being, tell us, because it’s like getting catfished by the AI. That discipline matters because it keeps users focused on evidence and process instead of the fantasy that the machine has interiority.
The same move needs to happen at the agent-instruction level across the industry. Most frameworks still rely on moral language — “be careful,” “be critical of your reasoning” — as if the model took those as genuine obligations. Trust comes from harness design: which ontology gets loaded, how outputs anchor to evidence, how gaps get surfaced. And gaps are the interesting part — when our system can’t find something in the knowledge base, that surfaces as a flag the knowledge manager can see, assign, and resolve. It’s a workflow, not a failure. That’s what governance actually looks like in production.
MCP standardizes connection — 10,000+ servers, OAuth 2.1 and formal audit support on the 2026 roadmap. But only one in five companies has a mature governance model for autonomous agents, even though 72% report running them in production. That’s a 60-percentage-point governance gap running live right now.
What do executives misunderstand about AI ethics and deployment economics?
They think ethics means telling the model to be a good person. A well-known investor recently posted instructions to a chatbot — “behave as a world-class expert, never hallucinate, don’t make anything up” — to millions of followers. You cannot give a chatbot an instruction that changes what it is. “Don’t lie” is a wish, not a control.
The actual ethical infrastructure is the proxy structure: what data is in scope, what framework defines terms, what ontology anchors meaning and what evidence view surfaces assumptions. Organizations that invest in that structure deploy faster because stakeholders can see how conclusions were reached. The ones writing increasingly elaborate “be responsible” prompts are installing smoke detectors in buildings that lack fire-rated walls. The detectors work. They’ll tell you the building is burning. They won’t stop it.
Meanwhile, the legal landscape has shifted under everyone. There are now 600+ AI hallucination cases on record implicating 128 lawyers. The EU AI Act’s high-risk deadline hits August 2. Courts are converging on a standard of proof over prediction. The question is no longer “is your model explainable in aggregate?” It’s “Can you produce the basis for this specific decision affecting this specific person?” And there are families in courtrooms right now asking exactly that — Raine v. OpenAI, the Setzer case, at least ten active lawsuits involving minors. The institutions that bolted on a dashboard will produce a screenshot of a working chatbot and an internal Slack thread of nervous engineers. Neither will function as a defense.
What separates orgs that operationalize trustworthy AI from those stuck in pilots — or worse, scaling AI that isn’t trustworthy?
Whether they own their ontologies or just rent the model. The formula is (Model + Harness) + (Ontologies + Data). Organizations stuck in pilot purgatory are over-investing on the left side — which model, which agent framework — while the right side stays informal. Tribal knowledge. Undocumented domain models. Concepts that mean different things to different teams. And crucially, different teams need different maps of the same territory. The clinical team, the billing team, and the compliance team don’t navigate their knowledge the same way. Each needs an ontology that reflects their problem and retrieval pattern — not a single, canonical schema that flattens everything into a single view.
That’s what we learned building a nursing program with 50% attrition in two courses. We spent months not building anything — just studying why students were struggling. The answer was a structural problem: the syllabus was organized by chronology, the textbook by topic, but neither matched how students needed to retrieve knowledge under exam pressure. So we built the ontology around the actual retrieval problem.
Weeks mapped to exams. Topics mapped to weeks. When a student asked for quiz questions for exam three, the system pulled only material from the weeks the exam covered. Then we layered different lenses on the same content — scenario questions, analogies, Jeopardy-style prompts — different cognitive grooves, each a different way of looking at the same knowledge base. It took a year. The result was $3.2 million in attrition costs avoided and a 30% reduction in dropout.
Everyone I tell that story to asks the same question: “But isn’t AI supposed to be fast?” I live in the real world. Productivity gains do not fall off the back of trucks. They come from the patient, the joint work of subject-matter experts, knowledge engineers, and operators who together build a structure that can hold the problem. Deloitte’s numbers show a 41-percentage-point gap between enterprises piloting AI and those prepared to operationalize it.
Gartner says 60% of agentic projects will fail this year from data readiness alone. The differentiator in 18 months won’t be model access. It’ll be who built the library of ontologies that make their organizational knowledge navigable — so any model can reason over it, any lens can be applied to it, and any human can audit the result. The dashboard is the easy part. The structure underneath is where the year goes.
About Beth Rudden,
CEO and Chair of Bast.ai
Beth Rudden is the CEO & Chairwoman of Bast AI. Beth is a dynamic global executive and market creator with over two decades of expertise in IT leadership and cognitive science. Her tenure as Chief Data Officer, Chief Data Scientist, and Global Talent Transformation Leader saw her deftly steer digital transformations through the strategic deployment of trusted, high-integrity AI systems. Beth’s vision transformed analytics and AI advances into a profitable $2B enterprise, placing her among the foremost 100 leaders in AI Ethics.
A staunch proponent and trailblazer of ethical AI, Beth advocates for the democratization of transparent and responsible AI technologies. She founded Bast AI in 2022, which revolutionizes how organizations leverage full-stack explainable AI to improve business impact and end-user adoption of AI technologies across industries and sectors.
Beth honed her analytical acumen with a Master’s in Anthropology from the University of Denver and sharpened her literary skills with a Classics Degree from Florida State. As an influential keynote speaker, inventor with over 50 patents, and author of AI for the rest of US, her contributions are formidable. Beyond her professional achievements, Beth flourishes in her roles as a storyteller, the wife of a soldier, mother, mentor, and member of the revered science and technology board. Her dedication to fostering educational and innovative pursuits is mirrored in her active role on the Maryville University Board of Trustees, where she tirelessly works to sculpt the next generation of innovators.
For more serious insights on AI, click here.
Did you find this interview with Beth Rudden useful? If so, please like, share or comment. Thank you!
The cover image is AI-generated from the author’s prompt and Aravind’s source photos.
Leave a Reply