Fabrion Splits AI Reasoning From Action to Prevent Operational Errors

Democratising AI for the Enterprise With Fabrion — Ai4

Ai4

Placing generative models in direct control of transactional workflows causes 95% of enterprise pilots to fail, requiring companies to separate conversational intelligence from deterministic, permission-enforced execution layers.

AI-assisted summary of Democratising AI for the Enterprise With Fabrion — original source. Check the original for full context.

Source published: 2026-08-15

Detailed notes

## Background Published by Ai4 on YouTube on August 15, 2026, this presentation features Kunal Bhasin, Co-founder, CTO, and CPO of Fabrion. The video outlines Fabrion's dual-brain, full-stack AI platform designed to solve enterprise AI governance challenges by separating conversational reasoning from deterministic operational execution across complex industrial knowledge graphs.

## TL;DR Kunal Bhasin explores the core enterprise dilemma where conversational LLMs hallucinate while rigid deterministic systems cannot adapt, contributing to a 95% failure rate for enterprise AI pilots. Fabrion resolves this by introducing a dual-brain architecture that decouples exploratory conversational reasoning from an auditable, IAM-enforced operational execution engine. Grounded in a four-layer Knowledge Fabric and an open industrial meta-model, the platform maps complex manufacturing ontologies—from parts catalogs to warranty pipelines—into verifiable data graphs. This model-agnostic, on-premises system continually learns from human feedback and fine-tunes open-weights models to ensure enterprise data sovereignty and long-term vendor independence.

## Atmosphere & style The presentation carries the sharp, authoritative cadence of a technical founder keynote aimed squarely at pragmatic enterprise operators. Bhasin balances engaging anecdotes about viral LLM logic failures with dense, systems-level architecture diagrams and clear governance workflows. The tone is deliberate, engineering-focused, and reassuringly realistic, cutting through generative AI hype to deliver a disciplined blueprint for production-grade industrial automation.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=LN_xY92LLmE&t=4s) Enterprise AI faces a fundamental conflict between fast, conversational LLMs that hallucinate and rigid, deterministic legacy systems that cannot hold a dialogue. - [01:36](https://www.youtube.com/watch?v=LN_xY92LLmE&t=96s) MIT research indicates that approximately 95% of enterprise AI pilots fail to reach production deployment due to governance and reliability challenges. - [01:36](https://www.youtube.com/watch?v=LN_xY92LLmE&t=96s) The operational and financial penalties of executing incorrect transactional enterprise actions are substantially higher than merely generating incorrect conversational text. - [01:36](https://www.youtube.com/watch?v=LN_xY92LLmE&t=96s) Attempting to solve AI governance by simply using a secondary LLM to monitor the primary model without deterministic structural boundaries is an engineering anti-pattern. - [03:42](https://www.youtube.com/watch?v=LN_xY92LLmE&t=222s) Fabrion's dual-brain architecture decouples conversational exploration (Brain 1) from operational validation and execution (Brain 2). - [03:42](https://www.youtube.com/watch?v=LN_xY92LLmE&t=222s) Operational guardrails must be enforced through programmatic identity, access controls, and structured schemas rather than soft prompt instructions. - [03:42](https://www.youtube.com/watch?v=LN_xY92LLmE&t=222s) The underlying Knowledge Fabric links four core layers: granular IAM identity, semantic data ontologies, specialized domain agents, and deterministic system integrations. - [06:34](https://www.youtube.com/watch?v=LN_xY92LLmE&t=394s) Industrial enterprises with high complexity (such as automotive manufacturing with 10,000+ parts) require dynamic, AI-legible graph ontologies rather than static dashboards. - [08:03](https://www.youtube.com/watch?v=LN_xY92LLmE&t=483s) The platform's open meta-model standardizes messy enterprise data across Action Types, Certified Measures, Roles & Rules, and Analysis Directives. - [08:03](https://www.youtube.com/watch?v=LN_xY92LLmE&t=483s) A rigorous four-stage governance lifecycle (Declaration, Binding, Verification, Certification) allows agents to establish trust by returning grounded fallbacks or 'I don't know yet' when data is uncertified. - [10:02](https://www.youtube.com/watch?v=LN_xY92LLmE&t=602s) Capturing manual human overrides and enterprise queries enables continuous fine-tuning of self-hosted open-weights models, increasing platform determinism over time. - [10:02](https://www.youtube.com/watch?v=LN_xY92LLmE&t=602s) Enterprise AI architecture must be model-agnostic and deployable on-premises, creating a durable semantic layer that outlasts individual LLM releases and vendor choices.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=LN_xY92LLmE&t=4s) The Enterprise AI Dilemma: Conversational Agents vs. Deterministic Systems

Kunal Bhasin, co-founder, CTO, and CPO of San Francisco-based startup Fabrion (backed by 8VC), outlines the fundamental challenge of bringing artificial intelligence into large, complex enterprises, particularly in heavy industrial manufacturing. He observes that modern enterprise AI deployments typically manifest in one of two distinct flavors. On one hand sits what he terms "the confident agent"—an LLM that is brilliant, incredibly fast, and never sleeps, yet occasionally invents supplier discounts or revenue numbers out of thin air. The true breakthrough of this paradigm is conversational interface design, capitalizing on the natural human inclination toward dialogue to make complex technology universally accessible.

On the other hand sit traditional enterprise software systems. These legacy platforms are strictly deterministic, auditable, reliable, and compliant, but they fundamentally cannot hold a conversation. In these rigid environments, answering a novel question becomes a heavyweight change event where innovation velocity is measured in months or quarters rather than days. Fabrion’s foundational thesis is to build from the bottom up to fuse these two conflicting natures—the non-deterministic, conversational flexibility of LLMs and the deterministic, audited reliability of enterprise platforms—into a unified platform.

- **Company Background:** Fabrion is a San Francisco startup backed by 8VC focused on complex enterprise verticals such as automotive and industrial manufacturing. - **The Core Dichotomy:** Fast, creative, conversational LLMs prone to hallucination versus rigid, deterministic, change-resistant enterprise applications.

---

[01:36](https://www.youtube.com/watch?v=LN_xY92LLmE&t=96s) The High Stakes of Action: Who Watches the AI?

Highlighting data from MIT research, Bhasin notes that roughly 95% of enterprise AI pilots fail to advance into production. The danger escalates dramatically as agents move beyond merely generating text into executing real-world business actions; taking an incorrect enterprise action carries far higher financial and operational penalties than producing incorrect words. When considering how to govern these systems, standard "one size fits all" policies fail. More critically, attempting to solve the governance dilemma by simply placing another AI tool in charge of watching the first AI is "a trap" rather than a viable engineering design pattern.

```mermaid flowchart TD accTitle: Dual Brain AI Architecture Overview subgraph Brain1["Brain 1: The Conversationalist"] direction TB A["Intent Pipeline"] --> B["Planner Loop"] B --> C["Knowledge Graph Reasoning"] C --> D["Proposed Pipe Plan"] end

subgraph Gate["Validation & Human-in-the-Loop"] direction TB E["Inspection Verdicts & Citations"] F["Human-in-the-Loop Approval"] end

subgraph Brain2["Brain 2: The Operator"] direction TB G["Attribute & Action IAM"] G --> H["Real Data Boundaries"] H --> I["Structured Tool Calls (Schemas)"] end

D --> E E --> F F --> G ```

To illustrate model vulnerabilities, Bhasin references classic riddles and logic puzzles where frontier LLMs famously stumble—such as calculating whether to drive or walk to a gas station, or the traditional riddle about a surgeon unable to operate on their son. While these slip-ups spark entertaining Reddit threads, translating that ungrounded reasoning into production quotes, invoices, or critical manufacturing sales figures creates unacceptable risk. LLMs cannot be permitted to act purely from raw parametric memory or pre-trained data. Surveying the audience on how many currently run AI pilots versus how many have achieved production deployment, Bhasin emphasizes that true production readiness demands rigorous guardrails rather than blind faith in foundation models.

- **MIT Research Metric:** 95% of enterprise AI pilots never make it to production environments. - **The LLM Governance Anti-Pattern:** Using a second LLM to monitor the first without structural, deterministic boundaries. - **Consequence Shift:** The transition from conversational hallucination (text errors) to operational execution (wrong transactional actions).

---

[03:42](https://www.youtube.com/watch?v=LN_xY92LLmE&t=222s) The Dual-Brain Architecture and the Knowledge Fabric

To solve the governance challenge, Fabrion implements a dual-brain architecture designed to separate generative reasoning from operational execution. "Brain One" is the Conversationalist: it reasons over enterprise knowledge graphs, explores hypotheses, and interacts with users, but is strictly prohibited from touching operational systems without validated guardrails. "Brain Two" is the Operator: engineered to be "boring on purpose," it validates all requests against concrete identity, permissions, and system roles. Rather than relying on a block of prompt text "begging the model to behave," the operator enforces real boundaries and permission loops.

Brain One runs an intent pipeline grounded in the enterprise ontology, producing candidate plans, evidence frames, and inspection verdicts alongside a human-in-the-loop planner loop. Brain Two enforces attribute-level and agent-action-level permissions through Fabrion’s native Identity and Access Management (IAM) engine. It mandates structured tool calls with strict schemas, completely eliminating unvalidated free-text commands.

Both brains operate on top of a four-layer **Knowledge Fabric**: - **Identity:** Controls who may see and execute actions via enforced programmatic controls rather than system prompts. - **Data Layer:** Certified business measures tied to canonical enterprise sources with deterministic thresholds, providing an auditable paper trail and citation for every single number. - **Agent Layer:** Specialist agents constrained to specific operational lenses within the ontology to propose actions strictly within their designated lanes. - **Integration Layer:** Direct bi-directional integration with ERP, MES, data warehouses, identity providers, and business intelligence systems.

---

[06:34](https://www.youtube.com/watch?v=LN_xY92LLmE&t=394s) AI-Legible Ontologies for Complex Manufacturing

Modern industrial enterprises are far too intricate for static dashboards and too critical for ungrounded AI. In sectors like automotive manufacturing, a single vehicle model involves over 10,000 distinct parts, thousands of tier suppliers, assembly plants, warranty claim pipelines, bills of materials (BOM), and complex sales codes. Fabrion models these interconnected relationships into an AI-legible ontology.

```mermaid flowchart LR accTitle: AI Legible Manufacturing Ontology Model subgraph Core["Physical & Operational Entities"] V["Vehicle Model"] --- P["Assembly Plants"] V --- S["Suppliers"] V --- W["Warranty Claims"] end

subgraph DataFabric["Deterministic Knowledge Layer"] BOM["Bill of Materials / Parts Catalog"] KPI["Business Goals & KPIs"] SC["Sales Codes"] end

Core --- DataFabric DataFabric --> CQ["Certified Recorded Queries"] CQ --> AG["Monitored Agent Actions"] ```

Within this framework, every business object and relationship is grounded in certified, recorded queries originating from the underlying data fabric graph. Agents monitor and operate against this real-world operational graph with built-in data certification gates, eliminating the need for engineers to hand-craft validation logic from scratch. This allows conversational agents to "think in graphs" and navigate domain-specific business taxonomy while keeping enterprise operators firmly in control.

- **Manufacturing Scale:** Designed to handle vehicles and industrial assemblies spanning 10,000+ parts and thousands of global suppliers. - **Ontological Grounding:** Replaces static dashboards with dynamic graph representations linked to canonical parts catalogs, warranty records, and plant configurations.

---

[08:03](https://www.youtube.com/watch?v=LN_xY92LLmE&t=483s) The Open Meta-Model and Governance Lifecycle

Establishing an enterprise-wide guardrail layer requires an extensible, open-source approach based on industry standards rather than proprietary vendor silos. Fabrion structures its ontology around a core meta-model designed to ingest and normalize messy, fragmented enterprise data. Everything originates from a foundational "Concept" (such as a plant, factory, engineering change order, or KPI), which encapsulates four critical dimensions:

- **Action Types:** Permissible actions that human operators and autonomous agents can execute. - **Measures:** Numerical calculations, dimensions, sums, and aggregates bound to canonical data. - **Roles and Rules:** Precise security boundaries and conditional access controls. - **Analysis Directives:** Contextual business taxonomy explaining how data must be interpreted.

```mermaid flowchart TD accTitle: Open Meta Model to Deployment Architecture subgraph MetaModel["Core Meta-Model Concept"] AT["Action Types"] M["Measures"] RR["Roles & Rules"] AD["Analysis Directives"] end

MetaModel --> IndOnt["Industry Standard Ontology (e.g., Manufacturing)"] IndOnt --> Tenant["Tenant Instance (On-Premises Deployment)"] Tenant --> Engine["Model-Agnostic Execution Engine"] Engine --> Mod1["Frontier LLMs"] Engine --> Mod2["Fabrion Open-Weights Models"] ```

This meta-model maps onto standardized industry ontologies, which are then deployed inside the tenant’s private environment. Fabrion runs entirely on-premises or within the enterprise's sovereign infrastructure, ensuring sensitive operational data never leaves company walls. The execution engine remains completely model-agnostic, supporting frontier commercial LLMs as well as Fabrion-trained open-weights models.

The governance lifecycle strictly enforces four stages: **Declaration**, **Binding** to canonical sources, **Verification** against ontology guardrails, and final **Certification** by human operators. When an agent encounters uncertified data or ambiguous metrics, it explicitly outputs "I don't know yet" or provides a grounded fallback, establishing trust rather than fabricating answers.

---

[10:02](https://www.youtube.com/watch?v=LN_xY92LLmE&t=602s) Continuous Learning, Domain Skills, and Platform Tenets

Fabrion utilizes continuous learning backed by open-weights models to steadily enhance enterprise determinism. Every user query, decision, and manual human override is captured as an enterprise signal. Specialized "data agents" handle automated ingestion, enrichment, and cleansing pipelines. Through the **Fabrion Lab**, these signals are leveraged to fine-tune self-hosted open-weights models on internal enterprise taxonomy and operational workflows, continually evaluated against real agent actions and registered in a centralized agent registry. Over time, this feedback loop renders the overall platform increasingly deterministic.

```mermaid flowchart LR accTitle: Continuous Learning and Operational Skills Loop subgraph Feedback["Continuous Learning Loop"] direction TB Sig["Queries, Overrides & Exceptions"] --> DA["Data Agents (Ingest/Clean)"] DA --> Lab["Fabrion Lab (Fine-Tuning)"] Lab --> Reg["Agent Registry & Open Weights"] Reg --> Dual["Dual-Brain Platform"] Dual --> Sig end

subgraph Skills["Pre-Built Domain Skills"] direction TB S1["Parts & Bill of Materials"] S2["Warranty Claims Analysis"] S3["Plant Operations & S2P"] S4["Command Center Orchestration"] end

Dual -.-> Skills ```

On top of the ontology sits a modular **Skills Layer** offering pre-built domain manifests for manufacturing and automotive use cases, including Parts & BOM management, Warranty Claims analysis, Plant Operations, and Source-to-Pay (S2P) supply networks. These operational capabilities are managed through an enterprise **Command Center**, allowing business users to launch governed actions across their data fabric.

Bhasin concludes by reiterating Fabrion’s foundational architectural pillars: - **On-Premise Deployment:** Complete data sovereignty where internal assets never leave the client environment. - **Model Agnosticism:** Freedom to plug in external frontier models or self-hosted open-weights architectures with zero vendor lock-in. - **Guardrail-Driven Ontology:** A durable semantic asset that outlasts individual model releases.

The decisive question for enterprise leadership over the coming decade is not "which AI vendor should we pick," but rather: "How do we build an AI layer that adapts to our organization, gives us full control to accelerate our roadmap, and outlives any vendor we'll ever pick?"

## Notable quotes > Every enterprise AI platform today has one of two problems: either they put the LLM in charge (fast and creative, but hallucinates and can't be trusted with real enterprise actions) or they put rigid workflow tools in charge (reliable but inflexible, can't hold a conversation). — Kunal Bhasin

> The modern industrial enterprise is too complex for static dashboards and too critical for ungoverned AI. — Kunal Bhasin

> The enterprises that will dominate the next decade aren't asking 'which AI vendor should we pick', they're asking 'how do we build an AI layer that adapts to our organization and people, gives us full control to accelerate our AI roadmap, and is agnostic to every vendor we'll ever pick.' — Kunal Bhasin

> [01:36](https://www.youtube.com/watch?v=LN_xY92LLmE&t=96s) 95% of enterprise AI pilots fail to advance into production. — Kunal Bhasin

> [01:36](https://www.youtube.com/watch?v=LN_xY92LLmE&t=96s) taking an incorrect enterprise action carries far higher financial and operational penalties than producing incorrect words. — Kunal Bhasin

> [01:36](https://www.youtube.com/watch?v=LN_xY92LLmE&t=96s) attempting to solve the governance dilemma by simply placing another AI tool in charge of watching the first AI is 'a trap' rather than a viable engineering design pattern. — Kunal Bhasin

> [03:42](https://www.youtube.com/watch?v=LN_xY92LLmE&t=222s) Brain Two is the Operator: engineered to be 'boring on purpose,' it validates all requests against concrete identity, permissions, and system roles. — Kunal Bhasin

> [03:42](https://www.youtube.com/watch?v=LN_xY92LLmE&t=222s) Rather than relying on a block of prompt text 'begging the model to behave,' the operator enforces real boundaries and permission loops. — Kunal Bhasin

> [08:03](https://www.youtube.com/watch?v=LN_xY92LLmE&t=483s) When an agent encounters uncertified data or ambiguous metrics, it explicitly outputs 'I don't know yet' or provides a grounded fallback, establishing trust rather than fabricating answers. — Kunal Bhasin

## Mentioned Kunal Bhasin, Fabrion, 8VC, Ai4, MIT, Reddit, Fabrion Lab

## Video description Kunal Bhasin - Co-founder, CTO, CPO, Fabrion

Every enterprise AI platform today has one of two problems: either they put the LLM in charge (fast and creative, but hallucinates and can't be trusted with real enterprise actions) or they put rigid workflow tools in charge (reliable but inflexible, can't hold a conversation). Nobody has solved the governance problem: who watches the AI? Fabrion's dual-brain, full-stack AI platform answers this by separating conversational intelligence from operational execution; one brain reasons and proposes, the other validates and acts within real permissions, real data boundaries, and real human-in-the-loop controls, all connected through a full-stack knowledge-fabric of data, agents, integrations, and identity that plugs into the tools and systems your teams already use. The modern industrial enterprise is too complex for static dashboards and too critical for ungoverned AI. ‍We believe that manufacturing lines, supply chains, product structures, and compliance regimes are complex knowledge graphs, and AI that thinks in graphs and works with people inside guardrails will master them. The enterprises that will dominate the next decade aren't asking "which AI vendor should we pick", they're asking "how do we build an AI layer that adapts to our organization and people, gives us full control to accelerate our AI roadmap, and is agnostic to every vendor we'll ever pick.

[Source](https://www.youtube.com/watch?v=LN_xY92LLmE)

中文简报

Fabrion解耦AI推理与行动以防范操作失误让生成式模型直接控制事务型工作流导致95%的企业试点项目失败，这要求企业将对话智能与实施权限控制的确定性执行层分离开来。 Kunal Bhasin深入探讨了企业的核心困境：对话式LLM存在幻觉问题，而僵化的确定性系统又缺乏适应能力，这也导致95%的企业AI试点项目走向失败。Fabrion通过引入双脑架构解决了该难题，将探索性对话推理与具备审计能力且受IAM权限控制的操作执行引擎相分离。该平台立足于四层知识织构（Knowledge Fabric）与开放工业元模型，可将零部件目录至保修流程等复杂制造业本体映射为可验证的数据图谱。这一独立于具体模型且支持本地部署的系统，能够持续从人工反馈中学习并微调开源权重模型，从而保障企业数据主权并实现长期的供应商独立性。企业级 AI 面临核心两难困境：对话式大模型虽灵活敏捷却易出现幻觉，而传统企业系统虽然确定可审计却过于僵化且无法进行自然交互。 (0:04) MIT 研究表明约 95% 的企业 AI 试点未能成功投入生产；在缺乏确定性结构边界的情况下单纯用第二个大模型来监控主模型，属于工程反模式。 (1:36) Fabrion 提出“双脑”架构（Dual-Brain），将负责对话推理的 Brain 1 与负责严格 IAM 权限校验及结构化执行的 Brain 2 解耦，配合四层 Knowledge Fabric 实现全流程风控。 (3:42) 针对涉及上万种零部件与复杂供应链的工业制造场景，Fabrion 依托对 AI 可读的开放本体元模型，提供完全本地私有化部署并实现模型无关（Model-Agnostic）的敏捷接入。 (8:03)

Reading summary

Kunal Bhasin, co-founder of Fabrion, argues that enterprise artificial intelligence stalls because probabilistic language models cannot meet the rigorous compliance and deterministic requirements of mission-critical business systems. To bridge this divide, Fabrion proposes separating generative interaction from system execution through a dual-brain architecture underpinned by structured domain ontologies. Execution Risks and the Failure of Pilot Governance (1:36) According to research from MIT cited by Bhasin, approximately 95% of enterprise AI pilots fail to advance into production environments. The primary barrier is not conversational performance, but the transition from generating natural language to executing operational tasks. While a hallucinated figure in a chat interface is inconvenient, executing incorrect enterprise actions—such as modifying quotes, issuing invoices, or misrouting supply chain orders—creates severe financial, legal, and operational consequences. Bhasin contends that conventional enterprise guardrails often fail because organizations rely on an architectural anti-pattern: deploying a second model to monitor the first. Frontier models frequently stumble on basic logic and ungrounded parametric assumptions. Consequently, relying on another probabilistic model or embedding natural language prompts that instruct an agent to behave does not establish valid operational boundaries. Dual-Brain Architecture and Enforced System Boundaries (3:42) To prevent unauthorized actions, Fabrion decouples reasoning from transactional execution using a dual-brain architecture. 'Brain One' acts as the conversationalist, handling intent parsing, hypotheses, and knowledge graph exploration without direct access to core operations. 'Brain Two' serves as the operator, deliberately designed to be deterministic. It validates every proposed plan against identity and access management (IAM) permissions down to the individual attribute and agent-action levels. This operational framework relies on a four-layer Knowledge Fabric spanning identity access, semantic data, specialist agents, and direct enterprise integrations. Instead of passing unconstrained text prompts to application interfaces, the operator enforces rigid, schema-validated tool calls. Certified business measures are pinned directly to canonical enterprise sources such as enterprise resource planning (ERP) and manufacturing execution systems (MES), ensuring every metric maintains a verifiable audit trail. AI-Legible Ontologies and the Verification Lifecycle (8:03) Large industrial sectors, such as automotive manufacturing, manage interconnected physical and business entities encompassing over 10,000 distinct parts per vehicle, thousands of suppliers, complex warranty records, and sales codes. To make this complexity navigable, Fabrion maps data into an open meta-model concept defining four dimensions: permitted action types, canonical measures, access roles and rules, and contextual analysis directives. Data within this graph must advance through a strict four-stage governance cycle: declaration, binding to canonical enterprise records, verification against ontological guardrails, and explicit human certification. When an autonomous agent encounters ambiguous metrics or unverified data paths, the system forces a fallback or outputs that the information is unknown, eliminating fabricated assumptions in domain workflows. Continuous Adaptation Through Sovereign Feedback Loops (10:02) To enhance system reliability over time, Fabrion captures operational queries, user decisions, and human overrides as structured training signals. Specialized data agents clean and route this telemetry into internal model-tuning pipelines. Organizations use these operational signals to fine-tune self-hosted, open-weights models specifically on proprietary business taxonomies, systematically raising determinism across common domain skills like bill-of-materials management and plant logistics. Bhasin highlights deployment sovereignty as a core requirement, stating that sensitive manufacturing and enterprise data must remain on-premises or within private infrastructure. Because the execution engine is decoupled from underlying foundation models, enterprises can switch between proprietary frontier LLMs and self-hosted open-weights alternatives without rewriting business logic. Bhasin concludes that enterprise AI strategy over the coming decade should focus on establishing sovereign, model-agnostic semantic ontologies rather than selecting specific foundation model vendors. The key unresolved test for this approach is whether complex industrial enterprises can sustain the rigorous governance, data certification, and human oversight required to maintain these ontologies across dynamic operational environments.

精炼阅读摘要

Fabrion 联合创始人兼首席技术官 Kunal Bhasin 指出，大语言模型具备对话灵活性却易生幻觉，而传统企业系统可审计却高度僵化；为了推动 AI 从实验走向生产环境，企业必须构建将二者解耦并融合的双脑架构与语义本体底座。企业级 AI 落地瓶颈与执行风险 (0:04) Bhasin 引用 MIT 的研究数据指出，约 95% 的企业 AI 试点项目未能成功投入实际生产。在工业制造等复杂垂直领域中，大语言模型展现出极快的响应速度与自然语言交互能力，但存在凭空捏造供应商折扣或财务数据的风险。相比之下，传统的 ERP 和 MES 等遗留软件系统虽然严格确定、可审计且符合合规标准，却缺乏对话交互能力，使得响应新业务需求的周期往往长达数月乃至数个季度。当智能体从单纯的文本生成转向执行真实的业务操作时，系统性风险将显著加剧。产生错误的交易操作相比输出错误的文本，会给企业带来高得多的财务和运营损失。Bhasin 警告称，试图通过引入另一个 AI 模型来监督前一个 AI 是一种工程设计上的陷阱；前沿大模型在经典逻辑谜题上的频繁失误表明，不能允许企业级智能体仅依赖基础模型的原始参数记忆或未经校准的数据来触发生产指令。双脑分离架构与四层知识织网 (3:42) 为从根本上解决治理难题，Fabrion 提出了将推理与执行严格物理隔离的“双脑架构”。“脑一”作为对话者，负责理解用户意图、遍历企业知识图谱并推演假设，但被硬性禁止直接访问业务系统底层；“脑二”作为操作员，专为确定性设计，负责核验具体的身份凭证、权限边界与系统角色。该架构摒弃了单纯依靠 Prompt 约束模型行为的做法，强制要求通过具有严格 Schema 定义的结构化工具调用取代自由文本指令，并引入人在回路（Human-in-the-Loop）的审批闭环。该双脑体系运行在四层“知识织网”（Knowledge Fabric）之上：第一层是身份管理（IAM），在属性级别和智能体操作级别实现细粒度访问控制；第二层是语义数据层，将经过认证的业务指标绑定到规范数据源，保留完整的计算阈值与引用证据链；第三层是专业智能体层，限定各垂直智能体在指定的本体子图内提供建议；第四层是集成层，直接与企业的 ERP、MES、数据仓库和商业智能系统建立双向连接。复杂制造业的语义本体与治理生命周期 (6:34) 现代重工业制造涉及庞杂的实体网络，例如单款汽车即可包含超过 10,000 个零部件，并关联数千家供应商、总装工厂、保修索赔管道与销售代码。Fabrion 将这些物理实体与运营关系建模为 AI 可理解的图结构本体，使业务对象直接绑定到数据织网中经过认证的记录查询，免除工程师手动编写验证逻辑的负担，让对话系统能够在清晰的业务分类法中展开推理。为避免供应商锁定，该方案采用开放元模型设计，涵盖操作类型、度量指标、角色与规则以及分析指令四大核心维度。企业的治理生命周期严格执行“声明、绑定、验证和认证”四个阶段：指标必须绑定到规范源，经本体护栏校验，并由业务操作员最终认证。当智能体遇到未经认证的缺失数据或模糊指标时，系统会明确输出“我尚未知晓”或提供确定性的降级说明，杜绝虚假推理。持续学习机制与领域技能沉淀 (10:02) Fabrion 通过构建持续学习反馈回路来增强平台的确定性。平台将每一次用户查询、业务决策和人工覆盖操作转化为企业专属信号，利用专门的数据智能体自动完成摄取、清洗与数据富集。在 Fabrion Lab 环境中，这些数据被用于针对企业内部专业分类法和工作流程微调自托管的开放权重模型，经过实际智能体动作评估后纳入集中注册表，使系统运行随着时间推移愈发稳定可控。在此之上，系统提供模块化的技能层，内置面向汽车与制造业的预制领域清单，覆盖零部件及物料清单（BOM）管理、保修索赔分析、工厂运营以及寻源到付款（S2P）供应链网络。业务用户可通过企业指挥中心协同调度这些受到严格治理的操作能力。 Bhasin 总结认为，企业在未来十年面临的核心命题并非挑选哪一家具体的 AI 供应商，而是如何打造一个完全本地化部署、模型无关且由防御性语义本体驱动的基础架构，从而在确保数据主权与全权控制的同时，使企业级 AI 资产能够跨越底层基础模型的迭代周期长期发挥价值。

Detailed notes — zh

## Background 该内容由频道 Ai4 于 2026 年 8 月 15 日发布在 YouTube 平台，主讲人为 Fabrion 联合创始人兼 CTO/CPO Kunal Bhasin。视频简介指出当前企业 AI 普遍面临“失控幻觉的 LLM”与“无法对话的僵化工作流”两大割裂困境，并介绍了 Fabrion 如何通过解耦对话智能与操作执行的双脑架构及全栈知识织网，为复杂工业制造等企业打造安全受控、模型无关的 AI 平台。 ## TL;DR 演讲指出高达 95% 的企业 AI 试点因缺乏确定性与有效护栏而无法上线投产，并强调依赖“AI 监督 AI”是错误的治理陷阱。Fabrion 提出了创新的“双脑架构”，将负责推理提议的“对话脑”与严格执行权限及 Schema 校验的“操作脑”分离，并依托全栈知识织网（Knowledge Fabric）打通身份、数据语义与企业系统集成。针对工业制造等复杂业务，平台构建了标准化的核心元模型与数据认证闭环，在支持本地私有部署和模型无关的前提下，帮助企业构建安全、可控且摆脱单一供应商锁定的长期 AI 架构。 ## Atmosphere & style 演讲整体基调冷静务实、专业严谨且直击企业落地痛点，彻底摒弃了对大模型的盲目狂热与概念炒作。Kunal 通过现场互动与犀利的行业数据，生动勾勒出企业在 AI 试点跨越到生产级应用时的普遍困境与焦虑。整个分享逻辑层层递进，从架构缺陷的尖锐批判延伸到底层本体元模型的硬核拆解，展现出极强的工业实战技术底蕴与前瞻性的战略掌控力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=LN_xY92LLmE&t=4s) 当前企业 AI 部署存在断层：自信型 Agent 具备自然对话能力但容易产生幻觉，而传统企业级系统具备确定性与合规性却无法进行灵活交互。 - [01:36](https://www.youtube.com/watch?v=LN_xY92LLmE&t=96s) 根据 MIT 2026 年研究，高达 95% 的企业 AI 试点项目无法投产，核心瓶颈在于 Agent 执行错误操作的风险远高于单纯说错话。 - [01:36](https://www.youtube.com/watch?v=LN_xY92LLmE&t=96s) 试图通过“用 AI 监督 AI”或依赖长 Prompt 的治理模式是不可行的设计陷阱，无法为生产级业务提供可靠的确定性保障。 - [03:50](https://www.youtube.com/watch?v=LN_xY92LLmE&t=230s) Fabrion 提出双脑架构（Dual-Brain Architecture），将负责图谱推理的对话脑与负责权限、边界硬校验的操作脑完全解耦。 - [03:50](https://www.youtube.com/watch?v=LN_xY92LLmE&t=230s) 操作脑严禁自由文本即兴发挥，必须经由严格的 Schema 校验发起结构化工具调用，并嵌入预算执行和安全护栏。 - [03:50](https://www.youtube.com/watch?v=LN_xY92LLmE&t=230s) 底层全栈知识织网（Knowledge Fabric）贯穿身份访问层（IAM）、数据语义层、记忆与专家 Agent 层及 ERP/MES 等集成系统。 - [06:34](https://www.youtube.com/watch?v=LN_xY92LLmE&t=394s) 面向工业制造等复杂场景，AI 原生业务本体（Ontology）通过概念、动作类型、度量和规则等核心元模型将企业数据转化为图谱化知识。 - [09:34](https://www.youtube.com/watch?v=LN_xY92LLmE&t=574s) 数据治理遵循严格的“声明-绑定-验证-认证”生命周期，在指标未通过人工认证前 AI 绝不胡乱捏造，从而建立企业信任。 - [09:34](https://www.youtube.com/watch?v=LN_xY92LLmE&t=574s) 平台建立由数据 Agent、Fabrion Lab 本地微调和技能层构成的闭环系统，将用户决策与人工覆盖转化为持续迭代的异常信号数据。 - [11:28](https://www.youtube.com/watch?v=LN_xY92LLmE&t=688s) Fabrion 支持本地私有化部署且保持模型无关性（Model-agnostic），确保企业核心机密数据不出域并彻底规避单一供应商锁定风险。 - [11:28](https://www.youtube.com/watch?v=LN_xY92LLmE&t=688s) 企业 AI 的核心命题不在于挑选哪家具体供应商，而在于构建一套能深度适配组织、保持完全控制力且生命力超越单一厂商的底层架构。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=LN_xY92LLmE&t=4s) 企业 AI 的困境：对话能力与确定性系统的断层 Fabrion 联合创始人兼 CTO/CPO Kunal Bhasin 介绍了公司的使命：作为一家获得 8VC 投资的旧金山初创企业，Fabrion 致力于让 AI 真正赋能大型复杂企业，尤其是工业制造等复杂垂直领域。当前企业部署 AI 主要呈现出两种割裂的形态： 1. **“自信型 Agent”（Confident Agent）**：聪明敏捷、从不睡觉，但偶尔会捏造数字或供应商折扣。它的核心价值在于“对话能力”，因为人类本质上是对话型物种，对话是使用技术最自然的交互界面。 2. **传统企业级应用（Traditional Enterprise Applications）**：极具确定性（Deterministic）、可靠、具备审计追踪与合规能力，但完全无法进行自然对话。在传统系统中，每一个新提问都相当于一次复杂的系统变更（Change Event），创新迭代周期以月或季度计算，而非几天。 Fabrion 的核心理念是从底层出发，将非确定性的对话智能与确定性的企业系统融合在单一平台中。 ### [01:36](https://www.youtube.com/watch?v=LN_xY92LLmE&t=96s) 治理难题：“谁来监督 AI？” 根据 MIT 在 2026 年的研究数据，高达 **95% 的企业 AI 试点项目（AI Pilots）无法成功走入生产环境**。随着 Agent 正在从单纯的文字生成演进到实际执行操作（Actions），执行错误动作所带来的灾难性后果远甚于说错一句话。传统的“一刀切”治理模式注定失败，而企业面临的核心问题是：如何让 AI 跨越“写邮件”的初级阶段，真正进入生产环境接管业务？针对“谁来监督 AI（Who watches the AI?）”这一命题，Kunal 强调：**“如果你的答案是用另一个 AI 工具去监督，那就是一个陷阱，它根本不是一个合格的设计模式。”** 大语言模型（LLM）经常在逻辑与常识推理上翻车。例如经典的加油站路线决策（开车还是步行更合理），或是外科医生无法为儿子做手术的经典伦理谜题（医生是母亲还是父亲），LLM 往往会给出荒谬的答案。如果在企业环境中，模型凭借自身训练记忆随意回答销售额、发票或报价等关键业务问题，后果不堪设想。在演讲现场互动中，当 Kunal 询问有多少人正在进行 AI 试点并保持举手直到上线生产时，绝大多数举起的手都放下了，这直观印证了生产级 AI 护栏的匮乏。 ```mermaid flowchart TD accTitle: AI Pilot Failure vs Production Requirements subgraph S1["当前企业 AI 痛点"] A["LLM Agent 试图执行行动"] --> B["幻觉与推理缺陷 (基于自身训练记忆生成)"] B --> C["95% 试点项目无法投产 (高风险/缺乏确定性)"] end subgraph S2["错误治理陷阱"] D["AI 监督 AI (Prompt Begging)"] -.->|无效护栏| B end subgraph S3["生产级解决方案"] E["确定性本体与系统权限"] --> F["双脑架构 & 真实护栏"] F --> G["可靠上线生产环境"] end ``` ### [03:50](https://www.youtube.com/watch?v=LN_xY92LLmE&t=230s) 双脑架构与全栈知识织网（Knowledge Fabric）为了彻底解决治理与执行问题，Fabrion 提出了**双脑架构（Dual-Brain Architecture）**： * **脑 1：对话脑（The Conversationalist）**：基于知识图谱进行推理，充满好奇心且反应敏捷；但它绝不允许直接触碰任何生产数据或系统，所有行为必须经过验证。其前端由意图管线（Intent Pipeline）构成，锚定在业务本体（Ontology）上，并通过规划器循环（Planner Loop）始终保持人在回路（Human-in-the-loop）监督，输出管线计划、证据帧（Evidence Frames）与审查结论。 * **脑 2：操作脑（The Operator）**：刻意设计得极其“枯燥”，负责对照真实的企业身份权限、系统角色与数据边界进行硬性校验。它摒弃了“写一段长 Prompt 祈求模型老实听话”的做法，只允许发起具备严格 Schema 校验的结构化工具调用（Structured Tool Calls），杜绝自由文本即兴发挥，并在全平台各层嵌入预算执行与安全护栏。双脑系统构建在自底向上的**全栈知识织网（Full-Stack Knowledge Fabric）**之上，由四层核心能力组成： 1. **身份与访问管理层（Identity / IAM）**：权限细粒度下沉到具体属性级别和 Agent 动作级别，用系统强制执行权限而非依赖 Prompt 提示词。 2. **数据语义层（Data & Semantic Layer）**：由业务本体与数据本体组成，绑定权威标准数据源（Canonical Sources），并设有门禁阈值，确保每个数据指标都有据可查（Paper Trail）并附带引用来源。 3. **记忆与专家 Agent 层（Memory & Agents）**：结合业务上下文、向量化数据与数据目录，专职 Agent 仅在自身限定领域（Lane）内提供方案。 4. **集成层（Integrations）**：无缝对接 ERP、MES、数据仓库、IDP、BI 仪表盘及遗留数据湖等确定性工具。 ```mermaid flowchart TD accTitle: Fabrion Dual Brain Architecture subgraph UserInteraction["用户交互层"] UI["人类操作员 / 业务提问"] end subgraph DualBrain["Fabrion 双脑架构"] direction TB subgraph Brain1["脑 1: 对话脑 (The Conversationalist)"] B1["意图管线 + 知识图谱推理 (规划计划 / 证据帧)"] end subgraph Brain2["脑 2: 操作脑 (The Operator)"] B2["权限与边界硬校验 (结构化工具调用 / 预算拦截)"] end B1 -->|提议与校验请求| B2 B2 -->|执行结果反馈| B1 end subgraph KnowledgeFabric["全栈知识织网 (Knowledge Fabric)"] L1["1. 身份层: IAM 细粒度权限管控"] L2["2. 数据层: 权威数据源 + 认证指标 + 审计追踪"] L3["3. Agent 层: 领域专业 Agent 提议"] L4["4. 集成层: ERP / MES / 数据仓库 / BI 系统"] L1 --- L2 --- L3 --- L4 end UI <--> B1 B2 <--> KnowledgeFabric ``` ### [06:34](https://www.youtube.com/watch?v=LN_xY92LLmE&t=394s) 面向工业制造的 AI 原生本体（Ontology）现代大型工业企业拥有极高的复杂度，例如一辆汽车包含 10,000 多个零部件并涉及数千家供应商，静态仪表盘无法满足运营需求。AI 必须学会以图谱和业务分类学（Business Taxonomy）的方式思考。 Fabrion 在平台内部内置了针对复杂对象的认证查询机制。以汽车制造本体为例： - **核心实体**：车辆型号（Vehicle Model）、组装工厂（Plant）、供应商（Suppliers）、保修索赔（Warranty Claims）。 - **确定性关联数据**：零部件目录（Parts Catalog）、物料清单（BOM Catalog）、KPI、业务目标及销售代码。 ```mermaid flowchart LR accTitle: Automotive AI-Legible Ontology Example VM["车辆型号 (Vehicle Model)"] -->|"组装于"| PL["制造工厂 (Plant)"] VM -->|"供应自"| SP["供应商 (Suppliers)"] VM -->|"关联"| WC["保修索赔 (Warranty Claims)"] subgraph DeterministicData["确定性数据目录"] BOM["BOM 清单"] PARTS["零部件目录"] KPI["业务 KPI & 销售代码"] end DeterministicData -.->|"认证指标与规则绑定"| VM ``` 为了让混乱的企业内部数据有序落地，Fabrion 推行开源且遵循行业标准的**核心元模型（Core Meta-Model）**，包含四大要素： - **概念（Concept）**：业务对象载体（如工厂、工程变更单 ECO、KPI 等）。 - **动作类型（Action Types）**：人类和 Agent 可执行的操作。 - **度量（Measures）**：聚合、求和、维度等数据指标定义。 - **角色与规则（Roles & Rules）**：用于下发权限与安全边界。 - **分析指令（Analysis Directives）**：定义数据背后的业务上下文与分类学。该元模型映射到行业标准本体（如制造行业），再适配到具体租户系统。Fabrion 平台直接私有化部署在企业内部，**数据永不离开客户环境**；同时系统具备模型无关性（Model Agnostic），既可对接顶尖闭源前沿模型，也能接入基于开源权重微调的专用模型。 ### [09:34](https://www.youtube.com/watch?v=LN_xY92LLmE&t=574s) 持续学习闭环、技能层与指挥中心数据治理遵循严格的生命周期流程：**声明（Declaration） $\to$ 绑定权威源（Binding） $\to$ 规则验证（Verification） $\to$ 人工认证（Human Certification）**。在数据未完成认证前，AI 不会胡编乱造，而是明确说明“暂未掌握该数据”以建立企业信任。生产运行中的每一次用户提问、系统决策和人工覆盖操作（Override），都会转化为企业内部的信号异常数据。这些数据被用于支撑全闭环的持续优化系统： - **数据 Agent（Data Agents）**：专门处理数据摄取（Ingestion）、富化（Enrichment）与数据清洗（Cleansing）。 - **Fabrion Lab**：在企业内部针对私有数据、流程和业务术语对开源权重模型进行微调与训练，并在每次 Agent 操作中持续评估。 - **Agent 注册表（Agent Registry）**：管理企业自托管的各类机器学习模型。 - **技能层（Skills Layer）**：预构建了免二次开发的工业技能清单，覆盖零部件与 BOM、保修索赔、工厂运营、供应网络及采购到付款（Source-to-Pay）。 - **指挥中心（Command Center）**：提供统一的业务本体操作界面，允许管理人员直接对全局数据与业务本体下达执行动作。 - **95% 的试点失败率**：援引 MIT 2026 年研究，绝大多数企业 AI 方案卡在落地门槛之外。 - **严格架构规范**：操作脑只执行具备明确 Schema 约束的结构化工具调用，禁止一切无格式自由文本交互。 - **本地部署与隐私**：Fabrion 平台完全运行于客户边界内，保障核心制造机密数据不出域。 ### [11:28](https://www.youtube.com/watch?v=LN_xY92LLmE&t=688s) Fabrion 的核心原则与面向未来的企业 AI 战略在演讲的最后，Kunal Bhasin 总结了 Fabrion 平台立足的核心原则（Core Tenets）。首先，在数据安全与主权方面，Fabrion 支持本地化私有部署（On-premise deployment），确保企业敏感数据永远不会流出自身的组织边界。其次，平台在架构上保持完全的模型无关性（Model-agnostic），并且所有能力均构建在基于护栏的本体模型（Guardrails-based ontology）之上。这种本体结构不仅是一套技术方案，更是企业构建未来所有 AI 应用的一项长期战略资产，能够以完全即插即用（Plug-and-play）的方式运作，彻底消除未来的供应商锁定风险。基于这些核心原则，演讲最后向所有正在探索智能化转型的企业提出了一个发人深省的核心命题：当今企业面临的关键问题根本不是“我们究竟该选择哪一家 AI 供应商”，而是一个更加关乎长远生存与竞争力的命题——“你该如何构建一套既能深度契合自身组织运作、又能让你时刻掌握完全控制权，并且其生命力能够超越你所选择的任何单一供应商的 AI 体系架构？”（"How do you build AI that adapts to your organization, keeps you in control, and outlives any vendor you'll ever pick?"）。 - **数据绝对可控**：支持本地化（On-premise）部署，保障企业数据资产与隐私不外流。 - **技术解耦与灵活性**：保持模型无关性（Model Agnostic）与完全即插即用（Plug-and-play）特性，实现真正的零供应商锁定（Zero lock-in）。 - **长期价值沉淀**：基于安全护栏的本体系统（Guardrails-based ontology）为企业沉淀可复用、可持续演进的底层 AI 能力资产。 ## Notable quotes > Every enterprise AI platform today has one of two problems: either they put the LLM in charge... or they put rigid workflow tools in charge. — Kunal Bhasin > Nobody has solved the governance problem: who watches the AI? — Kunal Bhasin > [01:36](https://www.youtube.com/watch?v=LN_xY92LLmE&t=96s) 如果你的答案是用另一个 AI 工具去监督，那就是一个陷阱，它根本不是一个合格的设计模式。 — Kunal Bhasin > The modern industrial enterprise is too complex for static dashboards and too critical for ungoverned AI. — Kunal Bhasin > We believe that manufacturing lines, supply chains, product structures, and compliance regimes are complex knowledge graphs, and AI that thinks in graphs and works with people inside guardrails will master them. — Kunal Bhasin > The enterprises that will dominate the next decade aren't asking 'which AI vendor should we pick', they're asking 'how do we build an AI layer that adapts to our organization and people, gives us full control to accelerate our AI roadmap, and is agnostic to every vendor we'll ever pick.' — Kunal Bhasin > [11:28](https://www.youtube.com/watch?v=LN_xY92LLmE&t=688s) How do you build AI that adapts to your organization, keeps you in control, and outlives any vendor you'll ever pick? — Kunal Bhasin > [01:36](https://www.youtube.com/watch?v=LN_xY92LLmE&t=96s) 95% 的企业 AI 试点项目（AI Pilots）无法成功走入生产环境。 — Kunal Bhasin ## Mentioned Kunal Bhasin, Fabrion, 8VC, MIT, Ai4, Fabrion Lab, Command Center, Agent Registry ## Video description Kunal Bhasin - Co-founder, CTO, CPO, Fabrion Every enterprise AI platform today has one of two problems: either they put the LLM in charge (fast and creative, but hallucinates and can't be trusted with real enterprise actions) or they put rigid workflow tools in charge (reliable but inflexible, can't hold a conversation). Nobody has solved the governance problem: who watches the AI? Fabrion's dual-brain, full-stack AI platform answers this by separating conversational intelligence from operational execution; one brain reasons and proposes, the other validates and acts within real permissions, real data boundaries, and real human-in-the-loop controls, all connected through a full-stack knowledge-fabric of data, agents, integrations, and identity that plugs into the tools and systems your teams already use. The modern industrial enterprise is too complex for static dashboards and too critical for ungoverned AI. ‍We believe that manufacturing lines, supply chains, product structures, and compliance regimes are complex knowledge graphs, and AI that thinks in graphs and works with people inside guardrails will master them. The enterprises that will dominate the next decade aren't asking "which AI vendor should we pick", they're asking "how do we build an AI layer that adapts to our organization and people, gives us full control to accelerate our AI roadmap, and is agnostic to every vendor we'll ever pick. [Source](https://www.youtube.com/watch?v=LN_xY92LLmE)

Source quotes

Kunal Bhasin: Every enterprise AI platform today has one of two problems: either they put the LLM in charge... or they put rigid workflow tools in charge. Kunal Bhasin: Nobody has solved the governance problem: who watches the AI? [1:36] Kunal Bhasin: 如果你的答案是用另一个 AI 工具去监督，那就是一个陷阱，它根本不是一个合格的设计模式。 Kunal Bhasin: The modern industrial enterprise is too complex for static dashboards and too critical for ungoverned AI. Kunal Bhasin: We believe that manufacturing lines, supply chains, product structures, and compliance regimes are complex knowledge graphs, and AI that thinks in graphs and works with people inside guardrails will master them. Kunal Bhasin: The enterprises that will dominate the next decade aren't asking 'which AI vendor should we pick', they're asking 'how do we build an AI layer that adapts to our organization and people, gives us full control to accelerate our AI roadmap, and is agnostic to every vendor we'll ever pick.' [11:28] Kunal Bhasin: How do you build AI that adapts to your organization, keeps you in control, and outlives any vendor you'll ever pick? [1:36] Kunal Bhasin: 95% 的企业 AI 试点项目（AI Pilots）无法成功走入生产环境。

Source quotes

Kunal Bhasin: Every enterprise AI platform today has one of two problems: either they put the LLM in charge (fast and creative, but hallucinates and can't be trusted with real enterprise actions) or they put rigid workflow tools in charge (reliable but inflexible, can't hold a conversation). Kunal Bhasin: The modern industrial enterprise is too complex for static dashboards and too critical for ungoverned AI. Kunal Bhasin: The enterprises that will dominate the next decade aren't asking 'which AI vendor should we pick', they're asking 'how do we build an AI layer that adapts to our organization and people, gives us full control to accelerate our AI roadmap, and is agnostic to every vendor we'll ever pick.' [1:36] Kunal Bhasin: 95% of enterprise AI pilots fail to advance into production. [1:36] Kunal Bhasin: taking an incorrect enterprise action carries far higher financial and operational penalties than producing incorrect words. [1:36] Kunal Bhasin: attempting to solve the governance dilemma by simply placing another AI tool in charge of watching the first AI is 'a trap' rather than a viable engineering design pattern. [3:42] Kunal Bhasin: Brain Two is the Operator: engineered to be 'boring on purpose,' it validates all requests against concrete identity, permissions, and system roles. [3:42] Kunal Bhasin: Rather than relying on a block of prompt text 'begging the model to behave,' the operator enforces real boundaries and permission loops. [8:03] Kunal Bhasin: When an agent encounters uncertified data or ambiguous metrics, it explicitly outputs 'I don't know yet' or provides a grounded fallback, establishing trust rather than fabricating answers.

Original transcript

[0:04] My name is Kunal. I'm a co-founder and CTO at Fabriqon. We are a San Francisco-based startup [0:09] backed by 8 VC um and our mission is to uh really make AI accessible for large complex [0:17] organizations and enterprises uh specifically in complex verticals like industrial manufacturing. [0:23] And I appreciate everybody attending today. So, uh let me start with um [0:31] this the state of enterprise AI today. And in most enterprises, if you look at how AI [0:36] is being deployed, it's being deployed in two flavors. Uh one, I call it the confident agent, [0:41] brilliant, fast, never sleeps, occasionally invents numbers or supplier discount, uh but it's [0:48] conversational. That's the thing that's really unlocked AI is uh humans are conversational beings [0:54] and it's a very accessible interface uh to technology is having a conversation. [1:00] Um on the other hand, uh you have your traditional uh enterprise applications um which are very [1:06] deterministic in nature, uh reliable, auditable, compliant, but cannot hold a conversation. And [1:12] every question is a change event and innovation is measured in quarters or months, not days. And what [1:20] we're doing at Fabriqon is really uh thinking from the bottom up, how do you combine these two um [1:28] uh deterministic and non-deterministic uh nature of applications uh in one platform. [1:36] So, uh let's look at some real numbers. Um in uh in 2026, uh per uh MIT research, 95% of all um [1:47] uh AI pilots don't go into production. Um agents are moving beyond hallucination into actions and [1:55] taking a wrong action has much higher consequences than just wrong words. Oh, and as far as the [2:00] governance model, we know that one size fits all doesn't work. That's that's a recipe for failure. [2:05] And the big question is, how do you upgrade AI from just writing emails to actually taking [2:11] actions in production? So, the big question is who watches the AI? And if your answer is um [2:20] another AI tool, that's a trap. Uh it's not a design pattern. [2:24] And over here I have a couple of examples of some well-known examples, you know, [2:30] that where LLMs uh get it wrong. One of them is where you want to drive to a gas station and um [2:38] and you know, you you better you drive or you walk. The other one is uh you're asking, you know, [2:45] as a surgeon that you cannot do a surgery on your son, whether you're a mother or father. [2:49] It's a riddle. And um most LLMs have gotten this wrong. At some point there are long Reddit threads [2:55] on this. And if you uh really convert that in the enterprise to sales numbers or invoices or quotes, [3:03] the consequences are really, really high. So, you don't [clears throat] want the LLMs responding [3:06] from their memory or trained data. And what this shows you is the LLMs and the models are only [3:11] as good as their trained data. Um and the big question becomes um who watches the AI? Right? [3:20] Like how many people in the room actually have an AI pilot running today? Can you raise your hand [3:26] in your organization? And keep your hand up if that pilot's in production. [3:32] You know, very few. So, that's the big question we're trying to answer at at Fabriqon with a lot [3:37] of research is how do you watch the AI and how do you set the right guardrails [3:42] um that work in production. So, we've come up with our own architecture, which we call the dual brain [3:50] architecture. Brain one is the conversationalist. it reasons over your knowledge graphs, [3:54] it's brilliant, it's curious, but it cannot touch anything without validation and guardrails. And [4:01] brain two is the operator. It's boring on purpose, it validates everything against real permissions, [4:07] your actual identity, your system roles. It's not just a paragraph or a page of prompt begging [4:15] the model to behave. They're real guardrails, real boundaries, and real permission loops. [4:20] And um the conversational brain starts with the intent pipeline. It's grounded in what we call [4:26] your ontology, and it has a planner loop which always has human in the loop for guardrails. [4:32] So, it has pipe plans, evidence frames, inspection verdicts, and the execution brain is [4:37] gated by your permission system. We have our own identity and access management system [4:42] that sets permissions down to every attribute level and agent action level. [4:47] There are real data boundaries that are compiled into your ontology, and we and the system only [4:53] allows you to make a structured tool calls with real schemas, no free text, no improvisation. [4:58] And the guardrails and budget enforcement is built in every layer of the product. And both [5:03] of these brains sit on top of what we call our knowledge fabric which starts with the identity, [5:08] which is IAM, identity and access management, the semantic layer, which is your business ontology, [5:14] data ontology sitting on top of human measures. And memory, which is your business context along [5:21] with vectorized data and a data catalog, and then integrations where you can connect to your ERPs, [5:27] MES, legacy data lakes, and any data or application source within the organization. [5:35] So, connected by our industry standard ontologies which I'm going to talk about more in a minute, [5:41] the full stack knowledge fabric has four layers. Number one, gated by identity, everything happens [5:47] to permissions. So, who may see and do what enforced rather than prompted. The next layer [5:52] is the data, which is certified measures bound to actual canonical sources. And these measures and [5:59] are are gated by thresholds that are actually compiled into your ontology into your data layer. [6:06] So, every user doesn't have to worry about is the measure right? Is the data deterministic? [6:11] So, every number has a paper trail and a citation. The next layer is agents, which are specialists [6:17] that propose within their lane. So, you can think of an agent working on a specific lens on your [6:22] ontology or use case on your ontology. And lastly, there's integrations that plug into existing [6:27] deterministic tools, your data warehouses, IDPs, business intelligence dashboards, and MES systems, [6:34] and so on. So, here's a example of what an AI-legible ontology would look like. We're, [6:40] you know, big in the manufacturing space, automotive space, where you're building [6:45] very complex products like a vehicle which has 10,000 plus parts and thousands of suppliers. [6:50] And what would the ontology look like? So, at the center of it is your vehicle model, [6:54] assembled at a plant, supplied by some suppliers, [6:58] claimed on warranties. And then you have data, which is which is deterministic, which is your [7:05] parts catalog, bill of materials catalog, KPIs, business goals, sales codes, and all of these [7:12] ontology objects are gated in and certified in Fabriq on. So, there's a certified recorded query [7:20] starting with the data data fabric in the graphs. And on that data fabric and graphs, [7:25] your agents can monitor and act on your real-world ontology because the gate the gates [7:31] and the certified certification of these measures is built into the platform. So, you don't have to [7:35] hand-build it from scratch. And then on top of that, we have the operational control. So, [7:40] we feel like most complex organizations are too complex for static dashboards and to get the real [7:46] value out of the conversational agents, um you need uh agent that can think in graphs, that can [7:52] think in your business taxonomy, and works with people with guardrails inside your organization. [7:58] So, the next question is, how do you even start and think about building an ontology, right? And [8:03] and having this grounded layer, guardrails layer that your agents and AI can work on. So, the way [8:08] we've innovated at Fabriq AI on is we think it should be open source and it should be grounded in [8:13] industry standards. And the way you make that happen is you need a core meta model that can [8:18] adapt to your organization and and your different fragmented data sources and most data in big [8:23] organizations is messy. So, the meta model helps ground it in uh in an object model that scales and [8:29] it all starts with what we call a concept. So, a plant, a factory, uh uh engineering change order, [8:35] uh uh a KPI are all concepts. And each concept, because we've built this ontology to be AI native, [8:40] has um action types, which is actions that that humans and agents can take, [8:46] measures, which is what does the data look like, aggregates, sums, you know, uh dimensions, [8:53] uh roles, which is how you roles and rules, which is how you apply permissions, [8:57] and an analysis directive, which is how do you analyze the data, what is the context behind [9:02] the data, what is the business taxonomy behind the data. Um and then we map our core meta model onto [9:07] what we call an industry standard um ontology. In this case, we're showing a manufacturing ontology, [9:13] uh which is parts, plants, dealers, vehicles, and then we map that industry ontology to a tenant, [9:19] which is uh you know, a customer. And Fabriq AI deploys on on as a platform within your [9:24] organization, so none of the data ever leaves your organization. Um and we are model agnostic, [9:29] so you can connect to any frontier model or a Fabriq AI trained open weights model. [9:34] So, what govern actually looks like, you start with uh declaration and then you bind [9:39] it to a canonical source, verified through the guardrails that are built into the ontology, [9:43] and then finally certified by a human before it goes live into a production production system. So, [9:51] and when the data isn't certified yet, AI that says I don't know yet or gives you a [9:55] grounded answer when it doesn't know something builds trust versus making it up from memory. [10:02] So, for us the answer is open weights and continuous learning. Uh your and every query, [10:08] decision, overrides becomes an exception and is a signal and enterprise signal in the system, [10:14] which works with data agents. And data agents are are trained just to work with data operations like [10:20] ingestion, enrichment, cleansing. And then all all of that is connected to our Fabri on lab, [10:26] where we train and tune open weights models on your internal data, on your internal processes, [10:31] uh on your internal taxonomy, and then evaluate that with every agent action. And then there's an [10:36] agent registry that hosts self-hosted models and self-hosted, you know, machine learning models. [10:41] And our dual brain uh platform uh works with you what what should be conversational and what [10:46] should be deterministic. And the whole thing is a continuous loop, which is constantly improving. [10:50] So, with every action, every prompt, uh your your system is getting more deterministic uh [10:55] even though it's using AI. And then I also want to talk about on top of the ontology sits our skills [11:00] layer, and skills are manifest that don't need uh for you to deploy and build uh applications. [11:06] We have some skills that we built from the ground up for the manufacturing and automotive industry, [11:10] parts and bill of materials, uh warranty claims, plant operations, uh supply networks, source to [11:16] pay. Um and then we also have a business ontology layer called the command center, where you can [11:20] launch actions on top of your uh data and business ontology uh within the organization. So, I really [11:28] want to end with like what are the core tenants tenants of Fabri on, which is uh we are deployed [11:33] on on on premise so your data never leaves your organization. We are model agnostic. Everything [11:38] is built on a guardrails based ontology. That's That's a long-term investment for your company [11:43] to build AI applications and we have full plug and play with zero lock-in in the future. So the big [11:48] question is, you know, not which vendor should you pick for AI? The big question is, how do you build [11:54] AI that adapts to your organization, keeps you in control, and outlives any vendor you'll ever pick?