To prepare for the continued expansion of profitable AI, Chief Data Officers (CDOs) must realize that a fundamental architectural shift is taking place: the move from the data fabric to the AI fabric. The data fabric served its purpose for human-consumed analytics, but as enterprise operations transition toward autonomous, machine-speed execution, older frameworks are reaching their breaking point. CDOs that begin to make this shift now will outpace their competitors on the journey toward the autonomous enterprise.

From Human Consumers to Agent Consumers

The data fabric provides unified, metadata-driven access across data engineering, governance, and analytics. However, the data fabric was built primarily to curate and govern structured data for analytics and human decision making. Autonomous agents now utilize insight from databases, documents, and text. The data fabric treated structured and unstructured data separately. However, operational AI requires the unification of data engineering and data governance for all data.

The good news is that the metadata from the data fabric builds a useful foundation for the AI fabric in three ways. One, it becomes the blueprint by which even more data engineering and analytical insight can be automated using AI. Two, it feeds the semantic layer enabling conversational AI to deliver curated insight without having to touch the data. Three, it provides the basis on which unstructured data can interact with structured data insight. The result is that in the shift to the AI fabric there will be an increase in the level of automation at the data layer, until the data layer is fully or quasi-fully automated. Ultimately, structured and unstructured data will be unified and AI enabled for constant adaptation to new business models and new AI requirements, especially for agentic AI.

From Unified Data to Universal Realtime Context

Achieving deep data engineering autonomy also requires solving an architectural friction point within today’s enterprise technology stack. The closest we have come to unified data engineering and unified data governance is looking at structured data engineering vendors who have purchased unstructured data engineering vendors, who are now working to unify the two into a single offering. The challenge is that the requirements for structured versus unstructured data are vastly different. When you add in the real time requirement for agentic AI, integration becomes even more difficult to merge the two.

The secret will be at the metadata layer, where a matrix of metadata is programmed to understand the intersection between structured and unstructured data. For example, hidden in C-level monthly and quarterly reports is the key to understanding causal relationships within a specific organization, and every organization is unique. When causality can be mapped and set as an overlay above the structured data, suddenly the conversational AI layer can make more strategic sense of the two sets of data combined.

For agentic AI, the necessity of real time context adds more complexity to AI control plane success. Historically, active metadata catalogs did not require low latency. General business context was enough to guide analytics. However, as we move into the agentic age where thousands of AI agents are acting on behalf of the organization, not just writing code, the demand for real time context and insight becomes even more critical. When agents run the business operations, real time business, data, and technical context are all mandatory. The need for active metadata becomes the need for real time active metadata; the need for context becomes the need for universal context.

From AI Fabric Control Plane to Causal Intelligence

To orchestrate this real time environment across a complex enterprise, organizations require a dedicated control plane. There are three requirements for the AI Fabric Control Plane that are accentuated in the move to agentic AI, and to address the fast approaching causal intelligence and causal AI. One, composability. The depth of integration for all components managed within the control plane must go beyond connectivity to plug and play. Two, interoperability. All aspects of each composable part must be fully interoperable with all other related composable parts. Three, adaptability. Since each composable part will be constantly changing and growing, the control plane must be AI enabled to automatically adjust for continuous interoperability.

This adaptable foundation is critical because enterprise decision making can no longer rely on simple correlation. In the business world, “correlation” is the analytical equivalent of the leader’s “best guess.” It doesn’t have the accuracy necessary for leaders to continue making strategic decisions for their new world operations that now includes the familiar “people” and the growth of AI agents.

Causal intelligence is the next step in equipping leaders to make strategic decisions and to give leaders the confidence they need to guide their new armies of agents that work 24/7 and transact at the speed of a machine. While it is possible to use probabilistic AI to discover and document cause and effect within the enterprise, the control plane must support the engineering of causal maps that are confirmed by human analysis along with the support of AI agents, then hard coded into the control plane.

From AI Governance to Algorithmic Business Assurance (ABA)

In the Ferraro Consulting POV Paper, The AI Governance Manifesto, I laid out how both generative AI and agentic AI inevitably break down the guardrails set up by traditional governance tools. Enterprise AI is failing its first major operational stress test. Not because the technology is too smart, but because the solutions designed to “govern” it are inadequate.

In addition, because AI agents act on behalf of people and organizations, AI governance cannot belong to an isolated compliance committee or an IT security silo. True AI governance must be a real time business discipline.

From Retrospective Paperwork to Deterministic Execution Proofs: We must stop treating model cards and compliance reports as actual safety. They are static artifacts, outdated the moment they are compiled. True assurance requires deterministic runtime testing of probabilistic actions. Tests must be automatically generated at runtime and dynamic throughout their life. If a model or an autonomous agent cannot mathematically prove it operated within explicit behavioral and regulatory boundaries during a transaction, the execution is dynamically invalidated before it can cause any harm.

To govern highly autonomous multiagent environments, we must move beyond outdated perimeter guardrails by embedding governance natively within the execution engine, infusing trust from the inside out. In addition, the AI control plane must apply continuous governance, actively monitoring execution intent and reasoning divergence. This necessitates real-time financial risk intervention, where every agentic decision is assigned an instantaneous economic liability score, allowing the system to dynamically restrict or reroute high-risk actions before they occur, effectively shifting from retrospective auditing to proactive, deterministic execution control.

The ultimate owner of the control plane must be the business unit leader carrying the P&L. Policies, guardrails, and risk tolerances must be managed in pure, natural business intent language, giving nontechnical executives decision power over the algorithms and the incidents. Without business ownership, agents cannot have true accountability.

Accelerating the AI Value Pyramid

When anchored by real time metadata, hardcoded causality, and real time governance, the AI fabric with its AI Fabric Control Plane builds a pyramid of value. At the data layer, the current 25% autonomous will grow to 50% over the next 1 to 2 years [1]. Because the data becomes autonomous, the AI layer will grow at a similar pace over the next 1 to 3 years, from 10% autonomous to 40% [2]. The real value multiplier comes when the business realizes the gains from data layer and AI layer autonomy [3]. The ability to launch new AI enabled business models and adapt quickly to changing market conditions will be the hallmark of leaders.

The way to become a truly autonomous enterprise, is to continue growing the autonomy of your data and AI layers, and moving to the AI control plane. As you do, you will outpace your competitors in the speed of AI expansion, the agility of business adaptability, the accuracy of AI enabled business operations. Ultimately, you will pave the way for the use of causal intelligence to make strategic decisions at the senior executive level.

The Leader’s Growth Mindset

To lead this transformation and become an AI first enterprise, the CDO needs to make two shifts: an architectural shift and a design shift.

The architectural shift focuses on an AI control plane that unifies platforms across the entire data and AI estate. Abandon old paradigms built around single platforms and opt for a metadata centric architecture designed for deep integration and disambiguation of competing definitions within your diverse platforms. For example, if your AI control plane can extract signals from your operational streams, disambiguate any conflicting definitions, and provide a rich semantic context, you may not need a data warehouse.

The design shift focuses on a more human-centric approach to AI. Business requirements are inconspicuously hidden in emails, documents, and presentations. The conversations that take place in the board room and around tables are accompanied by searchable artifacts. Your human-centric approach spans two arenas: One, search where people have communicated before to find the hidden treasure of causality. Two, promote enterprise wide use of AI and use the logs as an even better guide to what humans need to better do their jobs, from the CEO all the way out to individual contributors.

Three Red Flags

As CDOs execute these shifts, they must hold AI software vendors to an entirely new standard. There are three red flags. One, watch out for marketecture from vendors who craft a story of automation and agentic, then require you to build out the automations and agents in order to make use of them in data engineering, AI applications, and AI agent operations. Two, beware of AI without a brain from vendors who lack the maturity in their semantic layer to support the complexities of your enterprise context. Without the merging of technical and business metadata with dynamic, real time interactivity, your AI expansion will stall. Three, take time to calculate “automation value”, and compare vendors based on total automation value. Ask a simple question, “Tell me everything in your platform or system that is fully automated.” Then quantify the value of every automation. Vendors who have slapped AI on top of their antique software will come out at the bottom of the list every time. Leading vendors will have a measurable roadmap of additional automations and AI agents.

CDOs have an opportunity to help your organizations move from data fabric architectures to an AI fabric control plane in preparation for the coming agent proliferation. Rather than seeing the move as purely architectural, think of it as the means by which you will guarantee a quicker return on your AI investments and safer guardrails for potential AI risks. Your initial returns will come from the speed of AI deployment, the ability to fine tune AI autonomy, and the precision at which you can operate.

NOTE TO AI VENDORS: The data layer is being commoditized. Even though enterprise buyers continue to spend on efficient data engineering, the funds are shifting toward enabling the autonomous enterprise. As investments shift, so will the expectations of a measurable and growing return on AI investments. The time is ripe to make a shift toward the AI fabric. Your technology will no longer be measured by data efficiencies; you will be compared to other vendors based on your automation value and business value creation. The most important shift for your messaging will be toward business control over both risk and value creation. In this regard, the AI fabric does what the data fabric could never do.

Subscribe now

[1] Gartner forecasts that up to 60% of data and analytics leaders will automate core operational data tasks over the next 2–3 years. Gartner Data & Analytics Trends

[2] Gartner predicts that 40% of enterprise applications will embed task-specific AI agents by the end of 2026 (up from under 5% in 2025). Gartner Data & Analytics Trends

[3] S&P Global and McKinsey research reports ~31% of enterprise organizations running AI agents in production workloads. Gartner & McKinsey AI Agent Statistics 2026

The AI governance market has devolved into an expensive exercise in corporate and startup theater. Organizations are spending millions on software that promises protection but requires massive development or slow human intervention to make it operational. It is time to call out the rebranded data governance platforms, the partially operational startups, and the static frameworks for what they actually are: insufficient attempts to govern AI in a fast-changing, real time, non-deterministic world.

The Three Scandals of the Incumbent Market

1. The Rebranded Data Governance Hustle

Yesterday they tracked SQL tables and GDPR compliance; today, they slapped an “AI” sticker on the box and now claim they can manage autonomous multi-agent systems. Tracking data inputs and outputs is data governance. AI governance is the management of emergent behavior, probabilistic risk, model drift, and execution intent. You can’t fix an algorithmic intent shift by auditing data.

2. The Paperwork Factory

Nations and well-meaning associations are pumping out new AI governance frameworks as fast as ChatGPT can produce new content. Legacy GRC platforms have built automated workflow systems to generate compliance documentation, model cards, and system cards. They function like cloud-based spreadsheets, relying on human-filled intake and update. They flag incidents and measure risk, but they were not built for the operational requirements of AI governance. Paperwork and process recommendations will never provide the evidence necessary to operationalize data governance.

3. The Peripheral Illusion

The current market relies on API proxies, AI gateways, webhooks, or boundary-level constraints. These perimeter fences are designed for deterministic software. In an ecosystem of interconnected, self-improving agents, a superficial gateway is either easily bypassed or introduces latency that kills business execution. These legacy approaches to governing software will show well in a chatbot demo, but they are inadequate when it comes to governing the intent of an army of AI agents working together on behalf of a complex business ecosystem.

The New Paradigm: Algorithmic Business Assurance (ABA)

Because AI agents act on behalf of people and organizations, we reject the notion that AI governance belongs to an isolated compliance committee or an IT security silo. True AI governance must be a real-time business discipline. We are drawing a line in the sand between the old guard of passive tracking and the new frontier of active, mathematical assurance.

The Five Declarations of Absolute Assurance

I. From Retrospective Paperwork to Deterministic Execution Proofs

We must stop treating model cards and compliance reports as actual safety. They are static artifacts, outdated the moment they are compiled. True assurance requires deterministic runtime testing of probabilistic actions. Tests must be automatically generated at runtime and dynamic throughout their life. They must be written to generate real evidence, not just estimates of risk. If a model or an autonomous agent cannot mathematically prove it operated within explicit behavioral and regulatory boundaries during a transaction, the execution is dynamically invalidated before it can cause any harm.

II. Out of the Gateway, Into the Kernel

Perimeter guardrails are a relic of old software design. To govern highly autonomous multi-agent environments, the governance plane must sit natively inside the execution engine. We advocate for kernel-level supervisor micro-agents that continuously monitor internal tool calls, enforce dynamic least-privilege constraints, and audit the multi-agent trust chains from the inside out, without influence from the agent itself.

III. Govern the Intent, Not Just the Infrastructure

Monitoring data drift is no longer enough. AI systems change in production not just because the inputs change, but because their internal reasoning path diverges. True AI governance must actively isolate model drift, algorithmic bias, and execution intent shift. We must govern what the system is trying to accomplish, not just the pipeline it travels through.

IV. Real-Time Financial Risk Intervention

A qualitative risk score or a visual heatmap means nothing to a CEO during an incident. If risk cannot speak the language of liquidity, it is useless. Every single API call or agentic decision loop must be assigned an instantaneous economic liability score using dynamic exposure modeling:

Algorithmic Financial Risk = Likelihood of AI Intent Shift (%) × Negative Value Score × Total Dollar Value at Stake

When an autonomous agent attempts to execute a contract, process an insurance claim, or trade an asset where the real-time financial liability triggers a threshold breach, the system dynamically restricts or reroutes that action in stride.

V. Absolute Business Ownership

If an AI governance console requires a computer science PhD or a data science background to interpret, it can’t function in real time. The ultimate owner of the control plane must be the business unit leader carrying the P&L. Policies, guardrails, and risk tolerances must be managed in pure, natural business intent language, giving non-technical executives immediate, absolute decision power over the algorithms and the incidents. Without business ownership, agents cannot have true accountability.

The Ultimatum to the Enterprise

The era of treating AI governance as a defensive corporate box-ticking exercise is over. You cannot manage non-deterministic, fast-evolving AI systems with tools designed for static databases and point-in-time regulatory reporting.

The choice before enterprise executives is stark:

Continue deploying “governance bots” that document failures after they happen.

OR

Move to Algorithmic Business Assurance and architect systems that are mathematically incapable of violating your business intent.

Stop auditing your AI after it fails. Govern its execution in real time.

There are the “imposters,” old BI companies that snapped generative AI onto their products and call them agentic; the “innovators,” vendors that were smart enough to invest in AI-enablement and semantic layers all the way back in the mid-2010s; the “incoming,” more recent startups that have bold and flashy agentic AI capabilities built into their products but lack enterprise-readiness; the “inert,” who are so bogged down by legacy architecture they won’t survive the decade; and finally, the “inventors,” the AI-first companies whose very architecture was built from the ground up to support native agentic, headless, and enterprise-ready analytics.

There are several things that make Golden Analytics unique as an “inventor”:

AI First Architecture: With AI at the core of Golden Analytics, every part of the analytics flow in Golden is AI-driven, including data discovery, data preparation, automated insights, dashboard creation, descriptive narrative, storytelling, and collaboration. MCP makes Golden’s curated insights accessible and sharable.

The “Slider of Autonomy”: This is a critical shift in human-AI collaboration. Rather than having AI automatically do everything, the platform allows the human to define the level of AI intervention. Not only does this ensure that users retain control over the final analysis, it is a great bridge between traditional business intelligence and agentic analytics for users who are new to the game.

Automatic Analytic Workflow: Everyone has access to insight when the platform automatically profiles data at the point of connection, cleanses data, separates the signals from the noise, and gives users enough insight to start the conversation.

Analysis Optimizer: In the same way that databases optimize queries, Golden Analytics has a built-in optimizer that understands how people ask questions of data, routes the question through the best model, and helps users get to the answer they need more quickly.

Stories and Collaboration: Golden Analytics includes workspaces where users collaborate, and storytelling baked into the platform with a little help from analyst-grade explanations from the AI.

@Ferraro Consulting POV: Francois Ajenstat has an advantage right out of the gates. He was responsible for Tableau at the height of its fame and fortune. That means he understands customers, knows how to deliver what customers want, and understands what it takes to turn users into fanatics. Where Tableau inspired a whole generation of data people, Golden Analytics has the potential to ignite a movement of analysts and agents.

I was most impressed by Francois’ unique understanding of analytics users, and the way he has added a human element to platform design that takes into account what people need to transition from legacy BI to agentic analytics. Instead of forcing users to abandon their familiar territory and adopt something completely new, Golden Analytics gives people a natural pathway from legacy BI to modern autonomous analytics.

My favorite takeaway: Look at this ‘What’s New’ page: It makes me feel like the start of an old Tableau conference, only better: https://goldenanalytics.com/product/whats-new

Subscribe now