Former CEO Tableau Mark Nelson: AI Agents Don't Need a Better Model; They Need Your Business Context

Your AI knows everything about the world and nothing about your company

The most capable models available today have read a substantial fraction of everything humans have written down. They have also never seen your customer hierarchy, your revenue recognition policy, or the three competing definitions of "active user" currently circulating in your BI tool.

Enterprises keep treating agent failures as model quality problems, then buying a better model... then watching the same failures recur. The forecasts point at the same misdiagnosis: Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027¹, and attributes those cancellations to escalating costs, unclear business value, and inadequate risk controls. Yet the real problem goes deeper: A system with no conceptual model of your business will produce fluent, confident, unusable answers no matter how good the underlying model gets, and supplying that model of the business is work only the enterprise can do.

On the season four premiere of AI Radicals², host Satyen Sangani spoke with Mark Nelson, Venture Partner at Madrona and former CEO of Tableau, about this challenge exactly. Nelson's characterization of it is the most useful starting point for data leaders trying to figure out what they own in this shift.

The parrot problem, restated for the enterprise

Nelson is direct about what large language models are and are not.

"If you imagine the current state of the technology and what LLMs have brought us, it's not intelligence," he says. "These things still don't understand the world. They don't have judgment." What they do have is coverage: "It is the most amazing encapsulation of knowledge that we've ever seen." A beat later he lands the image. "It is still fundamentally a parrot, and it's an amazing parrot, but it's a parrot."²

He reaches for an analogy he credits to computer scientist Yejin Choi, from her 2023 TED talk on why LLMs are simultaneously brilliant and inept³, and extends it: skyscrapers transformed cities, but no quantity of additional floors gets you to the moon. Different destination, different machine.

The version of that argument worth having inside an enterprise is narrower and more actionable. Set aside whether models will ever achieve general understanding. Inside your company, there is a specific and immediate understanding gap, and it has nothing to do with reasoning capability. The model has read thousands of articles about churn and has never encountered how your company calculates it. It knows what a data contract is in the abstract and does not know which of your tables are governed and which are somebody's abandoned experiment from 2023.

A world model is the structured understanding of entities, relationships, definitions, and rules that lets a system reason about a domain rather than pattern-match against text about it. Foundation models ship with a general one, assembled from public text. They ship with none of yours. Every enterprise deploying agents is discovering the difference between those two things, usually in production.

Judgment looks unteachable until you notice how much of it is context

Sangani made an observation in that cuts against the prevailing anxiety about AI and knowledge work. Information stopped being scarce, and demand for judgment went up rather than down. "All I do is, I only have more work to do, because everybody's sending us more information and more judgment is required, not less," he said, landing on judgment as the actual bottleneck.²

Nelson agrees on the mechanism, and identifies the roles most exposed as the ones built on codified procedure. As he puts it, the jobs at risk are the ones where you "learn a process, learn knowledge, and then regurgitate that process or that knowledge at the right time."²

Nelson's answer to what remains scarce is taste, the judgment to know what to build. Yet taste is personal and tough to transfer, and a large share of what passes for judgment in a data organization is not taste at all. It is accumulated institutional context: knowing that the finance team's revenue number is the one that ties to the board deck, that the marketing attribution table has been unreliable since the replatform, that a spike in one metric usually means a pipeline failure rather than a business event.

Institutional context behaves very differently from taste. Taste lives in individuals and resists documentation. Context can be captured, curated, versioned, and made queryable. That distinction determines what a data team can actually build. You cannot install taste into an agent. You can absolutely give it the definitions, ownership, freshness signals, and relationships that a competent analyst carries around in their head, and doing so converts a fluent parrot into something closer to a colleague who has read the runbook.

Non-deterministic systems fail in a new and more expensive way

Nelson uses one word in passing that deserves far more attention from governance teams than it usually gets. He describes AI as "encapsulated knowledge, non-deterministic, kinda human attributes, kinda not."² The hedge at the end is the whole point. These systems are close enough to human reasoning to be trusted like a colleague and different enough to fail in ways no colleague would.

Deterministic software fails loudly. A broken pipeline throws an error, a failed job pages someone, a malformed query returns nothing. Agents fail differently. They return an answer that is well-formed, appropriately hedged, correctly formatted, and wrong, and the failure surfaces weeks later in a decision nobody thought to audit.

This changes what oversight has to accomplish. Testing whether the system runs is no longer sufficient, because it always runs. The questions become whether the answer was grounded in governed sources, whether those sources were current, whether the agent had permission to see what it read, and whether anyone can reconstruct the path afterward. Analysts are beginning to size that exposure directly: Gartner projects that by 2030, half of AI agent deployment failures will trace to insufficient runtime enforcement in AI governance platforms⁴, and warns that in the near term, ungoverned decisions made with LLMs will produce financial or reputational loss for enterprises.

That is the practical substance of AI governance once agents are in production, and it depends on infrastructure most organizations built for human consumers rather than machine ones. Data lineage becomes the mechanism for tracing an assertion back to a source. Knowing which fields carry real consequence, the discipline of identifying critical data elements, determines where evaluation effort belongs. And the accumulated practice of model risk management turns out to transfer better to agentic systems than most teams expect.

Nelson also flags the flip side of cheap software creation, which is a governance problem wearing a productivity costume. On the prospect of everyone generating their own applications, he is unsentimental: "Wait until you have 20 CRMs bouncing around your company. You're gonna find out that that's not what you wanted. It's not what you wanted at all. And by the way, most of those people don't understand how to do security, don't understand about taking care of PII."²

The assembly language moment moves value beneath the interface

One of Nelson's sharper predictions concerns abstraction. "I think most programming languages are gonna become like assembly language was to us," he says, noting that he learned assembly, never shipped it professionally, and still drew on it as an architect.² The layer where humans express intent keeps rising. The need to know what should be built does not go away with it.

Apply that to enterprise applications and the strategic consequence is significant. Nelson points to Salesforce's Headless 360⁵ — announced in April 2026, exposing the platform's capabilities as APIs, MCP tools, and CLI commands — and calls it the right instinct, because, in his words, "no one ever wants to see a terrible lightning-based CRM screen ever again in their life."² If interfaces can be generated on demand, the interface stops being the moat. What remains durable is the logic and the governed information underneath: the pipeline model, the entity definitions, the permissions, the historical record.

"Do I really want to have to go rebuild the logic around what it takes to track a sales pipeline?" Nelson asks. "There's still a core of CRM that I'm better off paying someone to build for me than doing myself."²

Salesforce makes a version of the same argument from the other side of the table. Describing what Headless 360 gives an agent, the company notes that a coding agent wired to a raw database has no idea that a customer has an open escalation, a renewal due in 30 days, or a breached support SLA — because "that context took years to accumulate."⁵ It is worth registering who is making that case: the vendor whose screens Nelson says nobody wants to see again is arguing that the accumulated context underneath them is the durable asset.

This reframes a decade of data platform investment. The data catalog was originally justified as a way to help humans find and trust data. With the rise of AI, it can also serve as the machine-readable description of what a business contains and how its parts relate, which is exactly what an agent lacks.

That same logic elevates data products from an organizational nicety to a supply mechanism, because a curated, owned, documented product is something an agent can be pointed at safely while a raw table is not. Teams that already run a deliberate data product operating model are further along here than they realize.

Context needs an operating system, not a portal

If the enterprise world model is the thing agents are missing, it has to live somewhere with specific properties. It needs to describe entities and their relationships, not just list assets, which is why the debate about knowledge graphs and catalogs matters more than it appears. It needs semantics, meaning agreed definitions with owners attached. It needs policy encoded rather than documented, so permissions travel with the data. It needs freshness and lineage so an agent can weigh what it retrieves. Most importantly, it has to be readable at runtime by a machine, in milliseconds, rather than by a human browsing a portal during onboarding.

The analyst community has started to treat that layer as infrastructure rather than tooling. Gartner now expects universal semantic layers to be treated as critical infrastructure alongside data platforms and cybersecurity by 2030⁴, and describes semantic capability as a nonnegotiable budget line for data and analytics leaders rather than a discretionary one.

That set of requirements is the argument behind treating context as platform infrastructure, and it is the bet underneath Alation's AIOS. Agents are only as capable as the context they can retrieve, which is why trusted data determines whether agent programs succeed far more than model selection does. Nelson's own investment thesis runs in a compatible direction: he describes putting money into verticalized AI solutions for specific domains and industries, on the view that the rules of what buyers want — and where value accrues — are being rewritten across categories.² Inside a single enterprise, the domain being rewritten is your business, and nobody else can encode it for you.

What this means for data and AI leaders

The uncomfortable implication of Nelson's parrot framing is also the encouraging one. If models cannot supply understanding of your business, then no vendor can hand you a shortcut past the work of describing your own operations well. The organizations that get durable value from agents over the next two years will be the ones whose definitions, ownership, lineage, and policies exist as structured, queryable assets rather than tribal knowledge distributed across Slack threads and retiring employees.

That work was always worth doing. It used to be justified on the grounds of analyst productivity and audit readiness, which made it easy to defer. It now determines whether every agent you deploy has a foundation for reasoning about your company, which makes deferring it considerably more expensive.

Curious how a governed context layer and strong data governance foundations can make your agents reliable enough to trust with real decisions? Book a demo with us today.

Sources & notes

Every external claim on this page is independently verifiable. The public sources are listed here.

1. Over 40% of agentic AI projects forecast for cancellation by the end of 2027, attributed to escalating costs, unclear business value, or inadequate risk controls. — Gartner, 25 June 2025 ↗ https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027

2. All Mark Nelson and Satyen Sangani quotations are from AI Radicals, Season 4, Episode 1, "Why Enterprise AI Is Entering Its ROI Era." Full transcript available at the episode page. — Alation ↗ https://www.alation.com/podcast/episodes/enterprise-ai-roi-mark-nelson-madrona/ Disclosure: Madrona, where Nelson is a venture partner, is an investor in Alation.

3. Yejin Choi, "Why AI is incredibly smart and shockingly stupid," TED2023, April 2023. Nelson credits the moon-and-building analogy to this talk and extends it in the episode. — TED ↗ https://www.ted.com/talks/yejin_choi_why_ai_is_incredibly_smart_and_shockingly_stupid

4. Predictions that 50% of AI agent deployment failures will stem from insufficient AI governance runtime enforcement by 2030, and that universal semantic layers will be treated as critical infrastructure by 2030. — Gartner, 11 March 2026 ↗ https://www.gartner.com/en/newsroom/press-releases/2026-03-11-gartner-announces-top-predictions-for-data-and-analytics-in-2026

5. Salesforce Headless 360 announced 15 April 2026, exposing platform capabilities as APIs, MCP tools, and CLI commands; context-accumulation argument from the same announcement. Accessed 13 August 2026. — Salesforce ↗ https://www.salesforce.com/news/stories/salesforce-headless-360-announcement/

Analyst attributions & disclaimers

Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, Anushree Verma, 25 June 2025.

Gartner, Gartner Announces Top Predictions for Data and Analytics in 2026, Rita Sallam, 11 March 2026.

Gartner does not endorse any vendor, product or service depicted in its research publications and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner research publications consist of the opinions of Gartner's research organization and should not be construed as statements of fact. Gartner disclaims all warranties, expressed or implied, with respect to this research, including any warranties of merchantability or fitness for a particular purpose. GARTNER and Magic Quadrant are registered trademarks and service marks of Gartner, Inc. and/or its affiliates in the U.S. and internationally and are used herein with permission. All rights reserved.