The Problem Nobody in AI Marketing Will Admit

Every enterprise AI pitch promises accuracy. The reality, according to Snowflake EVP of Product Christian Kleinerman, is blunter: "There are a lot of tools out there that you can ask questions, you get a very confident answer, but whether it's correct or not is different."

AI is already on the job, and it keeps lying with complete confidence.

This isn't a model problem. The models are fine. The problem is that the word "revenue" means one thing in a BI dashboard, something slightly different in a SQL table, and something else entirely in an agent's instruction set. Three agents, same database, three answers. Nobody owns the definition.

Snowflake Swings at the Root Cause

At Snowflake Summit 2026 in San Francisco, Snowflake announced a two-layer system called Horizon Context and Cortex Sense — built specifically to give AI agents a single, governed definition of business logic across every retrieval stack.

Horizon Context is built on Snowflake's acquisition of Select Star. It pulls metadata from Postgres, SQL Server, Tableau, and Power BI into one catalog. Every agent, every BI tool, every external system draws from the same governed definition — not its own raw schema interpretation.

Cortex Sense goes a step further. It automatically builds and enriches context from usage patterns, without requiring manual maintenance.

According to VentureBeat's VB Pulse Q1 2026 survey — drawn from organizations with 100 or more employees — hybrid retrieval intent tripled from 10.3% in January to 33.3% in March. That's the fastest-growing strategic position in their dataset. More enterprises are running more complex retrieval architectures. More complexity means more opportunities for agents to return different answers to the same question.

Your Finance Team Is Already Uploading Contracts Into Personal ChatGPT Accounts

Meanwhile, Zip — the AI procurement platform valued at $2.2 billion — announced at its AI Summit in New York that it's addressing a problem procurement chiefs describe in private but rarely say publicly.

Employees are already using AI for sensitive financial work. They're just doing it in personal, unmonitored accounts. Spend data going into Claude. Contracts being redlined inside ChatGPT. Internal financial analyses generated through personal Gemini accounts. No audit trail. No compliance controls. No record.

The consequences are serious. SOX violations carry fines up to $25 million. Executives can face prison time. Public companies that fail compliance audits can be delisted.

Zip's answer is a suite of five AI "Superagents" — capable of reviewing contracts, coding invoices, and negotiating with vendors — but constrained inside Zip's own governance framework. The company also announced a procurement-native implementation of the Model Context Protocol (MCP) that pipes Zip's data directly into Claude and ChatGPT without the data ever leaving controlled, audited systems.

The launch included speakers from Anthropic, OpenAI, Datadog, and Humana — companies whose tools employees are already using unsupervised.

Gartner predicts 40% of enterprise applications will include task-specific AI agents by end of 2026, up from less than 5% today. The compliance window to get ahead of this is closing fast.

Healthcare: Where Wrong Answers Cost More Than Money

The stakes get higher in healthcare. MIT Technology Review published a sponsored piece from Ema on June 2 detailing how agentic AI is being deployed across hospital systems — with some striking numbers that deserve wider attention.

At Hospital for Special Surgery (HSS) in New York, AI agents now complete 1,100 insurance claims per month. They've cut the appeals process from 45 minutes to five. Appeals success rate jumped from 65% to 100% over nine months.

Dr. Ashis Barad, HSS's Chief Digital and Technology Officer, says the difference between agentic AI and prior digitalization attempts is that agents handle nuanced, complex scenarios without defaulting to a human every time a case falls outside a rigid framework.

The WHO has warned that global healthcare worker shortfalls will reach 11 million by 2030. AI handling back-office claims processing isn't a luxury — it's triage.

But there's tension here: Snowflake just spent its entire summit explaining how AI agents return confidently wrong answers when definitions aren't governed. Healthcare is not the place to find out your agent defined "prior authorization" differently than your payer does.

What Mainstream Coverage Is Missing

Most AI coverage is still framing this as "AI coming for jobs" or "AI ethics." Both miss the immediate operational problem.

Data governance at the context layer is the story — and almost nobody outside of enterprise tech trade press is covering it. The companies that deploy AI agents without solving the shared-definition problem aren't going to get sued for using AI. They're going to get sued because their AI gave a confident, auditable, completely wrong answer to a compliance question, and nobody caught it.

Small businesses face a different version of the same issue. MIT Technology Review profiled London-based tutor Sam Finnegan-Dehn, who uses Notion AI for invoicing, meeting summaries, and goal-setting. It works. But he's one person with low-stakes data. Scale that to a 500-person procurement team with SOX obligations and you have a different problem entirely.

What's Next

Enterprise AI just hit its next wall. Faster models didn't fix it. Cheaper vector search didn't fix it. The fix is boring, unglamorous data governance work — defining what your data means before you let an agent act on it.

Companies that skip that step aren't innovating. They're building a liability.