Cartoon of confused robots and office workers pointing at each other under speech bubbles asking Context and What is right — VentureBeat's illustration for its October 6, 2026 report on AI agent context. Image: VentureBeat.

Ask an AI agent for last quarter's revenue and it will give you a number. It will sound sure. The problem is that it may not know what your company means by "revenue." A semantic layer is supposed to fix that, but new survey data shows most agents are not using one.

On October 6, 2026, VentureBeat published the latest results from its VB Pulse tracker. Two-thirds of enterprises now say they run, pilot or build a governed semantic layer. Yet only 13% name it as the main place their agents get business context.

I'm Nitin from Debugged Pro. We build AI features and data dashboards for clients in India and the USA. The gap in this survey is one we see on real projects almost every month. Here is what the numbers say and what I think teams should do about it.

What Is a Semantic Layer, in Plain Words?

A semantic layer is a shared set of business definitions that sits between your raw data and the people or tools that use it. It answers simple but important questions.

What counts as revenue: booked, billed or collected?
Which customer record is the current one when there are three?
Does "active user" mean logged in this week or this month?

BI teams have used this idea for years so that two dashboards do not show two different answers. Now the same need has moved to AI agents. An agent that reads a table without these rules can return a figure that is technically correct and still wrong for the business.

What the VB Pulse Survey Found on Semantic Layer Adoption

The full report, written by Sean Michael Kerner, is worth reading. You can find it in VentureBeat's semantic layer survey coverage. The tracker has run three waves since June, and the trend is clear.

Survey wave	Run, pilot or build a semantic layer	Use it as agents' main context source
June 2026	58%	21%
July 2026	63%	19%
August 2026	67%	13%

So adoption is going up. Its use as the main source for agents is going down. Even inside the group that has a semantic layer, only 20% name it as the primary source of business context for their agents.

Wrong Answers Are Still Common

The survey also asked about confidently wrong answers. In August, 64% of respondents traced a confidently wrong agent answer to missing or inconsistent business context in the past six months. The figure was 68% in July and 57% in June.

That is not a model problem. It is a data meaning problem. The agent did what it was asked. It just did not know the rules.

Where AI Agents Get Context Instead

If agents are not reading the semantic layer, what are they reading? The August wave gave a clear breakdown.

Retrieval over documents: 32%. The agent searches files, wikis and PDFs and pulls in the closest matches.
Direct queries to live systems: 21%. The agent runs SQL or calls an API or an MCP server and reads what comes back. This was 11% in July, so it nearly doubled in a month.
Long-context loading: 19%. Teams paste a big set of documents or records straight into the model's context window.
A governed semantic layer: 13%.

VentureBeat makes a fair point here. Direct queries and long-context loading do not carry business definitions by default. A query can return a column called "rev_adj" without telling the agent what it means. Either method can be paired with a governed layer, but the survey does not show how many teams actually do that.

The fast rise of direct queries matches what we hear from clients. MCP servers are easy to set up. You connect a database, and the agent can read it in an afternoon. The business rules usually come later, if at all.

Snowflake Horizon Context diagram with three steps labeled Collect, Enrich and Activate above a Security and Governance bar — Snowflake describes Horizon Context as active context for AI and BI that collects, enriches and activates business definitions. Image: Snowflake.

Why Teams With a Semantic Layer Report More Failures

This is the most surprising part of the data. Enterprises that run, pilot or build a semantic layer report confidently wrong agent answers far more often. In August it was 78% for adopters against 37% for those still evaluating or with no plans.

A separate July group of 101 respondents showed the same pattern: 89% against 35%. The gap held even after removing companies that do not run agents on their own data or do not track root causes.

VB Pulse did not ask why. The report offers two possible reasons:

Better detection. A semantic layer gives the team a correct definition to check against. So these teams simply notice more of their own wrong answers.
Reverse cause. Teams that already had bad answers may be the ones who went and built a semantic layer.

The survey cannot tell which one is true. My own view leans toward the first reason. When we add clear metric definitions to a client project, someone usually finds an old dashboard that was wrong for a year. Errors do not appear because of the layer. They become visible because of it.

How Vendors Are Handling the Semantic Layer Gap

The big data and AI platforms are not taking the same road. VentureBeat spoke to three of them.

OpenAI: Build Context on the Fly

OpenAI's Data agent launched inside ChatGPT Work in September. It connects to warehouses like Snowflake, Databricks and BigQuery. It also pulls context from Slack, BI dashboards and file storage in the same pass.

It does not build a lasting context layer. Arpan Shah, general manager of the enterprise technology vertical at OpenAI, told VentureBeat: "It combines all the various contexts in their state, as opposed to generating a net new context layer itself."

Keewano: Add Meaning When Data Arrives

Keewano launched KeewanoDB in September. Agents read raw event data directly through MCP. But co-founder and CEO Mark Kardashov said the company builds a semantic layer during ingestion, using large language models to attach context to each event as it lands. So a direct query here is not a query without meaning behind it.

Snowflake: Prepare Context Before the Model Works

Snowflake's Cortex AI Gateway sends a task to a smaller model first and only moves up to a bigger one when needed. That only works if the task arrives with context already prepared.

Baris Gultekin, Snowflake's vice president of AI, explained in August that without good context, a model has to explore on its own. It writes SQL, tests it, searches the data and retries. That process costs more and usually needs a stronger model. Snowflake says its Horizon Context governed context layer and Cortex Sense tools prepare that context in advance, under its existing access controls.

Analyst Sanjeev Mohan, principal at SanjMo, summed it up: "Snowflake isn't really selling routing, it's selling routing that never leaves the governed data boundary."

What This Means for Agencies and Businesses in India and the USA

Most of our clients are not Fortune 500 firms. They are startups, mid-size companies and Global Capability Centres in India, plus growing product teams in the USA. Here is how this data applies to them.

The cheap path is often the risky one. Hooking an agent to a database through MCP is quick and impressive in a demo. Without shared definitions, it can give a CFO the wrong number in a board meeting.
Cost is tied to context. Gultekin's point matters for budgets. An agent that has to explore the data on its own burns more tokens and may need a pricier model. Clear context can make smaller models good enough.
Clients will ask about grounding. US buyers already ask how we stop AI from making things up. Indian enterprises in finance and healthcare are starting to ask the same thing. "We use RAG" is no longer a full answer.
You do not need an expensive platform to start. A semantic layer can begin as a well-kept metrics file and a few views in your warehouse. The tool matters less than the agreement on what each term means.

We wrote earlier about why business context matters for enterprise AI. This survey puts hard numbers on that idea.

OpenAI graphic with the words Now everyone can put data to work on a green and blue gradient background — OpenAI's launch graphic for its Data agent in ChatGPT Work, which builds context on the fly instead of keeping a separate layer. Image: OpenAI.

A 6-Step Plan to Ground AI Agents With a Semantic Layer

Here is the plan we use when a client wants an AI agent to answer questions about their own data. It works whether you are on Snowflake, BigQuery, Postgres or a mix.

List the 20 terms that cause fights. Revenue, active customer, churn, margin, lead. Ask finance, sales and product to agree on one definition each. Write them down.
Put definitions where agents can read them. Use metric views, a semantic model file or your platform's catalog. A Google Doc that the agent never sees does not count.
Route agent queries through the layer first. If you use MCP, expose governed views and metrics to the agent, not raw tables. Keep raw access for engineers.
Keep documents for the "why," not the numbers. Retrieval is fine for policies and how-to questions. Numbers should come from governed definitions.
Track wrong answers by cause. When an agent gets something wrong, log whether it was the model, the data or the definition. This is how you learn which fix pays off.
Review access rules. An agent should see only what its user is allowed to see. The NIST AI Risk Management Framework is a free and practical guide for setting these rules.

For step five, tooling helps a lot. Our post on AI observability tools for agent monitoring covers how to trace what an agent read and why it answered the way it did.

Quick Comparison of Agent Context Sources

Context source	Share of respondents (Aug 2026)	Good for	Main risk
Document retrieval	32%	Policies, guides, FAQs	Old or conflicting files
Direct queries (SQL, API, MCP)	21%	Live, fresh numbers	No business meaning by default
Long-context loading	19%	One-off analysis	High token cost, no shared rules
Governed semantic layer	13%	Metrics and KPIs	Setup effort and upkeep

Our Take

I don't think the lesson is "buy a semantic layer." Plenty of companies already have one and their agents still skip it. The lesson is to connect the layer to the path your agents actually use.

Direct queries through MCP are growing fast because they are easy. That is fine, as long as the agent hits governed views with clear names and definitions. The teams that do this well will get fewer wrong answers and lower AI bills.

Planning an AI agent on top of your company data? Our software development consulting team can help you map the definitions and data flow before you build.

Key Takeaways

67% of enterprises now run, pilot or build a semantic layer, up from 58% in June.
Only 13% use it as the main business context source for AI agents.
64% traced a confidently wrong answer to missing or inconsistent context in the past six months.
Direct queries through SQL, APIs and MCP jumped from 11% to 21% in one month.
Start small: agree on key terms, expose governed views to agents and track every wrong answer by cause.

About the Author

Nitin Kumar is the Lead Project Coordinator at Debugged Pro. He works with development teams and clients in India and the USA to plan software, data and AI projects, set delivery processes and keep quality high. He writes about practical AI adoption for agencies and growing businesses.