We want to believe that we can hand tasks over to AI agents and have them execute quickly and flawlessly on our behalf. But how many times have you gotten exasperated trying to re-explain what you’re looking for, knowing that what that agent has given you isn’t it? At a certain point, it’s easier to just do it yourself — which defeats the purpose.
The problem is that your agents can’t see your richest AI asset: context.
It starts with the data catalog. Your team spent years curating a catalog that contains the most complete record of what your data means, who owns it, and what it’s certified for. But then it sat idle, useful to humans who knew where to look but invisible to agents.
The impact of broken agent communication is starting to show. S&P Global Market Intelligence reports that 42% of enterprise AI projects are scrapped before reaching production, up from 17% just a year prior.
The common thread is context. Agents surface the wrong definitions, outputs that require human intervention, and brittle workflows that don’t work if they don’t match the test set. The catalog may contain the answer, but nobody pointed the agent in its direction.
MCP changes that, making the catalog an API instead of a destination. It acts as the pipeline that moves context from your catalog (and other connected systems) to your agents. MCP allows context to flow into agents, making workflows more resilient and outputs more reliable. Agents no longer look in the wrong places because context comes directly to them.
Finally, you may have found your solution to the “I’ll just do it myself” problem — and it’s MCP.
Real use cases, built on MCP
Permalink to “Real use cases, built on MCP”When you connect Atlan’s MCP server to any surface that supports the protocol, your catalog becomes queryable by machines. Any agent that can make an MCP call can ask: What is this table? Who owns it? What’s upstream and downstream? Is it certified, and for what? What does this column mean, and does it carry a sensitivity classification?
The answer comes back governed, current, and pulled from the context your team has built, with no humans in the loop and no one opening Atlan to look it up.
What does that look like for real companies?
A new hire at a financial services firm used Claude with Atlan MCP to navigate the company’s data estate. Without any support from the data team, the employee surfaced the right tables for an analysis. Instead of browsing an unfamiliar catalog, which could have taken hours, the catalog was answering questions for them.
A media company’s engineers now see how a dbt change ripples downstream before they push the commit. The IDE calls Atlan via MCP and returns the full impact map. Answering the question “How many tables would this break?” now takes seconds, not a day of tracing dependencies by hand.
A financial institution built a semantic search agent that breaks down user prompts, identifies the most likely certified term from the glossary, and enriches queries before generating any SQL. Analysts get trustworthy answers without the governance team having to carve out time for oversight and review.
A cybersecurity team fed 104 requested fields through Claude and Atlan MCP in a single pass and got back a structured audit: 60 fields exist, eight are missing source documentation, and twelve are calculable from existing fields. What used to be a multi-day discovery exercise became a one-shot query.
A semiconductor company began running a workflow where an internal LLM bulk classifies hundreds of assets for PII, SOX requirements, and sensitivity in a single sweep. Instead of taking months to do this manually, the company’s governance team now simply needs to review and approve classifications.
A global bank built AI agents grounded in a semantic layer curated in Atlan via MCP. The agents don’t hallucinate metric definitions because the definitions come from the source of truth, not just from whatever data the model was trained on.
These use cases run in IDEs, LLMs, Slack bots, custom agents, and BI tools. The common thread is that the AI could read the catalog. Without the MCP connection, none of it is possible.
See Atlan's MCP in action
Watch a live demo of Atlan's MCP server connecting your catalog to AI agents across IDEs, LLMs, and custom workflows.
Watch Context Layer LiveWhen it comes to MCP, most teams stop too soon
Permalink to “When it comes to MCP, most teams stop too soon”Most teams using MCP today have stopped at the obvious use cases: talk-to-data, glossary building, and asset tagging. But that barely scratches the surface.
What if an agent monitored your data estate for schema changes and automatically notified the right owners, complete with a lineage-traced impact summary, before anyone filed a ticket?
What if a migration workflow had an agent classify every asset in your source system, check it against your governance rules, and tag each one Migrate, Review, or Deprecate before a human got involved?
What if a root cause analysis for a broken dashboard took the form of an agent tracing lineage upstream through every transformation, surfacing the first point of failure and writing the incident report?
None of these require smarter AI; they require AI that’s informed. MCP is the distinction that makes informed AI possible.
And the cycle gets better over time. Every agent interaction that includes governed context feeds back into the catalog. Every term a steward corrects gets reinforced. Every certification an agent applies gets validated and written back. The system compounds with use, with context as the moat and MCP as the delivery mechanism.
Why prompt-based context isn’t the answer
Permalink to “Why prompt-based context isn’t the answer”We’ve heard from customers: “Can’t you hardcode business context into system prompts?” It’s a valid question, and for a single agent on a stable use case, the answer is yes.
But hardcoded context doesn’t update when definitions change or scale across dozens of agents. It fails when a prompt template drifts or a model changes behavior. Every new agent must be built from scratch, re-encoding your catalog’s knowledge.
Teams we work with that are successfully scaling AI have one thing in common: they’ve given context its own layer, separate from the agents that consume it. A catalog is a source of truth, but prompts are volatile. Building on prompt-based context is like building analytics on CSV exports: functional at small scale, brittle as you grow, and expensive to maintain.
Is your AI compounding or just helping?
Permalink to “Is your AI compounding or just helping?”We all want AI to be helpful, but that’s a low bar. There’s a difference between organizations where AI assists individual workflows and organizations where its intelligence compounds across the whole stack. That difference depends on whether your business context is accessible to machines.
The organizations moving toward compounding AI are asking infrastructure questions, like “what does every agent in our stack read from, and do we own it?” MCP is the answer to those questions.
But it’s more than just an infrastructure solution. MCP can accelerate onboarding, make teams more efficient, and automate many of the manual processes that used to slow organizations down. It can finally free you from having to answer questions yourself, simply because you’ve memorized what lives in your catalog. The teams who figure this out early won’t be easy to catch.
Find out more about Atlan’s MCP here.
FAQs about MCP and business context
Permalink to “FAQs about MCP and business context”What is Atlan’s MCP server?
Permalink to “What is Atlan’s MCP server?”Atlan’s MCP server connects your data catalog to any surface that supports the Model Context Protocol, including IDEs, LLMs, custom agents, and BI tools. It makes your catalog queryable by machines, so AI agents can retrieve governed metadata like table definitions, lineage, certifications, and PII classifications without human intervention.
Why do AI agents fail without catalog context?
Permalink to “Why do AI agents fail without catalog context?”AI agents built without catalog access must rely on hardcoded prompts or incomplete schema information, leading to unreliable outputs and eroded trust. S&P Global Market Intelligence reports that 42% of enterprise AI projects are scrapped before production — a pattern often traced to agents that cannot reach the organizational context already available in the catalog.
What types of context does Atlan’s MCP expose to AI agents?
Permalink to “What types of context does Atlan’s MCP expose to AI agents?”Atlan exposes six categories of context via MCP: asset and table metadata, data lineage, business glossary and certified terms, column-level metadata, governance classifications like PII and SOX scope, and semantic context including SQL templates and metric definitions.
Sources
Permalink to “Sources”- CIO Dive / S&P Global Market Intelligence, “42% of AI projects scrapped before completion,” 2024. https://www.ciodive.com/news/AI-project-fail-data-SPGlobal/742590/
