Key takeaways
- Pick one domain with one owner, such as revenue or support, and finish it before starting another.
- Customer facts and deal notes belong in warehouse tables as data, and the context layer is the map that tells agents how to read them.
- Write down the metrics people already argue about, then certify them with the person who owns each number.
- Ownership and access come before agents, so the first agent inherits rules that already work.
- Federation, conflict resolution and audit are what “enterprise” adds, and they can wait for the second domain.
- 01Inventory
Systems, tables and what is still queried
- 02Metrics and entities
Certified definitions, joins and usage rules
- 03Ownership and access
An owner per metric, roles and row rules
- 04Gateway to agents
Each agent gets a role, a token and a log
What is an enterprise context layer?
A context layer tells people and AI agents how to read your company data. It records which systems feed the warehouse, what each table and column means, how tables join, which metric definitions are certified, and which rules apply to whom. When someone asks for revenue by segment, the context layer says which table, which filters and which definition to use.
It is a map of the data, and it should stay a map. A fact such as “John is the CFO and the main buyer” is data about a deal, and it belongs in a table in the warehouse, where it can be queried and kept current. Memory files and markdown folders that copy those facts drift away from the source within weeks. We go further into that distinction in our post on what context actually is.
Why do AI agents need a context layer?
An agent pointed at raw tables has to guess. It sees three columns that might be revenue, two customer tables and a status field with values nobody documented, and it picks one. The answer comes back in a confident sentence, so a wrong join is easy to miss.
With a context layer, the agent reads the certified definition and the usage rules before it writes SQL, and it queries through a role that limits what it can see. The answer is reproducible because two people asking the same question reach the same metric.
Where should you start?
Start with one domain that has a clear owner and a question people ask every week. Revenue reporting, support performance and payments are good candidates, because the numbers already exist and somebody already disputes them. Avoid starting with a company-wide graph or a platform tour. Both take months and produce a diagram before anyone gets a correct answer.
Write down ten to twenty questions the domain owner wants answered. They become the test set for every later step.
How do you inventory what you already have?
List the source systems for the domain, the tables they land in and the reports that read them. Then check usage. Query history shows which tables people still read, and BI usage logs show which dashboards anyone opens. In a merged education software group Rig worked with, an overnight audit showed that only about a third of the Tableau workbooks were still in use, which cut the rebuild down to what people needed.
- List each source system and the warehouse schema it lands in.
- Mark each table as queried, stale or empty, using query history.
- Trace each live report back to the tables behind it.
- Describe every table and column the domain needs, starting with the busiest.
How do you define metrics and entities?
Take the questions from the first step and find the numbers behind them. For each metric, write the definition in plain English, the SQL, the table it reads and the filters it applies. Compare the result with a figure the business already trusts, such as the finance pack, before you mark it certified.
Entities matter as much as metrics. Decide what one customer, one account and one order mean in this domain, and how they join across systems. Usage rules capture the exceptions that live in people’s heads, such as which deals need CFO sign-off or which rows are test data.
Who owns the definitions and access?
Every certified metric needs a named owner who approves changes. The data team controls definitions and access, and the business owner decides whether the number answers the question. Set roles and row rules at this stage: who sees all regions, who sees their own team, and which columns are masked. Test each role before anyone is given it.
How do you open the context layer to agents?
Connect agents last, through a gateway. Each agent gets its own role, copied from one you trust and narrowed, its own token and a daily call limit. MCP is the usual protocol, so Claude, Cursor and other clients can read the definitions and query through the agent’s role. Run the test questions from the first step through the agent and fix any wrong answer at its source, in the definition, the usage rule or the table description.
What does “enterprise” add later?
The second and third domains bring problems the first did not have. Two teams define active customer differently, so you need a way to record both definitions and agree which one is certified for board reporting. Several warehouses need one set of agent roles. Audit becomes a regular task, with someone checking which agents ran which queries each month. Plan for these, and solve them when a second domain makes them concrete.
How does Rig do it?
Rig Map is Rig’s context product, and it holds the ontology, the semantic layer and the context layer on the warehouse you already run. Rig drafts table and column descriptions from sample rows, detects joins, and builds a glossary and draft metrics, which your team reviews and certifies. Areas group tables, metrics and rules by business domain, and Rig Map updates as schemas change. A context score and replayed test questions show which parts of a domain still answer badly.
Gateway gives each person and agent a role with schema, column and row rules, and gives every agent its own token, call cap, owner and audit trail. Cleo’s data team certifies the metrics and table descriptions in Rig Map that its agents query across a 4,000-table warehouse.
Tools and approaches
- Atlan
- Atlan describes the context layer as the governed layer that connects AI agents to trusted data. Its how-to guide has five steps: pick one domain and one source, connect the source, enrich metadata with its context agents, have a domain expert review and certify, then expose the context through MCP and validate. A warehouse-native metric layer and a query-time gateway can sit alongside that catalog work.
- DataHub
- An Apache 2.0 open-source metadata platform with column-level lineage, discovery and governance features. It suits teams who want to run their own catalog for inventory, lineage and ownership.
Common questions
How long should the first domain take?
Plan for the time it takes to answer the domain owner’s test questions correctly through an agent. Keeping to one domain and one owner lets you measure progress against those questions each week.
Do I need a new catalog?
You do not need one to start. A catalog helps with inventory and lineage across many systems. The first domain needs table descriptions, certified metrics, usage rules and access, which can sit next to any catalog you already run.
Can I start without MCP?
Yes. Definitions, ownership and access are useful to analysts and BI tools on their own. MCP matters when you connect agents, and it should go through a gateway that gives each agent its own role.
How does a context layer sit next to a semantic layer?
The semantic layer is the set of certified metric definitions. The context layer includes it, and adds table meanings, joins, entities, usage rules and access, which agents need for questions no single metric covers.
What does “enterprise” actually add?
It adds more domains, warehouses and owners, which bring conflicting definitions to resolve, one set of agent roles across platforms, and regular audit of agent queries.
Sources
Vendor documentation checked on 20 November 2026. Confirm current features and terms with the vendor.