Context OS is one layer of a Governed AI stack.
Updated 31 October 2026
Short answer: AI agents need more than a catalog that describes data or a semantic layer that defines metrics. They need current, governed context that says which facts apply to a decision, under which rules, and with what evidence. A data catalog, a semantic layer, a knowledge graph and a context operating system each solve part of that problem. The choice depends on which decision the agent will make and what the regulator will ask about it afterwards.
Why this question now
When a person asks a question of a dashboard, a mistake is visible and usually reversible. When an agent acts, it reads customer records, calls tools and changes things. The data it sees must be right at that moment, and the bank or enterprise must be able to show later what it saw.
Most enterprises already own some of the four categories. The question is which gap they are trying to close. Buying a fifth tool rarely closes it. Working out which job each existing tool does, and which job nobody does, usually does.
The four categories, defined
Data catalog. A catalog records what data exists, where it lives, who owns it and how it moves between systems. Its strength is discovery and lineage. It describes data; it does not usually decide whether an agent may use a record for a particular action.
Semantic layer. A semantic layer defines business terms and metrics once, so that "revenue" or "active customer" means the same thing in every report. It keeps analytics consistent. It is built for queries, and it does not record the decision an agent made with the result.
Knowledge graph. A knowledge graph stores entities and the relationships between them, such as a customer, their accounts, their guarantors and the products they hold. It is strong for reasoning across connected facts. It does not by itself enforce who may see or change those facts, nor keep a record of each decision.
Context operating system. A context operating system sits in the runtime path. It supplies governed context to an agent at the moment of action, checks the action against policy, routes exceptions to a person, and records each decision with the context it used. It is the newest of the four categories and the least standardised, so buyers should test each claim against their own workflow.
Figure: Four categories compared by the job each does for an AI agent. Source: ElixirData analysis, 31 October 2026; not a benchmark.
Where each one falls short for agents
- A catalog can tell an agent where a customer's income figure is stored, but not whether the agent may use it for a credit decision today.
- A semantic layer can give consistent numbers, but it cannot show which numbers an agent relied on when it closed an alert.
- A knowledge graph can link a customer to every related account, but it needs a separate control to stop the agent from using a restricted link.
- A context operating system can enforce the rule and keep the record, but it works best alongside a catalog and a semantic layer, not instead of them.
In practice, the four are complementary. The mistake is to expect one of them to do all four jobs.
A comparison by job
| Question | Data catalog | Semantic layer | Knowledge graph | Context OS |
|---|---|---|---|---|
| What data exists and who owns it? | Yes | Partly | Partly | Yes, via integration |
| What does a business term mean? | Partly | Yes | Partly | Yes, via integration |
| How are entities related? | No | No | Yes | Yes, via integration |
| Is the action allowed under a policy? | No | No | No | Yes |
| Is there a trace of the decision? | No | No | No | Yes |
A practical way to choose
- Pick one agent workflow, such as alert triage or customer onboarding, and write down the decision it makes.
- List what the agent must know to make that decision: the data, the definitions, the relationships and the rules.
- Ask which of the four covers each item. Mark any item that no category covers.
- Ask what the regulator will expect as evidence. Use the primary regulatory text, not a summary. The evidence requirements decide how much runtime recording you need.
- Test the vendor claim against one real case. Ask the vendor to show the decision record, not a slide.
Check where each capability runs. If policy is evaluated in a document that the agent never reads, the control is not at the point of action. If traces are generated by the model rather than by the platform, they may not be complete. Ask the vendor to show both at runtime.
What a supervisor will ask the context layer
A supervisor reviewing an agent that moves money or changes a customer record will ask a short list of questions. Which data did the agent use, and was it current? Which rule governed the action, and which version was in force? Who approved it, and on what information? Could the decision be reproduced? A context layer that answers these from its own records meets the question. A catalog, a semantic layer or a graph can contribute to the answers, but none of them records the decision by default.
The questions also show why the four categories are not interchangeable. The catalog can say where the data came from. The semantic layer can say what a term means. The graph can say how entities relate. Only the decision record answers what was done with them.
Common mistakes when buying
The first mistake is buying by category label. Two products that both say "context" can cover very different functions. A short demonstration of one agent action will reveal more than a feature list.
The second mistake is treating traces as optional. Teams often add logging after go-live and discover that the logs do not link inputs, policy versions and approvals. Require the trace format in the contract, not as a later enhancement.
The third mistake is ignoring change control. A policy that can be edited without approval, or a definition that changes silently, undermines every decision made after the change. Ask how changes are versioned and who signs them off.
The fourth mistake is assuming that one platform removes the need for ownership. Definitions, relationships and policies each need a named owner. A context layer without owners drifts out of step with the business within a few quarters.
Regulatory context
Regulators do not yet define a separate category called a context OS, and none of the instruments below names one. They do set expectations for the outcome a context layer must support: control of data use, documented rules, human oversight and records. Binding status differs by instrument, so the list below states it for each.
- European Union. Regulation (EU) 2024/1689 (AI Act), Article 14, requires that high-risk AI systems can be effectively overseen by people. It is binding law. Regulation (EU) 2026/1744, which amends the AI Act to simplify its implementation, entered into force on 27 July 2026. Read the amended text for high-risk application dates. EUR-Lex: Regulation (EU) 2024/1689; EUR-Lex: Regulation (EU) 2026/1744.
- Germany. BaFin's guidance on ICT risks in the use of artificial intelligence describes itself as non-mandatory advice. DORA, Regulation (EU) 2022/2554, is binding and applies from 17 January 2025. BaFin guidance page; EUR-Lex: Regulation (EU) 2022/2554.
- United Kingdom. The Prudential Regulation Authority's supervisory statement SS1/23 sets model risk management principles for banks. Its current version is effective from 23 April 2026. The PRA page says its scope covers banks with internal model approval for regulatory capital, and also building societies and PRA-designated investment firms. AI is in scope to the extent it is a model. Separately, the ICO's guidance on automated decision-making under UK GDPR Article 22 is under review following the Data (Use and Access) Act 2025, which came into law on 19 June 2025, and the ICO says it may change. PRA SS1/23; ICO: automated decision-making.
- United Arab Emirates. The Central Bank of the UAE's guidance note on consumer protection and responsible adoption and use of AI and ML is guidance. It states that it shall supplement, and not replace, any laws, regulations or directives. The Federal Decree by Law No. 45 of 2021 on personal data is binding law and has been in force since 2 January 2022. For firms in the Dubai International Financial Centre, the DFSA's AI survey is a useful reference, but it is not a rule. CBUAE rulebook; UAE legislation portal.
- Saudi Arabia. The SDAIA AI Ethics Principles, version 1.0 (September 2023), set seven principles, including accountability and human oversight across the AI lifecycle. The document describes registration as optional, and its binding status is unclear. We did not find an AI-specific rule from the Saudi Central Bank for banks in the primary sources reviewed. SDAIA AI Ethics Principles.
Across the jurisdictions the pattern is consistent. Expectations are stated for oversight, documentation and data protection, and agents are not named as a separate category. A context layer that records its decisions meets these expectations more easily than one that does not.
Where ElixirData fits
ElixirData positions Context OS as the runtime layer that applies governed context and policy to agent actions and records each decision. It is designed to work with the catalog, semantic layer and knowledge graph a buyer already has. Buyers should evaluate it against the same criteria as any other option. Product information is at https://www.elixirdata.co/.
Disclosure: ElixirData publishes this article and sells Context OS.
Related reading
For the criteria to use when comparing platforms, read Which Platforms Give AI Agents Governed Business Context? For the runtime controls, see the forthcoming articles on how to enforce policy checks before an AI agent acts and how banks keep an audit trail of every AI agent decision. For background on what agents need from context, read Context Engineering for AI Agents: From Retrieval to Decision Context and The Complete Guide to Bank AI Agent Controls.
Frequently asked questions
What is a context OS?
A context OS is a layer that supplies governed context to AI agents and records how they used it. Governed context has a source, an owner, an access rule and a version. The layer also keeps each decision's inputs, the policy applied, any approval and the outcome. No standard body defines the term.
Is a context OS a replacement for a data catalog?
No. A context OS can use catalog metadata for ownership and lineage. It serves agents that take action, while a catalog serves people who search for trusted data. Most organisations keep the catalog and add the context layer on top of it for agent use.
Where does a knowledge graph fit?
A knowledge graph can be one input to a context OS. It supplies relationships between customers, accounts and products. The context layer then applies policies and approvals to those relationships and records the decisions made.
Sources
- European Union, Regulation (EU) 2024/1689 (AI Act), Article 14. Binding law: EUR-Lex.
- European Union, Regulation (EU) 2026/1744 (Digital Omnibus on AI), entered into force 27 July 2026: EUR-Lex.
- European Union, Regulation (EU) 2022/2554 (DORA), applies from 17 January 2025: EUR-Lex.
- BaFin, guidance on ICT risks in the use of AI. Non-mandatory advice: BaFin.
- Bank of England and PRA, SS1/23, effective 23 April 2026: PRA SS1/23.
- Information Commissioner's Office, automated decision-making and profiling guidance, under review after the Data (Use and Access) Act 2025: ICO.
- Central Bank of the UAE, guidance note on consumer protection and responsible adoption and use of AI and ML: CBUAE rulebook.
- SDAIA, AI Ethics Principles, version 1.0, September 2023: SDAIA PDF.
- NIST, AI Risk Management Framework 1.0, 26 January 2023. Voluntary: NIST AI RMF.
