Context mining is the discipline of extracting usable context, such as schemas, code, logs, tickets, conversations, and runtime signals like query history and agent traces, before that raw material gets structured for an AI agent to use. Atlan names it Capability 1 of a five-part Context Layer model it built and uses internally, pulling from four categories of enterprise systems plus runtime signals to produce three outputs a reader will not find broken down this precisely anywhere else on the public web.
This page scopes “context mining” to that enterprise-AI-agent meaning, not the unrelated pervasive-computing research that also uses the phrase, and not process mining, an older discipline that analyzes event logs from business processes rather than extracting context for agents.
| Field | Value |
|---|---|
| What it is | The extraction stage that pulls usable context out of enterprise systems before it is structured for an AI agent |
| Key benefit | Mined context stays current, since it draws on runtime signals, not just documents someone remembered to write |
| Best for | Teams building the context layer up front, not bolting context onto agents after the fact |
| Not the same as | Process mining (event-log analysis) or the 2000s pervasive-computing usage of the phrase |
| Core inputs | Systems of record, data, knowledge, and work, plus runtime signals |
| Core outputs | Data graph, semantics and ontology, and skills |
Context mining vs. process mining vs. data mining
Permalink to “Context mining vs. process mining vs. data mining”Google’s results for “context mining” mix in process mining and data mining, two older, real, but different disciplines. Queries surface process-mining literature built around analyzing event logs to find bottlenecks, tied to researcher Wil van der Aalst and unrelated to feeding an agent. Data mining is closer in name only: it finds patterns within a dataset you already have; context mining pulls context that lives outside any single dataset.
Atlan’s five-source-type model, detailed next, is specific about what gets pulled and from where: an agent does not need a list of process bottlenecks, it needs the business context that tells it what a pattern means.
| Term | What it examines | Discipline age | Relevant to enterprise AI agents? |
|---|---|---|---|
| Context mining (this page’s meaning) | Enterprise systems and runtime signals, for agent context | 2024 to 2026, thin | Yes, directly |
| Process mining | Event logs from business processes | 2000s to present, established | Not directly, a different discipline entirely |
| Data mining | Patterns within an existing dataset | 1990s to present, established | Adjacent, not the same activity |
How does context mining work?
Permalink to “How does context mining work?”Context mining works by connecting to the systems where business context already lives and reverse-engineering it, rather than starting from a blank page, producing three outputs a context engineering layer then structures and delivers.
Prukalpa Sankar, co-founder of Atlan, puts it plainly: “The first job of the [context] layer is to connect to that fragmented reality and reverse-engineer business operations from it” (Prukalpa Sankar, “What an Enterprise Context Layer Actually Is,” Atlan, 2026). Mining is Capability 1 of Atlan’s five-capability context layer model, shown below against the other four.
| Capability | What it does | Atlan product surface |
|---|---|---|
| Context mining | Reverse-engineers business context from existing systems | Connectors, context agents, and data lineage |
| Context development lifecycle | Build, test, review, approve, deploy, learn | Context Engineering Studio |
| Compounding learning loops | Memory, feedback, and traces feed back into the shared layer | Platform-wide |
| Context activation and retrieval | Delivers context via MCP, A2A, SQL, and REST or GraphQL APIs | Context lakehouse |
| Context governance and observability | Quality, drift, lineage, versioning, and approval | Context layer reference architecture |
What systems does context mining pull from?
Permalink to “What systems does context mining pull from?”Four categories of enterprise systems supply the raw material, plus a fifth: runtime behavior, the signal most “RAG ingestion” framings miss. The types of metadata for AI agents underneath these categories are what mining extracts; the metadata layer for AI is where it lives once structured.
| Source category | Examples | What gets mined out of it |
|---|---|---|
| Systems of record | CRM, ERP | Ownership, definitions, business rules |
| Systems of data | Warehouses, lakes | Schemas, lineage, usage patterns |
| Systems of knowledge | Wikis, docs, tickets | Tribal knowledge, policies |
| Systems of work | Slack, email, PR history | How decisions actually get made |
| Runtime signals | Query history, agent traces, human overrides | What is actually being used and trusted right now |
Static documents answer what a system was designed to do; runtime signals answer what people actually do with it, and that gap is where most agent hallucinations live.
What does context mining produce?
Permalink to “What does context mining produce?”Three outputs come out of mining: a data graph that captures structure, a semantics and ontology layer that captures meaning, and skills that capture process.
| Output | What it contains | Concrete example |
|---|---|---|
| Data graph | Table descriptions, joins, lineage | “This table joins to opportunities on account_id” |
| Semantics and ontology | Metric definitions, relationships | “Revenue means closed-won bookings in Salesforce, recognized revenue in NetSuite” |
| Skills | Process maps, decision flows, exceptions | “How a reporting error actually gets escalated” |
The data graph is what teams reach for first, the closest analog to a traditional data catalog, but least useful alone: an agent context graph that knows two tables join but not what “revenue” means will answer confidently and wrong. Context graph tools for AI agents, context graph vs. vector database, and querying one with an agent cover what comes next.
Skills, the third output, are close cousins of agent memory: reusable process maps an agent draws on instead of re-deriving them each session (what is agent memory, agent memory architectures).
Manual vs. automated context mining
Permalink to “Manual vs. automated context mining”Manual context curation does not scale past a handful of systems and goes stale immediately. Automated mining treats runtime signals as a live source too, the same “just in time” loading Anthropic’s engineering team describes for agents that “dynamically load data into context at runtime” (Anthropic, September 2025), except mining is the earlier step that makes sure something worth loading exists.
| Aspect | Manual, documentation-led | Automated, mining-led |
|---|---|---|
| Coverage | A handful of systems, whatever someone wrote down | All connected systems, including runtime signals |
| Freshness | Stale the moment the doc is written | Updates as query patterns and agent traces change |
| Who does it | Whoever remembers to document | Connectors, context agents, and data lineage |
| Scales to | One team | Enterprise-wide |
The CIO's guide to context graphs
A practical breakdown of what a context graph needs to hold, who owns it, and how it differs from a traditional data catalog.
Get the CIO GuideContext mining vs. context engineering
Permalink to “Context mining vs. context engineering”Context mining and context engineering are adjacent, not identical: mining extracts the raw material, engineering structures and delivers it. That is Atlan’s internal distinction, not a settled industry split.
Context engineering is the established, analyst-and-lab-owned term. Anthropic’s canonical framing treats it as “the set of strategies for curating and maintaining the optimal set of tokens…during LLM inference” (“Effective context engineering for AI agents,” Anthropic, September 2025). A vendor glossary defines it as designing systems that decide what a model sees at each step, and an OpenAI community forum post calls it “assembling, summarizing, and delivering” context. None address where that material comes from, the question mining answers, upstream of context engineering vs. prompt engineering.
Most practitioners do not yet experience mining and engineering as separate jobs; they describe one workflow, building a context engineering framework or picking through context engineering platforms that claim to do both. Atlan names them as two separate capabilities internally, not a claim the industry has converged on. Whether running context engineering for RAG agents or context engineering for multi-agents, or asking what a context engineer actually does, mining comes first.
Where did the term “context mining” come from?
Permalink to “Where did the term “context mining” come from?”“Context mining” is not a term this page invents; it has a real, older lineage in a different field.
In 2018, a data scientist at Bisnode Analytics wrote: “Context appears through the relationship between disparate data…we mine the context of data” (Laurent Kinet, Bisnode Group, Medium, January 2018), describing text-mining applied to business data, different from this page’s meaning but proof the phrase itself is not new.
A separate, older lineage exists in pervasive computing: extracting situational data like location and device state from sensor streams. It resurfaces in unrelated work, a video-compression paper’s “temporal context mining” (arXiv:2111.13850) and a 2024 human-object-interaction paper, ContextHOI, extracting spatial context from images (arXiv:2412.09050), for the same underlying reason.
The enterprise-AI-agent usage here is 2024-to-2026 vintage with no Tier-1 owner: neither Gartner nor Anthropic has claimed “context mining,” both claim “context engineering” instead. Atlan already treats it as a named, defined capability internally; this page restates that position externally rather than inventing a coinage. Readers from the older data-governance vocabulary can see where the terms diverge in active metadata vs. context layer.
Why does context mining matter for AI agents?
Permalink to “Why does context mining matter for AI agents?”An agent is only as good as the context it can reach. Most of what it needs, such as which join to use or how an exception gets handled, never lived in one clean source; it is scattered across the systems named above, which is why mining treats runtime signals as a source in their own right. Without that stage, context drift sets in: the graph an agent reads from stops matching the systems it describes, and nothing catches the gap until it hallucinates.
Other builders reach for the same idea independently: Autostep, a Y Combinator company, for observing employee work patterns (“Autostep,” Y Combinator), and MineContext, a ByteDance/Volcengine open-source project, for a “proactive context-aware AI partner” (“MineContext,” GitHub, 2025). Neither operates at Atlan’s scale, but both confirm the activity is real.
Without mining, “context” for an agent tends to mean whatever got typed into a prompt, which does not survive an org chart change. That is why context drift detection and context quality testing exist downstream, catching what happens once mining stops running.
Find your context gaps
Run the Context Gap Calculator to see which systems your agents can already reach and which ones are still invisible to them.
Calculate Your GapHow Atlan approaches context mining
Permalink to “How Atlan approaches context mining”Most “context for AI” narratives start from a RAG pipeline pointed at whatever documents are easiest to reach, excluding runtime signals, the freshest source of what is true right now. That answers questions about last year’s documentation, not this quarter’s reality.
Context mining is Capability 1 of Atlan’s Context Layer model, connecting across systems of record, data, knowledge, and work, plus runtime signals, via connectors, context agents, and data lineage: connectors reach the systems, context agents extract the three output types above, and lineage traces transformations back to actual SQL and pipeline code, not memory. Atlan names mining as an explicit stage, distinct from context activation, the delivery stage after it.
Machine-mined context, refreshed continuously from lineage and usage patterns, degrades less than a wiki page nobody has opened since its author left. Teams evaluating whether to build this layer or adopt one can start with how to implement an enterprise context layer for AI: mine first, structure second.
Real stories from real customers: mining context at enterprise scale
Permalink to “Real stories from real customers: mining context at enterprise scale”Neither customer below uses the words “context mining,” Atlan’s internal name for the capability, not language customers reach for on their own. What they describe is the mechanism itself, in their own words.
"Atlan captures Workday's shared language to be leveraged by AI via its MCP server. As part of Atlan's AI labs, we're co-building the semantic layer that AI needs."
— Joe DosSantos, VP Enterprise Data & Analytics, Workday
Workday’s own language, the vocabulary its teams already use, is the kind of source-system material mining is built to reverse-engineer rather than have someone retype by hand. Feeding it to Workday’s MCP server is context activation, the stage after mining.
"Atlan is much more than a catalog of catalogs. It's more of a context operating system…Atlan enabled us to easily activate metadata for everything from discovery in the marketplace to AI governance to data quality to an MCP server delivering context to AI models."
— Sridher Arumugham, Chief Data & Analytics Officer, DigiKey
DigiKey’s “context operating system” describes the same sequence: mine what exists across marketplace discovery, governance, and quality signals, then activate it for people and AI models alike.
See context mining run live
Watch Atlan connect to real systems and mine a data graph, semantics layer, and skills in a live demo series.
Watch the Demo SeriesWhat mining actually settles, and what it does not
Permalink to “What mining actually settles, and what it does not”Context mining has to happen before context engineering can structure anything. It is the stage most “context for AI” narratives skip past, since it means connecting to messy, real systems rather than a clean document set. The term is thin and coined-but-real: no Tier-1 analyst has claimed it, independent builders reached for the same word without knowing about each other, and Atlan already treats it as a defined capability.
Whether “mining” stays separate from “engineering” industry-wide is still open. What mining produces, a data graph, a semantics layer, and reusable skills, is not: skipping the extraction stage for a RAG pipeline adds months to a deployment without improving accuracy. The context layer is the investment that compounds; the label on its first stage is secondary.
From here, the context development lifecycle covers what happens to mined context next, and context observability covers what happens after it is in use, whether it stayed current and traceable, or quietly drifted.
FAQs about context mining
Permalink to “FAQs about context mining”1. What does context mean in AI?
Permalink to “1. What does context mean in AI?”In AI, context is what a model or agent can access when it generates a response: the conversation so far, retrieved documents, tool outputs, and system instructions. For enterprise agents, it also includes business meaning: what a metric means to a team, who owns a dataset, how a process runs.
2. Is context mining the same as data mining?
Permalink to “2. Is context mining the same as data mining?”No. Data mining finds patterns in a dataset you already have. Context mining pulls context from outside any dataset, across docs, tickets, conversations, and runtime signals, so an agent has something usable to work with.
3. Is context mining the same as process mining?
Permalink to “3. Is context mining the same as process mining?”No, they share a search result, not a discipline. Process mining analyzes event logs for bottlenecks and deviations. Context mining extracts business context, such as definitions, ownership, and lineage, to feed an AI agent.
4. How is context mining different from context engineering?
Permalink to “4. How is context mining different from context engineering?”Context mining is the extraction stage, pulling context out of systems where it already lives. Context engineering is the structuring and delivery stage, assembling it into what an agent sees. Most practitioners treat these as one workflow.
5. What systems does context mining pull from?
Permalink to “5. What systems does context mining pull from?”Four categories: systems of record such as CRM and ERP, systems of data such as warehouses and lakes, systems of knowledge such as wikis and tickets, and systems of work such as Slack and PR history, plus runtime signals like query history.
6. Is context mining a manual or automated process?
Permalink to “6. Is context mining a manual or automated process?”It can be either, but manual curation does not scale past a handful of systems and goes stale quickly. Automated mining treats runtime signals as a live source, keeping context current as usage patterns change.
7. How do you build context for AI?
Permalink to “7. How do you build context for AI?”Start by mining what already exists across systems of record, data, knowledge, and work, plus runtime signals, before writing anything new by hand. That becomes the input for context engineering.
Sources
Permalink to “Sources”- Process Mining, Wikipedia
- MineContext, ByteDance/Volcengine, GitHub, 2025
- Effective context engineering for AI agents, Anthropic, September 2025
- What is Context Engineering?, euno.ai glossary
- Prompt Engineering Is Dead…, OpenAI community forum, July 2025
- What is Context Mining?, Bisnode Analytics / Medium, January 2018
- Temporal Context Mining for Learned Video Compression, arXiv:2111.13850
- ContextHOI: Spatial Context Learning for Human-Object Interaction Detection, arXiv:2412.09050, December 2024
- Autostep, Y Combinator company profile
- What an Enterprise Context Layer Actually Is, Prukalpa Sankar, Atlan, 2026
