If you follow conversations around Enterprise AI, you’ll have noticed one word cropping up more and more: context. So what does this mysterious (but hugely consequential) buzzword have to do with AI prompts, models, and agents? Where does context come from? And how can you put it to work for game-changing AI transformation? This glossary post’s job is to put context into, well, context. So let’s rattle through the key terms before pointing you in the direction of the kind of context that AI needs to succeed.
Summary
- Broadly speaking, context in AI describes the information that’s contained within a user’s prompt or retrieved from business systems and relevant documents to inform a decision.
- Context helps frame the instructions given to AI and guide it toward the right output and outcomes for the business and its strategy.
- The most important kind is operational context, which brings together the facts, relationships, rules, constraints, and process history that explain how the business runs.
What is context in AI?
If you take the standard definition of the word “context” and apply it to AI, you’re pretty much on the money. It’s an umbrella term for any information that explains the history and parameters of the environment in which AI is being deployed. In other words, the information that’s available to an AI model or solution when it generates a response, makes a decision, uses a tool, or takes an action.
This is where things get woolly, though. As the types of AI solutions have expanded, the term encompasses more varieties and applications of contextual information:
- Context for chatbots, copilots, and assistants can include the user’s prompt, prior conversation, system instructions, retrieved documents, tool outputs, structured data, business rules, process information, or operational signals.
- Context for AI agents also includes information from tools, systems, memory, workflows, and previous actions.
- Context for large language models is often discussed in terms of what fits in the context window (more on that in the next section).
But what we’re really talking about – what really matters in Enterprise AI – is operational context. This collates the business-specific facts about the historical actions, processes, and make-up of the organization, which AI needs to generate useful, trustworthy responses. This operational context comes down to two essential components:
- Process data – extracted from source systems (such as your ERP and CRM, courtesy of process mining) and user desktops (via task mining) to provide an objective end-to-end view of how work flows across the business.
- Business knowledge – KPIs, policies, goals, ownership, constraints, best practices, institutional know-how, and anything that contextualizes process data (see above) and helps AI understand the outcomes it’s aiming for.
When you see it laid out like that, it feels like it should be stating the obvious to say AI needs this surrounding context. But businesses began (and some, whisper it, still do) by developing and deploying AI solutions without it. It’d be like dropping you in a company and asking you to grow the business, without telling you what they do, their rules, KPIs, processes, and goals. You’d be able to recommend some generic growth levers, but nothing bespoke, nothing targeted.
Relevant context is what enables businesses to infuse AI with intelligence and orchestrate all their solutions at scale.
Different types of context in AI
Here comes that aforementioned array of different types of context for AI. But don’t worry: we’ve broken each of them down into a straightforward layman’s definition.
What is a context window?
A context window is the amount of information that large language models can consider at one time when generating an output. It includes the prompt, system instructions, conversation history, retrieved information, and other inputs provided to the AI model.
A larger context window size lets a model process more information at once – handy for long documents, extended conversations, and complex tasks. But for business use cases, the issue is often not “Can the model read more?” but “Does the model understand the business situation?” A bigger context window doesn’t automatically mean the model has relevant context. Better AI judgment comes from relevant, accurate, current, and well-structured information.
What is prompt context?
Prompt context is the information included directly in the instruction used to direct AI. It’s the extra detail and nuances provided to clarify their criteria (e.g. which industry or region it’s for), what they mean by a particular term or concept (e.g. how the business defines order cycle time), and any constraints (e.g. budget, timeframe, or deadline). Prompt context can also include examples, tone, and formatting instructions.
While this process of prompt engineering can improve the quality of AI responses, prompt context can be limited by what the user or system knows and remembers, which is unlikely to cover every relevant dependency, policy, KPI, exception, and downstream impact. AI might be prompted to “draft a response to a supplier about a delayed invoice”, for example, but unless the AI knows the invoice status, payment terms, supplier priority, goods receipt status, cash impact, and policy constraints, it can’t reliably take the best action.
What is retrieval context?
Retrieval context is information imported into an AI system from external sources – such as policies, product information, contracts, documents, knowledge bases, databases, search tools, or enterprise systems – to help ground the model’s response.
Retrieval-augmented generation (RAG) helps AI access relevant information that may not be in the model’s training data or current prompt. But retrieving documents doesn’t necessarily tell AI what’s happening right now across processes. RAG can retrieve a payment policy, for instance, but AI will still lack insight into which invoices are blocked and why, as well as which payment terms apply and which resolution path is allowed.
What is operational context?
Retrieval context and prompt engineering can give AI responses more information, but operational context helps AI reason and act with an understanding of the business, end to end. This eliminates AI’s operational blind spots and generalist decision-making, replacing them with a real-time representation of how the business runs, typically in the form of a digital twin.
Rather than just feeding AI more inputs, operational context translates that information and data into a language that AI understands and can use. By connecting live operational data, process history, relationships, rules, goals, constraints, rules, and current process states, AI can identify what follow-up action makes sense for any given scenario. With that, you have yourself truly autonomous and reliable AI – at scale.
How does context in AI work?
Let’s translate the theory of AI context into practice now, using an example use case from Finance.
- For our prompt context, AI is asked to explain and resolve a payment issue.
- The retrieval context would be the AI finding the business’ payment policy.
- The operational context would include the invoice block, vendor importance, goods receipt status, payment terms, working capital impact, so AI could determine a permissible resolution path.
The process of deciding what contextual information AI should receive, when it should receive it, and how that information should be structured, is called context engineering.
What is context engineering?
Effective context engineering goes beyond best practices for writing thorough prompts and core instructions. It can encompass retrieval, memory, tool outputs, user state, system instructions, formatting, structured data, and orchestration.
It’s particularly important for AI agents and multi-step workflows, which need to maintain process adherence and make decisions over time. AI hallucinations, irrelevant outputs, missed tool calls, and bad decisions can seriously add up here.
But context engineering falls short when it doesn’t consider what AI needs to understand about how the business runs. At Celonis, our Platform is designed to answer that question for every solution you deploy.
- Explore how context modeling works with our explainer post
How does Celonis help operationalize context for AI?
The Celonis Context Model (CCM) provides the operational context layer your people and Enterprise AI agents need to reason, decide, and act more reliably.
The Platform translates business reality into a language AI can understand. It does that by combining the process data and business knowledge we discussed earlier into a dynamic, real-time digital twin of business operations. But the CCM has something special up its sleeve, too: intelligence. Process intelligence fuels AI with hindsight and insight (about how the business currently runs), while decision intelligence drives foresight (into predictions about potential future states).
Context in AI might’ve begun as a technical discussion about prompts and token limits. But today, businesses have realized that AI agents and systems need more than LLM context windows, document retrieval, and prompt packaging. And Celonis can take you from talking the talk about operational context to walking the walk by embedding it in your Enterprise AI transformation.
Get in touch to take the first or next steps on your context-driven AI development journey.