> ## Content Index
> Fetch the complete content index at: https://blog.aperion.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Building Expert Systems with Celonis: Context, Authority and the Record
- URL: https://blog.aperion.ai/celonis-agent-governance-guide/
- Published: 2026-11-11T12:00:00.000Z
- Updated: 2026-11-11T12:00:00.000Z
- Description: Celonis supplies the knowledge base. An expert system that acts also needs a model chosen per call, a named person's authority and a record kept outside the agent. Seven steps.
- Author: Craig Alberino
- Tags: Guides, AI Agents, AI Governance, #pipeline

Updated November 11, 2026\. Part 9 of APERION's Agent Platforms series.

**Celonis supplies the knowledge base of an enterprise expert system. Its Context Model combines process data, business knowledge and intelligence from the systems where work runs. Its Agent Tools serve that context over MCP to agents built on platforms such as Claude, Microsoft Copilot Studio and Amazon Quick, and its Process Copilots and Orchestration Engine run agents inside Celonis. An expert system that acts in production needs three more things on every call: a model chosen for the data, the risk and the cost; the authority of the person the agent works for; and a record kept outside the agent. Controls that must survive a change to an agent's configuration belong on the path the calls take. Seven steps cover the build.**

Celonis, headquartered in Munich and New York, sells a platform that brings together process mining, process intelligence, decision intelligence and digital twin of the organization technology. It launched the Context Model on May 12, 2026, and describes it as combining "process data, business knowledge, and intelligence to create a living digital twin of the enterprise that AI can understand." Its MCP server for process intelligence launched in November 2025, the same month its Orchestration Engine became generally available. Celonis sums up its position as "no AI without PI." Celosphere 2026 runs December 1 to 3 in Munich. Facts in this guide were checked on October 20, 2026.

## What an Expert System Needs

An expert system emulates the decision-making of a human expert. The classic design pairs a knowledge base of facts and rules with an inference engine that applies them and can explain how it reached a conclusion.

MYCIN, built at Stanford in the early 1970s to recommend antibiotics for severe infections, ran on about 600 rules. In a blinded evaluation published in JAMA in 1979, eight infectious disease specialists from outside Stanford rated 65% of its prescriptions for ten meningitis patients acceptable; five Stanford faculty specialists, prescribing for the same patients, scored 42.5% to 62.5%. MYCIN was never used in routine care. It stood alone, so a user had to type all relevant information about a patient in answer to its questions, and observers raised ethical and legal questions about a physician's responsibility if it gave a wrong diagnosis.

XCON configured orders for Digital Equipment Corporation's VAX computers. It went into use at DEC in 1980, and by 1986 it had processed 80,000 orders with 95% to 98% accuracy. It was estimated to be saving DEC $25 million a year. A 1987 paper on its maintenance reported 6,200 rules, about half of which changed every year, and a system becoming more difficult to change.

Expert systems of that era ran into four problems: getting the knowledge in, keeping it current, connecting to the systems where work happens, and answering for what the system did.

The Context Model starts from what source systems record and adds what Celonis calls "the institutional know-how essential to every company." A language model can reason over cases no rule anticipated, and agents read the facts from the systems through tools. Answering for what the system did now comes up on every call: which model may see this data, whose authority the agent is using, and what record proves who decided.

| Part of the system | MYCIN and XCON                                                                  | An expert system built with Celonis                                             | What each call needs                                                             |
| ------------------ | ------------------------------------------------------------------------------- | ------------------------------------------------------------------------------- | -------------------------------------------------------------------------------- |
| Knowledge          | Rules entered by knowledge engineers: about 600 in MYCIN, 6,200 in XCON by 1987 | The Context Model, built from process data and business knowledge               | Data permissions that follow the person the agent works for                      |
| Reasoning          | A rule engine                                                                   | Models Celonis provides, or your own through bring your own model               | A model chosen for the data, the risk and the cost                               |
| Input              | A person typing answers to the system's questions                               | Agent Tools serving context over MCP                                            | One path for context, under one set of rules                                     |
| Action             | Advice on a screen; an order configuration                                      | Action Flows and Orchestration Engine steps that write to SAP and other systems | A named person's authority, checked before the action runs                       |
| Explanation        | MYCIN listed the questions and rules behind each conclusion                     | Process Copilot tool calls; Celonis audit and Action Flow logs                  | A record outside the agent tying context, model, policy and decision to a person |

![The expert system, 1979 and now: knowledge, reasoning, input, action and explanation compared across MYCIN and XCON, what is built with Celonis today, and what each call needs.](https://storage.ghost.io/c/56/dd/56dd2e81-b5ad-4845-bc3a-a53d4e7c40f8/content/images/2026/11/aperion-ap9-expert-system-1979-now-2026-11-05.webp)

Celonis supplies the knowledge base. Each call that acts needs three more things.

## What Celonis Ships for Agents

| Capability                 | What it does                                                                                                                                                                                                                                                   | What to know                                                                                                                                                                                                    |
| -------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Context Model              | Process data, business knowledge, and process and decision intelligence in one digital twin of operations. Launched May 12, 2026.                                                                                                                              | Celonis describes an open architecture that can integrate "any data source, AI model, or agent."                                                                                                                |
| Agent Tools (MCP)          | Serves Celonis data and actions to external agents over MCP or OpenAPI tool calling. Nine tools, among them Load Data, Search Data and Get Insights for reading, and Trigger Action Flow, Execute Orchestration and Resume Orchestration Execution for acting. | HTTP transport; OAuth 2.0 for production or application keys for testing; each client needs the use permission on the asset. Setup guides cover Claude, Claude Code, Microsoft Copilot Studio and Amazon Quick. |
| Process Copilots           | Agents built in Celonis Studio on a Knowledge Model, with tools switched on per copilot. They can run as AI Agent steps in the Orchestration Engine.                                                                                                           | Labeled Public Preview or limited availability. A copilot shows which tools it called, with which arguments, and lets each user query only data that user is permitted to see.                                  |
| Orchestration Engine       | Coordinates systems, applications, AI agents, bots and people in one process, in steps built from Action Flows, cloud functions or AI agents.                                                                                                                  | Generally available since November 2025\. People take part through Forms; a Resume Event pauses a process until a signal arrives.                                                                               |
| Action Flows               | Automations that write to source systems; SAP through Remote Function Calls and the Celonis Agent, an on-premises gateway.                                                                                                                                     | Started by data triggers, from Studio components, from a Process Copilot or through Trigger Action Flow. Some modules cannot be rolled back.                                                                    |
| Tasks                      | Bring a person into an automation that "requires the approval of a user, their manual input, or a decision before being executed."                                                                                                                             | An Action Flow can be connected to a Task.                                                                                                                                                                      |
| Bring your own model       | Admins connect models on AWS Bedrock, Azure OpenAI or an OpenAI-compatible API for their teams.                                                                                                                                                                | OAuth 2.0 client credentials. Celonis Support must allowlist a custom domain first.                                                                                                                             |
| Roles and data permissions | Admin, Analyst and Member roles; permissions at service, container and object level; data permissions on tables, columns and values.                                                                                                                           | A client that signs in as a user, such as the Claude connector, inherits that user's permissions.                                                                                                               |
| Audit logs                 | Team events with user role, date, message and, when enabled, IP address; CSV export or the Audit Log API, for a SIEM.                                                                                                                                          | Retention from 30 days to 10 years, or no deletion. Events include AI configuration changes, such as LLM providers, models and the AI kill switch, and agent registration; none is named for an MCP tool call.  |
| AI settings and SSO        | Admins switch AI services on or off, set per-asset consumption limits and can stop all AI services. SAML 2.0 single sign-on.                                                                                                                                   | Celonis says no third party may train generative AI models on customer inputs without express written consent.                                                                                                  |

## Where the Control Points Are

**Model calls.** Process Copilots run on the models enabled in a team's AI Settings. Bring your own model lets an admin add models on AWS Bedrock, Azure OpenAI or an OpenAI-compatible API, and an OpenAI-compatible endpoint added this way can be a model you host or a gateway that applies policy before the model. Agents built elsewhere call whatever model their platform is set to call: frameworks your teams run let you set that endpoint, and hosted agents keep model calls inside the vendor, as the [cross-platform guide](https://aperion.ai/blog/govern-ai-agents-across-platforms-guide/) maps.

**Context and actions.** External agents reach Celonis through Agent Tools, one endpoint per asset. One connection can carry reads and actions, and for Trigger Action Flow the model fills in the flow's declared inputs. Celonis controls who may call an asset through the OAuth scope, the use permission and the signed-in user's permissions. Inside Celonis, actions run as Action Flows and Orchestration Engine steps, and a Task can put a person in front of one.

**Records.** The records Celonis documents for agent work are the team audit log; Action Flow execution and configuration events, available through an API, which show who ran which flow, when, and what it did; and the tool calls a Process Copilot shows its user. SAP's own logs show the account Celonis used to connect. Tying an action to the person whose authority it carried, and to the model that proposed it, takes a record written on the call path.

## How to Build a Governed Expert System with Celonis in 7 Steps

### 1\. Start with one decision your experts make every day

Pick a decision Celonis already measures, with known rules and a visible cost when it goes wrong: releasing an order on credit block, paying an invoice that failed a three-way match, accepting a supplier's new delivery date. Write down who makes it today, under which limits, and what they check first. The agent will carry that person's authority.

### 2\. Put the knowledge in the Context Model, and scope who reads it

Model the objects, events and business rules the decision depends on, and the KPIs that show whether it went well. Set data permissions so each user, and each agent acting for one, sees the tables, columns and values that user may see. Give each integration its own OAuth client, as Celonis's setup guidance advises, so one leaked credential exposes one integration.

### 3\. Separate the tools that read from the tools that act

Enable on each Agent Tools asset the tools it needs and no others. An asset for analysis needs Load Data, Search Data and Get Insights. Put Trigger Action Flow, Execute Orchestration and Resume Orchestration Execution on a separate asset, with the use permission granted to the agents that act, so access to process data never brings the right to change it.

### 4\. Send every model call to an endpoint you control

Context from Celonis carries the enterprise's own process data, such as purchase orders, sales orders, prices and payment terms, and it travels in the prompt. Decide which models may receive which classes of data, then choose the model for each step by its data, its risk and its cost: one on infrastructure you control for steps that carry sensitive records, a frontier model for open-ended analysis of data allowed to leave. Point bring your own model, and the agents your teams build, at an endpoint that applies that rule.

### 5\. Bind every action to the authority of a named person

An agent that releases a payment or changes an order acts on someone's delegated authority. Record whose, and the limits that come with it: an amount, a supplier, a plant. Inside Celonis, connect the Action Flow to a Task where a person must approve. For external agents, check the call before Trigger Action Flow, Execute Orchestration or Resume Orchestration Execution reaches Celonis. An action inside the limits proceeds; one beyond them pauses until that person, or someone with approval rights, decides.

### 6\. Keep the record outside the agent

For each decision, keep the context the agent retrieved, the model that answered and where it ran, the policy that applied, what the agent proposed and who decided. Celonis's audit log and Action Flow events show who ran which flow, when, and what it did to the connected system. Write the rest at the model endpoint and the tool path, to your SIEM, with retention set to your longest regulatory obligation.

### 7\. Measure the decision in Celonis, then widen the authority

Celonis measures the process: cycle times, rework, duplicate payments, on-time delivery. The record from step 6 shows how often the agent's proposals matched what people decided. Raise an agent's limits one decision type at a time, where the process data and the decision record agree, and record each change.

## Operating Limits

- Celonis documents no approval step before an Agent Tools action runs. Approval for Trigger Action Flow, Execute Orchestration and Resume Orchestration Execution belongs on the tool path, or in a Task.
- The routes Celonis documents from a Process Copilot to other systems are Action Flows and Orchestration Engine steps.
- Celonis's audit log events cover AI configuration changes, such as LLM providers, models and the AI kill switch. None is named for an MCP tool call.
- Celonis's pages do not say whether a Process Copilot can select a model added through bring your own model. Confirm with Celonis.
- Availability labels differ. Celonis's February 2026 release notes call Agent Tools generally available, while its docs page carries a Private Preview notice, and Process Copilots are labeled Public Preview or limited availability. Confirm with Celonis what your team has.
- SAP's logs show the account Celonis connected with, so the person behind an action has to come from Celonis's records or yours.

## How APERION Fits

APERION is runtime governance for AI agents. It enforces policies and controls in real time across an enterprise's AI agents, maintains human accountability, and captures auditable evidence for every decision.

Owned intelligence, governed from the first call. APERION's Smartflow serves OpenAI-compatible and Anthropic-compatible endpoints, so an agent built on a framework your team runs can point its model calls at it, and its MCP gateway can front the MCP servers the agent calls, Celonis Agent Tools among them. Each agent gets its own key with model and route allowlists, so rented and owned models answer under one policy. Tool rules apply to every MCP call, an out-of-authority call goes to a person with approval rights, and every call lands in a hash-chained audit record your SIEM can ingest. Smartflow runs on-premises, including air-gapped.

A verified person, the authority that person delegated, the agent that used it, the policy that applied, the decision and the record of all of it. That chain is the product. [Talk to APERION](https://aperion.ai/contact).

## More in This Series

- [How to govern AI agents across platforms](https://aperion.ai/blog/govern-ai-agents-across-platforms-guide/)
- [CrewAI security and governance](https://aperion.ai/blog/crewai-security-governance-guide/)
- [Glean agents governance](https://aperion.ai/blog/glean-agents-governance-guide/)
- [Agentforce governance](https://aperion.ai/blog/agentforce-governance-guide/)
- [Amazon Bedrock AgentCore governance](https://aperion.ai/blog/aws-bedrock-agentcore-governance-guide/)
- [Microsoft Copilot agent governance](https://aperion.ai/blog/microsoft-copilot-agent-governance-guide/)
- [Google, ServiceNow, OpenAI, Anthropic and LangChain](https://aperion.ai/blog/google-servicenow-openai-anthropic-agent-governance-guide/)

[Watch all seven videos: the Agent Platforms playlist on YouTube](https://www.youtube.com/playlist?list=PLHG8axy3%5FFI0)

Related guides: [the AI agent governance guide](https://aperion.ai/blog/ai-agent-governance-guide/).

## Frequently Asked Questions

### What is an expert system?

An expert system is software that emulates the decision-making of a human expert. The classic design pairs a knowledge base of facts and rules with an inference engine that applies them and can explain its reasoning. MYCIN, built at Stanford in the 1970s to recommend antibiotics, is an early example. An expert system built with Celonis uses the Context Model as its knowledge base and a language model as its engine.

### Why did the first expert systems stall?

They ran into four problems: entering expert knowledge by hand, keeping it current, connecting to the systems where work happened, and answering for their advice. In a blinded evaluation published in 1979, MYCIN's prescriptions were rated acceptable 65% of the time, against 42.5% to 62.5% for Stanford faculty, yet it was never used in routine care. By 1987, about half of XCON's 6,200 rules changed every year.

### What is the Celonis Context Model?

The Context Model, launched on May 12, 2026, is the core of the Celonis Platform. Celonis says it combines process data, business knowledge and intelligence to create a living digital twin of the enterprise that AI can understand. In an expert system it plays the part of the knowledge base: what an agent consults about how the business runs before it reasons or acts.

### Does Celonis support MCP?

Yes. Celonis Agent Tools, formerly the MCP Server asset, serve Celonis data and actions to MCP clients over HTTP, and through OpenAPI tool calling. The nine tools include Load Data, Search Data and Get Insights for reading, and Trigger Action Flow and Execute Orchestration for acting. Clients authenticate with OAuth 2.0 or application keys and need the use permission on the asset.

### Can I bring my own model to Celonis?

Yes, through the bring your own model option in AI Settings. Admins connect models on AWS Bedrock, Azure OpenAI or an OpenAI-compatible API, authenticated with OAuth 2.0 client credentials, and make them available to their teams. Celonis Support must first allowlist a custom domain. Celonis's pages do not say which features, Process Copilots included, can select these models.

### Can a Celonis agent take action in SAP?

Yes, through Action Flows. Celonis reaches SAP through Remote Function Calls and the Celonis Agent, an on-premises gateway, with pre-built actions for objects such as purchase orders and sales orders. A Process Copilot shows its user a button to run an Action Flow, and an external agent can call Trigger Action Flow through Agent Tools. A Task can require a person's approval before an automation runs.

### How do you keep an agent inside a person's authority?

Record whose authority each agent carries and its limits, such as an amount, a supplier or a plant. Check each action against those limits on the path it takes, before it executes. An action inside them proceeds; one beyond them pauses until that person, or someone with approval rights, decides. Keep the decision in a record outside the agent, where a change to its instructions cannot remove it.

## Sources

- Celonis, [Celonis Launches the Context Model to Eliminate Enterprise AI's Operational Blind Spots, Agrees to Acquire AI Decision Intelligence Leader Ikigai Labs](https://www.celonis.com/news/press/celonis-launches-the-context-model-to-eliminate-enterprise-ais-operational-blind-spots-agrees-to-acquire-ai-decision-intelligence-leader-ikigai-labs), May 12, 2026\. Opened October 20, 2026.
- Celonis, [Global Enterprises Trust Celonis to Transform Operations, Industrialize their AI, and Achieve Billions in Value](https://www.celonis.com/news/press/global-enterprises-trust-celonis-to-transform-operations-industrialize-their-ai-and-achieve-billions-in-value), July 16, 2026, including the company description. Opened October 20, 2026.
- Celonis, [The Celonis Context Model](https://www.celonis.com/platform/context-model) and [Celonis Platform](https://www.celonis.com/platform). Opened October 20, 2026.
- Celonis, [Celonis Unveils Platform Innovations to Power the AI-Driven, Composable Enterprise](https://www.celonis.com/news/press/celonis-unveils-platform-innovations-to-power-the-ai-driven-composable-enterprise), November 4, 2025\. Opened October 20, 2026.
- Celonis, [Celonis Showcases Latest Process Intelligence and AI Innovations at Next 2025](https://www.celonis.com/news/press/celonis-showcases-latest-process-intelligence-and-ai-innovations-at-next-2025), May 13, 2025\. Opened October 20, 2026.
- Celonis Developer Center, [Agent Tools (MCP) Asset](https://developer.celonis.com/mcp-server-asset/overview/), [Tools for MCP Server/Agent Tools assets](https://developer.celonis.com/mcp-server-asset/tools/tools%5Ffor%5Fmcp%5Fserver), [Getting Started](https://developer.celonis.com/mcp-server-asset/getting-started), [Celonis Agent Tools (MCP Server) in Claude](https://developer.celonis.com/mcp-server-asset/integrations/claude-connector-agent-tools), [Celonis MCP Server Integration with Amazon Quick](https://developer.celonis.com/mcp-server-asset/integrations/mcp-amazon-quick), [Celonis MCP Server in Microsoft Copilot Studio](https://developer.celonis.com/mcp-server-asset/integrations/msft-connector) and [Audit Log API](https://developer.celonis.com/celonis-apis/audit-log-api/overview). Opened October 20, 2026.
- Celonis Docs, [Agent Tools (formerly MCP Server Asset)](https://docs.celonis.com/en/agent-tools--formerly-mcp-server-asset-.html) and [February 2026 Release Notes](https://docs.celonis.com/en/february-2026-release-notes.html). Opened October 20, 2026.
- Celonis Docs, [Process Copilots](https://docs.celonis.com/en/process-copilot.html), [Set up Process Copilots](https://docs.celonis.com/en/set-up-process-copilots.html), [System Card: Celonis Process Copilot](https://docs.celonis.com/en/system-card--celonis-process-copilot.html), [Trigger an Action Flow from a Process Copilot](https://docs.celonis.com/en/trigger-an-action-flow-from-a-process-copilot.html), [Adding Process Copilot to Process Orchestration steps](https://docs.celonis.com/en/adding-process-copilot-to-process-orchestration-steps.html) and [Creating Orchestration Engine overview](https://docs.celonis.com/en/creating-orchestration-engine-overview.html). Opened October 20, 2026.
- Celonis Docs, [SAP Action Flows](https://docs.celonis.com/en/sap-action-flows.html), [Action Flow execution, cycles, and phases](https://docs.celonis.com/en/action-flow-execution,-cycles,-and-phases.html), [Monitoring and auditing Action Flows](https://docs.celonis.com/en/monitoring-and-auditing-action-flows.html) and [Tasks](https://docs.celonis.com/en/tasks.html). Opened October 20, 2026.
- Celonis Docs, [Managing Celonis Platform permissions](https://docs.celonis.com/en/managing-celonis-platform-permissions.html), [Understanding data permissions](https://docs.celonis.com/en/understanding-data-permissions.html), [Viewing your audit logs](https://docs.celonis.com/en/viewing-your-audit-logs.html), [Audit logs - Event details](https://docs.celonis.com/en/audit-logs---event-details.html), [Configuring SAML single sign-on](https://docs.celonis.com/en/configuring-saml-single-sign-on.html) and [AI and LLM Settings](https://docs.celonis.com/en/ai-settings.html). Opened October 20, 2026.
- Celonis Documentation, [Orchestration Engine](https://documentation.celonis.com/en/orchestration-engine.html), [AI Settings](https://documentation.celonis.com/en/ai-settings.html), for bring your own model, and [2026 release notes](https://documentation.celonis.com/en/2026-release-notes.html). Opened October 20, 2026.
- Celonis, [Responsible AI](https://www.celonis.com/legal/trust-center/responsible-ai), [Celosphere 2026](https://www.celonis.com/celosphere) and the [celonis.com](https://www.celonis.com/) homepage. Opened October 20, 2026.
- Victor L. Yu, Lawrence M. Fagan, Sharon Wraith Bennett, William J. Clancey, A. Carlisle Scott, John F. Hannigan, Robert L. Blum, Bruce G. Buchanan and Stanley N. Cohen, ["An Evaluation of MYCIN's Advice"](https://people.dbmi.columbia.edu/shortliffe/Buchanan-Shortliffe-1984/Chapter-31.pdf), chapter 31 in Bruce G. Buchanan and Edward H. Shortliffe, eds., *Rule-Based Expert Systems* (1984), an edited version of the article in JAMA 242:1279-1282 (1979). Opened October 20, 2026.
- Elliot Soloway, Judy Bachant and Keith Jensen, [Assessing the Maintainability of XCON-in-RIME: Coping with the Problems of a VERY Large Rule-Base](https://mlanthology.org/aaai/1987/soloway1987aaai-assessing), AAAI 1987\. Opened October 20, 2026.
- Wikipedia, [Expert system](https://en.wikipedia.org/wiki/Expert%5Fsystem), [Mycin](https://en.wikipedia.org/wiki/Mycin) and [Xcon](https://en.wikipedia.org/wiki/Xcon). Opened October 20, 2026.