Celonis

Data comes in from any source, and actions land in connected systems. The Data Core "lets you extract data from any source and query billions of records at speed", and the Context Model is system agnostic, built from data across "systems, applications, devices, and interactions", with a Marketplace in the platform navigation.

Celonis also documents ingestion from SAP and Oracle, action flows that execute in third party systems, and a Celonis MCP Server that exposes process tools to agents. Its integrations that feed process context to agents built in Copilot Studio, watsonx Orchestrate and Bedrock are Celonis integrations, not capabilities of those platforms.

The connector catalog and action flow reference require a login, and no connector count is published.

Sourcecelonis.com/platform, home; July first-party blog and press read; readread 2026-09-17

AI driven processes on the platform combine agents, people and systems, and customers compose them in the Build Experience, which lets them "analyze, design, and operate composable, AI-driven business processes" and "build and run strategic, operational, business-critical solutions", with Design used to "define workflows, outcomes, guardrails, and best practices" and Operate to "orchestrate AI to work alongside people and systems" while monitoring adherence. The acquired Orchestration Engine handles end to end process workflows, and action flows execute the steps.

Branching, retry and versioning specifics are not on the public pages.

Sourcecelonis.com/platform; July first-party blog and press read; readread 2026-09-17

Agents ground on the Celonis Context Model (formerly the Process Intelligence Graph), a continuously refreshed model of the customer's business.

It "creates a dynamic, system-agnostic, real-time representation of your entire business, built on process data and business knowledge from across systems, applications, devices, and interactions", and "provides intelligence to determine root causes, generate predictions and recommendations, evaluate what-if scenarios".

It keeps improving, "mining new processes and agents as they execute", so AI agents have "the operational clarity to reason correctly". Beneath it, the Data Core extracts data from any source and queries billions of records. Knowledge enters by ingestion, not retraining.

Permissions per source and citation behavior are not publicly documented.

Sourcecelonis.com/platform, /solutions/ai, home; readread 2026-09-17

Customers define guardrails, but no approval step is documented on Celonis's own pages. The platform's Design capability is to "redesign operations to integrate enterprise AI... define workflows, outcomes, guardrails, and best practices", and Operate orchestrates AI to "work alongside people and systems" while monitoring adherence.

Where an approval would sit (node, workflow or policy layer), and whether autonomy can vary by risk, is not published. Governance planes in partner platforms, such as Microsoft's Agent 365, belong to those partners.

Sourcecelonis.com/platform, /solutions/ai; readread 2026-09-17

Celonis publishes its attestations and access controls on its own trust portal. The trust portal (hosted on SafeBase) lists ISO/IEC 27001 with a Statement of Applicability, ISO/IEC 27701, ISO 9001:2015, SOC 1, SOC 2, the CSA STAR registry, TISAX, Cyber Essentials and Cyber Essentials Plus, plus GDPR and HIPAA positions, with documents available on request and a SOC 2 report featured. Its product security section lists SSO and SAML, role based access and team management, alongside penetration testing and a subprocessor list.

The portal's detail pages and reports are available only with access, and the product documentation sits behind a login. Governance planes in partner platforms, such as Microsoft Agent 365 and Amazon Bedrock AgentCore, are not Celonis's own controls.

Sourcetrust.celonis.com; celonis.com/legal/trust-center; readread 2026-09-17

Celonis monitors agent activity inside the process record. The platform's Operate capability is to "continuously monitor process performance, adherence, and agent activity" and to orchestrate AI to work alongside people and systems, and the Context Model "mines new processes and agents as they execute", learning from how people and agents interact in real time. Each executed step is mined into the process record as it happens.

No per run trace view, prompt log or tool call inspector is publicly documented. Tracing in AWS AgentCore and observation in Microsoft Agent 365 belong to those partners.

Sourcecelonis.com/platform, home; readread 2026-09-17

Whether Celonis agents keep memory is not settled. The Context Model persists a digital twin of the business with agent outcomes written back, but that is the business record, not agent memory.

Celonis describes it as "continuously improving, mining new processes and agents as they execute, and learning from how your people and agents interact in real time": adaptation at the platform level, with no memory scope, lifetime or review surface for an individual agent.

The product documentation requires a login, and nothing public describes session or long term memory for the AI Assistant Builder or its agents.

Sourcedocs.celonis.com (login-gated); celonis.com/platform; triedread 2026-09-17

Delivery is cloud SaaS, and no hosting region, residency option or customer environment is published. The home, platform and AI solution pages describe the Data Core, Context Model and Build Experience as a cloud platform, and on premises, hybrid and residency options are not detailed. The trust portal's readable overview lists compliance frameworks and a subprocessor entry but no region choice.

The trust portal's detail sections and the product documentation both require access.

Sourcecelonis.com/platform, /solutions/ai; trust.celonis.com; readread 2026-09-17

Packaged solutions and partner built agents are both available to adopt. The platform navigation carries a Marketplace, and the Build Experience ships strategic, operational and business critical solutions a customer runs. Through AgentC, customers can use agents pre built by partners such as Rollio and Hypatos, alongside Solution Suites and assistant templates in the AI Assistant Builder. Each works on its own when adopted.

Sourcecelonis.com/platform; AgentC blog and press (read 8 July 2026); readread 2026-09-17

Work reaches Celonis agents from process events, without a person asking. Celonis's AgentC blog says agents act on process events, for example identifying orders ready for scheduling, and execute action flows against connected systems. The platform pages add that the Context Model "mines new processes and agents as they execute" in real time, and that Operate orchestrates AI alongside people and systems. Channels such as Teams and Microsoft 365 belong to partner surfaces.

The trigger catalog itself sits behind a login.

Sourcecelonis.com/blog AgentC post (read 8 July 2026); celonis.com/platform; datedread 2026-09-17

Whether Celonis customers choose a model is not settled. Celonis describes itself as platform and model agnostic because agents built in Microsoft Copilot Studio, IBM watsonx Orchestrate and Amazon Bedrock Agents choose their own models, but that choice belongs to those platforms, not to Celonis. The public pages describe the Context Model, the Data Core and the Build Experience but name no model, provider, routing or selection surface for Celonis's own AI (the AI Assistant Builder), and the product documentation is behind a login.

Sourcedocs.celonis.com (login-gated); celonis.com/platform, /solutions/ai; triedread 2026-09-17

Agents built elsewhere can call Celonis. Celonis's own AgentC announcement and blog say a Process Intelligence API opens the platform and a Celonis MCP Server "exposes process tools to agents", and AgentC supplies tooling and integrations for building agents in third party platforms.

Neither the API nor the MCP reference is public.

Sourcecelonis.com/news/press AgentC release and /blog AgentC post (read 8 July 2026); docs.celonis.com login-gated; datedread 2026-09-17

Processes can be simulated, but no way to test an agent's own behavior is documented. The Analyze capability lets customers "run process simulations, predictions, and what-if scenarios" and find where to deploy AI, and Operate continuously monitors process performance, adherence and agent activity. Simulation tests the process and forecasts outcomes rather than scoring or comparing what an agent does, and adherence monitoring after deployment watches operations rather than gating a change to the agent.

Sourcecelonis.com/platform, /solutions/ai; readread 2026-09-17

The agents work through data and integrations, not screens. Celonis extracts data from source systems, holds it in the Data Core, reasons over the Context Model, and acts through action flows and integrated platforms. Nothing published describes an agent operating a browser, a desktop session or another application's screens, and Celonis ships no bot or screen automation estate, so breakage when an interface changes is not a concern here.

Sourcecelonis.com/platform, /solutions/ai, home; readread 2026-09-17