> ## 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.

# Expert systems are back. Who answers for them?
- URL: https://blog.aperion.ai/expert-systems-are-back/
- Published: 2026-10-22T11:00:00.000Z
- Updated: 2026-10-22T10:59:59.000Z
- Description: In a study published in 1979, MYCIN's prescriptions outscored those of five Stanford specialists, and MYCIN was never used in routine care. Celonis and agents answer three of the four problems early expert systems ran into. The fourth comes up on every call.
- Author: Craig Alberino
- Tags: AI Agents, Runtime Governance, Enterprise AI, #pipeline

In a study published in 1979, eight infectious disease specialists from outside Stanford reviewed antibiotic prescriptions for ten patients with meningitis without knowing who had written them, or that one of the prescribers was a computer program. They rated 65% of the program's prescriptions acceptable. Five Stanford faculty specialists, prescribing for the same patients, scored 42.5% to 62.5%.

The program was [MYCIN](https://people.dbmi.columbia.edu/shortliffe/Buchanan-Shortliffe-1984/Chapter-31.pdf). It was never used in routine care.

## Four problems, two systems

MYCIN stood alone. A clinician had to type every relevant fact about a patient in answer to its questions, and observers asked who would be responsible if it gave a wrong diagnosis.

XCON got further. It configured orders for Digital Equipment Corporation's VAX computers from 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](https://mlanthology.org/aaai/1987/soloway1987aaai-assessing) reported 6,200 rules, about half of which changed every year, and a system becoming more difficult to change.

Between them, the two 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.

| Problem                   | MYCIN and XCON                                                   | With Celonis and agents                                                          |
| ------------------------- | ---------------------------------------------------------------- | -------------------------------------------------------------------------------- |
| Getting the knowledge in  | Rules entered by hand: about 600 in MYCIN, 6,200 in XCON by 1987 | The Context Model, built from source-system process data plus business knowledge |
| Keeping it current        | About half of XCON's rules changed every year                    | A digital twin Celonis describes as dynamic and real-time                        |
| Connecting to the work    | A clinician typed every relevant fact into MYCIN                 | Agents read context over MCP and act through Action Flows                        |
| Answering for what it did | Observers asked who would be responsible for a wrong diagnosis   | Raised on every call: which model, whose authority, what record                  |

## What changed

Celonis builds the knowledge base from the systems themselves. Its [Context Model](https://www.celonis.com/platform/context-model), [launched in May](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), starts from the process data in ERP, CRM and other source systems and adds business knowledge, which Celonis calls "the institutional know-how essential to every company." Celonis describes the result as "a dynamic, real-time digital twin of operations."

A language model can reason over cases no rule anticipated.

Celonis [Agent Tools](https://developer.celonis.com/mcp-server-asset/overview/) give agents access to Celonis data and insights over MCP, and Celonis publishes setup guides for [Claude](https://developer.celonis.com/mcp-server-asset/integrations/claude-connector-agent-tools), [Microsoft Copilot Studio](https://developer.celonis.com/mcp-server-asset/integrations/msft-connector) and [Amazon Quick](https://developer.celonis.com/mcp-server-asset/integrations/mcp-amazon-quick). The same tools can [trigger an Action Flow](https://developer.celonis.com/mcp-server-asset/tools/tools%5Ffor%5Fmcp%5Fserver) that writes to SAP, or start a process in Celonis's Orchestration Engine. The agent reads the facts through those tools and acts on them.

Three of the four problems now have answers.

## The fourth comes up on every call

MYCIN gave advice, and a physician decided. An agent with Agent Tools acts. When it calls Trigger Action Flow, the model fills in the flow's inputs, and the flow can write to SAP. Three questions come up on that call, and on every call like it.

**Which model may see this data?** Context from Celonis carries the enterprise's own process data: purchase orders, sales orders, prices, payment terms. It travels in the prompt to whichever model the agent is set to call, rented from a provider or running on infrastructure the enterprise controls.

**Whose authority is the agent using?** A payment release or an order change happens on someone's delegated authority, up to that person's limits. When an Action Flow writes to SAP, [SAP's logs show the account Celonis used to connect](https://docs.celonis.com/en/monitoring-and-auditing-action-flows.html). The enterprise has to record whose authority the agent carried.

**What record proves who decided?** Celonis logs who executed each Action Flow and when, and what the flow did to the connected system. Tying an action to the model that proposed it, the policy that applied and the person who approved it takes one record that covers all three.

Inside Celonis, a [Task](https://docs.celonis.com/en/tasks.html) can require a person's approval before an automation runs. An agent built outside Celonis needs the same approval for the same actions.

## Owned intelligence, governed from the first call

Enterprises will rent some intelligence and own more of it. An expert system built on Celonis can use both: Celonis's [bring your own model](https://documentation.celonis.com/en/ai-settings.html) option accepts models on AWS Bedrock, Azure OpenAI or an OpenAI-compatible API.

The knowledge in that system is the enterprise's process data and know-how. The choice of model, the authority each agent carries and the record of who decided belong with the enterprise too, at one control point inside its walls, from the first call. The accountable human decides.

*Related: [How to govern AI agents across platforms](https://aperion.ai/blog/govern-ai-agents-across-platforms-guide/) and [AI agent governance: MCP, A2A and the new attack surface](https://aperion.ai/blog/ai-agent-governance-guide/).*