The process before automation
Do not automate an unclear process; make decisions and exceptions visible first.
Map the trigger, input, transformations, decisions, actions, systems and waiting times. Record where value is created and where handovers lose information or efficiency.
Remove unnecessary steps before technology. Measure variation and exceptions; an unstable process requires standardization or limited assistance first.
Then choose how much technology is needed. With manually supplied source context, you paste approved text into the chat yourself, such as Noor’s fact sheet. With RAG (retrieval-augmented generation), a system first retrieves relevant source passages and uses them in its answer; check both the selection and the answer.
An agent can choose follow-up steps or tools within the permissions granted. MCP is a protocol through which applications can offer tools and context; it is not itself an agent or a quality guarantee. These concepts can be combined. A read-only connection carries different risks from one that sends messages.
- Trigger
- Input
- Decision
- System
- Exception
- Waiting time
Terms in plain language
- As-is / to-be
- The current way of working and the proposed way of working.
- RAG
- Building answers using source passages that a system has first retrieved.
- Agent
- An AI system that can choose follow-up steps or tools to carry out a goal, within configured limits.
- MCP
- Model Context Protocol: a standardised way to connect tools and context to AI applications.
An invoice flow slows down due to unclear approval boundaries, not because of text entry.
Model [process] as-is and to-be with trigger, steps, decisions, systems, waiting time, errors, and exceptions. Eliminate waste before AI.Draw the process for Noor’s draft replies. For the basic route, create ten fictional case cards: four ordinary product questions, two without a suitable source, two with conflicting source information, one with a missing link to the question, and one with a timeout. Walk through each card on paper and count the routes. Completion check: the trigger, review and fallback are recognisable and the counts agree. These are design simulations, not observed business cases. Real observation is a later step before actual implementation.
Source for this lesson
OpenAI — Retrieval
Technical background on retrieving relevant source passages. The course route uses manually supplied text and does not require this API.
Checked: 2026-09-08
OpenAI — MCP and Connectors
Connections to external services, tools and approvals; conceptual explanation, with no mandatory integration.
Checked: 2026-09-08