AI Automation & Agents

• What makes a good automation candidate: frequency versus value
• Rule-based work, language and judgement work, and human decisions
• Mapping the current process: triggers, inputs, actions and outputs
• Bottlenecks and handoffs
• Whether it is worth building, and where automation creates new risks

Exercise: map a repetitive process from start to finish before touching an automation tool.

• How visual automation platforms work: triggers, actions, filters and branches
• Variables, fields and structured data
• Connecting common workplace tools and scheduling
• Webhooks in plain English, credentials and account permissions
• Testing each step separately

Exercise: build a multi-step workflow that moves real information between at least two services.

• Classification, information extraction and summarisation
• Rewriting and drafting
• Structured AI output and routing based on it
• Supplying context and working with reference material
• Confidence, uncertainty and human review

Exercise: add an AI decision or transformation step and test it against several different inputs.

• Automation versus agent: goals, tools, context and memory
• Loops, planning and tool selection
• Permissions and human-in-the-loop design
• When an agent is unnecessary, and simple agent patterns

Exercise: create a constrained agent that chooses between a small set of tools or actions to complete a defined task.

• Failure states, invalid data, timeouts and duplicate actions
• Logging, error paths, notifications and manual fallback
• Sensitive information, prompt injection awareness, least-access permissions
• Cost control, monitoring, documentation and maintenance when a tool changes

Final project: one automation from your real work, for example enquiry to classification to draft reply to human approval, or meeting transcript to decisions to assigned tasks. It needs a clear trigger, at least two connected systems, an AI step that adds real value, human approval where appropriate, error handling and documentation.

An agent is not automatically better than a workflow. If a process has five known steps, build five known steps. Use more agentic behaviour when the system genuinely needs to choose between actions, search for information or adapt its route to reach a goal. Learning the difference saves a surprising amount of unnecessary complexity.

You work on a real process from your own work wherever possible. We use visual tools for most of the course; you will meet ideas such as APIs, webhooks and JSON, explained through the task.

AI tools change quickly. We review the teaching stack before each cohort; it typically includes n8n, Make, Zapier, the ChatGPT, Claude and Gemini APIs or connectors, and Google Workspace.