Teaching / Course 04
AI Automation
From process maps to agent-built automations a business can run
How to find the work in a business that is worth automating, have AI agents build the automation, and show afterwards that it saves more than it costs.
A practical course for people who automate business processes: choosing what to automate on evidence, building integrations with coding agents rather than by hand, and running the result with monitoring, cost control and a handover a team can own.
- 3 days · 21 contact hours
- Operations leads, automation specialists and consultants, analysts, technical product managers and engineers moving into automation.
- One real or realistic business process is mapped, scored, automated, hardened and handed over across the three days.
- One automated business process with a measured baseline, run cost, error handling and a handover pack.
Feasible when every team automates one process end to end against sandbox systems. Integrating with live company systems, or comparing every automation platform hands-on, would not fit into three days.
Outcomes
What a participant can do afterwards.
- Map a business process into tasks and score each task by frequency, time, cost of errors, data access and how easily its output can be checked.
- Write a business case for an automation: hours saved, the value of that time, build and run cost, maintenance and break-even, with the assumptions stated.
- Decide when a drag-and-drop builder (Zapier, Make, n8n, Power Automate), a coding agent (Claude Code, Codex), a small internal app or no automation at all is the right choice.
- Direct a coding agent to build integrations against APIs, webhooks and MCP connectors from a written specification, and review what it produced.
- Design approval points, idempotent steps, retries and fallbacks, so that a failed run is noticed and can be repeated safely.
- Produce automation results beyond text, such as filled documents, slide decks, images and short videos, and check them before they reach a customer.
- Monitor an automation in production through success rate, exceptions, cost per run and drift, with a named owner and an escalation path.
- Scope, price and hand over automation work with a proposal that states measurable outcomes, data-protection duties under the GDPR and the EU AI Act obligations that apply.
Depth
Realistic depth within three days.
The difference between these three columns is the difference between a course description and a course. Everything in the third column is real content; it simply isn't something anyone builds from scratch in 21 hours.
Implemented
Written by participants, from scratch or near it.
- Process audit and automation scorecard
- End-to-end automation built with a coding agent
- Monitoring, alerts and a cost report
Practised
Run, tuned and measured against prepared scaffolding.
- Agent-written API and MCP integrations
- Approval checkpoints, retries and idempotency
- Proposal, pricing and handover pack
Surveyed
Explained and demonstrated. Not built in three days.
- Enterprise RPA and process-mining suites
- Every drag-and-drop platform in depth
- Custom model training or fine-tuning
Day 1
Choosing work worth automating
Decide what to automate on evidence before building anything.
-
Orientation
Automation as the use of technology to reduce labour; what changed once agents started writing integrations; the course process and the capstone.
-
What agents automate today
Live demonstration: a coding agent builds a working integration from a one-paragraph request. Where it succeeds, where it fails, and why checking the result is now most of the work.
-
Process discovery
Interview a process owner, map the process, split it into tasks and record volumes, durations, error rates and the systems each task touches.
-
Scoring tasks
Frequency, time per case, cost of an error, data access, how checkable the output is and how much judgement it needs. Why digitised, frequent and verifiable tasks go first.
-
The economics of an automation
Hours saved and what they are worth, build cost, cost per run, maintenance, the cost of failures and break-even. Why most of the value usually sits in a few tasks.
-
Choosing the tool
Visual builders, coding agents, small internal web apps, off-the-shelf software, or leaving the task with people. Lock-in, ownership and who maintains the result.
-
Lab 1: business case
Score the task inventory, choose the automation target and write a one-page business case with its assumptions stated.
A process map, a scored task inventory and a one-page business case for the process the team will automate.
Day 2
Building automations with agents
Turn the chosen process into a working automation that an agent builds and a person reviews.
-
Writing the specification
Inputs, outputs, systems, edge cases, approval points and acceptance tests, written so that an agent can build from them and a reviewer can check against them.
-
Integration building blocks
APIs, authentication, webhooks, schedules, queues, MCP connectors and skills. Structured outputs and schema validation between steps.
-
Lab 2: agent-built integration
Direct a coding agent to build the first half of the process against the sandbox (for example intake, enrichment and classification), then review and test its code.
-
People in the loop
Approval checkpoints, confidence thresholds, escalation and audit trails, and how to keep an approver from becoming a rubber stamp.
-
Reliability
Idempotent steps, retries, timeouts, dead-letter queues, partial failures and re-running a single case safely.
-
Outputs beyond text
Generating documents, slide decks, images and short videos as automation results; templates and brand rules; checking generated media before it is sent.
-
Lab 3: completing the flow
Add the second half (for example a drafted proposal and a CRM or ticket update), the approval step and tests for every edge case in the specification.
A working end-to-end automation in the sandbox, with its specification, tests and an approval step.
Day 3
Running automations in a business
Turn a working automation into something a team can own, measure and pay for.
-
Monitoring and evaluation
Success rates, exception queues, sampling outputs for review, cost per run, and drift when inputs or upstream systems change.
-
Data protection and regulation
GDPR roles and data minimisation, processing agreements, personal data in logs, and the EU AI Act obligations that apply to organisations deploying AI systems.
-
Lab 4: monitoring and cost
Add logging, alerts and a cost report to the capstone and run it on a batch of test cases, including deliberately broken ones.
-
Scoping and pricing
Fixed-scope, time-based and value-based pricing; a proposal that names measurable outcomes, exclusions and maintenance; why a small scope that ships beats a large one that doesn't.
-
Handover and adoption
Runbooks, ownership, access, change requests, training the people whose work changes, and measuring the result after launch.
-
Capstone completion
Finish the handover pack and the measured before-and-after comparison.
-
Demonstrations and review
Team demonstrations against the business case, a cost and failure review, and next-step recommendations.
A monitored automation with a cost report, a handover pack and a proposal or internal decision paper.
Labs
Four deliverables, each one the input to the next.
- Process audit Process map, scored task inventory and a one-page business case.
- Agent-built integration Specification, agent-written integration code, review notes and tests.
- Approval and reliability The complete flow with an approval step, retries, idempotent steps and edge-case tests.
- Operations Monitoring, alerts, a cost report, the handover pack and a proposal.
Capstone specification
- A process map and scored task inventory that explain why this process was chosen.
- A baseline measured or estimated before automation: volume, time per case, error rate and cost.
- An automation built mainly by a coding agent from a written specification, with the specification and the review notes included.
- At least one human approval point, and documented behaviour for every failure the specification names.
- Idempotent steps, so any failed case can be re-run without duplicate side effects.
- A cost-per-run figure and a break-even estimate based on the baseline.
- A handover pack: runbook, owner, access list, monitoring view and data-protection notes.
Stack
The tools, and why each one is in the room.
- Mapping template and scoring sheet Most automation mistakes happen before anything is built; a simple scorecard makes the choice explicit and reviewable.
- Claude Code or Codex Writes integrations from a specification, which is how most new automations get built now. One is used hands-on, the other is compared.
- n8n (self-hosted) Shows where a visual workflow is still the clearer choice, and gives a fair comparison with agent-written code.
- REST APIs, webhooks and MCP servers The common ground every platform builds on; MCP lets agents and tools share the same connectors.
- Mock CRM, ticketing system and mailbox Realistic side effects without touching production data.
- Document and slide templates, image and video generation APIs More and more automation results are documents and media, not just fields in a database.
- Structured logs, alerts and a cost report An automation nobody watches fails silently, and cost per run decides whether it pays.
Discussed, not taught hands-on
- Zapier, Make and Power Automate are discussed and compared but not installed; n8n stands in for the visual builders.
- Enterprise RPA and process-mining suites are surveyed for context. They suit large, stable process estates more than the fast-changing processes this course targets.
- Lead generation and sales for automation agencies are covered only as far as scoping and pricing need them.
Logistics
Who it is for, and what it needs.
- Process and operations owners, automation specialists, freelance and agency automation consultants, business analysts, technical product managers and engineers who are asked to automate business work.
- Comfort with spreadsheets, web tools and reading short scripts. No programming background is required, but participants should be willing to read the code an agent writes and to work in a terminal with guidance. A laptop on which the starter environment can be installed.
- 10–18 participants per instructor; a teaching assistant above 14, or when most participants have never used a terminal.
-
- Each team brings one process from its own work or picks one from a prepared catalogue (lead intake, invoice handling, support triage, recurring reports).
- All labs run against sandbox APIs, a mock CRM, a ticketing system and synthetic data; no lab touches production credentials or live customer data.
- One coding agent (Claude Code or Codex) and one visual builder (n8n) are used hands-on; other platforms are discussed, not installed.
- A prepared starter repository provides connectors, test data and a deployment target, so lab time goes to the automation rather than to setup.
- Roughly 10 hours of guided building, 8 hours of concepts and demonstrations, and 3 hours of business cases, review and presentations.
The course teaches working proficiency in choosing, building and running automations. It does not certify compliance, and a capstone built against sandbox systems still needs a security review and a data-protection assessment before it touches production.
Book a training
Tell me which course and roughly when. You get a reply from me, not a sales sequence — dates, a quote, and an honest answer if the course is a poor fit for what your team actually needs.