Training staff to automate workflows with AI
By the Day Seven team · Updated 14 October 2026
Workflow automation training teaches people to spot repetitive steps, decide whether an AI assistant or an automation tool fits, build a small automation with human checkpoints, and measure whether it saves time. The best results come from training people on their own processes, starting with one low-risk workflow each.
Summary
- Map the workflow before automating it — most savings come from removing steps, not adding AI.
- Use an assistant (ChatGPT, Claude, Copilot) for judgement-heavy drafting; use automation tools for repeatable hand-offs.
- Keep a person at every point where an error would reach a customer or a record.
- Agree with IT which automation tools and connections are allowed.
Choosing a workflow to automate
Good candidates are frequent, rule-based, and currently involve copying information between systems or reformatting it. Poor candidates depend on judgement that is hard to describe, change often, or carry high consequences if wrong.
- Triage of a shared inbox into categories with a suggested reply.
- Turning meeting notes into actions in a task tool.
- Extracting fields from standard documents into a spreadsheet for checking.
- Weekly report assembly from data that already exists.
Assistant or automation tool?
Training should make the difference clear, because choosing the wrong one is the most common source of wasted effort.
| AI assistant | Automation tool (e.g. Power Automate) | |
|---|---|---|
| Best for | Drafting, summarising, analysing on request | Moving data between systems on a trigger |
| Who runs it | A person, each time | Runs automatically once set up |
| Typical skill needed | Prompting and checking output | Mapping steps, triggers and permissions |
| Main risk | Unchecked or inaccurate output | Errors repeated at scale, over-broad access |
| Human checkpoint | Before the output is used | Before anything leaves the team or changes a record |
Designing automations safely
Each automation needs an owner, a description of what it does, and a defined point where a person reviews the result. Grant the automation only the access it needs, and test it with dummy data first.
For automations that call AI models, the NCSC's guidelines on secure AI system development describe the risks to consider, including prompt injection from untrusted inputs such as inbound email.
What a training programme covers
A practical programme usually runs across a few sessions so people can build between them.
| Session | Focus | Output |
|---|---|---|
| 1 | Mapping a current workflow and finding wasted steps | A one-page process map |
| 2 | Using an AI assistant for the judgement steps | Tested prompts with a checking step |
| 3 | Building a simple automation in an approved tool | A working automation on test data |
| 4 | Governance, ownership and measurement | Owner, checkpoint and baseline agreed |
Measuring the result
Record how long the workflow takes before and after, including time spent checking, and how often the output needs correcting. An automation that saves time but increases rework is not a saving.
An illustrative example
This is a hypothetical scenario, not a Day Seven client result.
An operations team receives supplier queries in a shared mailbox. After training, one team member builds a flow that categorises each new email, drafts a suggested reply with an approved assistant, and places both in a review queue. Nothing is sent without a person approving it. The team times the process for two weeks before and after.
Frequently asked questions
Sources
Related
Want help applying this to your organisation?
Training is delivered on-site at your premises or live online, scoped after a short conversation. Individuals can join the in-person Leeds AI Fundamentals Workshop.