Course resource
Workplace Integration Patterns
The shapes that AI-in-your-workflow actually takes. Recognising the pattern tells you what to build and where it will break.
Week 5 covers how to build these in an automation tool. This is about choosing the right shape first.
The three-layer model
Every integration is the same three layers:
INPUT -> AI STEP -> OUTPUT
- Input — where the work arrives: an email, a form, a new row, a schedule, a message
- AI step — the judgement: classify, extract, summarise, draft, decide
- Output — where it goes: a draft, a row, a notification, a document
Most failures come from getting the input wrong — usually by not filtering it, so the AI runs on everything and you pay for it.
The six patterns
1. Triage — classify arriving work and route it.
Input: new message
AI: classify into [CATEGORIES], score urgency
Output: label, assign, or notify
Highest value, lowest risk, because a misclassification is visible and cheap. Start here.
2. Draft-first — AI writes, a human sends.
Input: message needing a reply
AI: draft using context and voice profile
Output: saved to Drafts, never sent
The workhorse. The output landing in Drafts rather than Sent is the entire safety mechanism.
3. Extraction — turn unstructured text into structured data.
Input: email, form, PDF, transcript
AI: extract named fields
Output: row in a sheet or CRM record
Very reliable when you constrain the output format and instruct it to mark unknowns rather than guessing.
4. Digest — many inputs, one summary, on a schedule.
Input: scheduled trigger + collected items
AI: summarise, rank, highlight what changed
Output: one message at a fixed time
Replaces the habit of checking something twelve times a day.
5. Enrichment — take a thin record and add context.
Input: new lead, new ticket, new candidate
AI: research, summarise, attach findings
Output: appended to the record
Watch for fabrication here. Instruct it to cite where each fact came from, and to leave fields blank rather than inventing.
6. Monitor — watch for a condition, alert when it occurs.
Input: scheduled check of a source
AI: judge whether the condition is met
Output: alert only when true
The value is in what it does NOT send. An alert that fires constantly gets muted within a week.
Choosing
| If the work is... | Pattern |
|---|---|
| Arriving and needs sorting | Triage |
| Arriving and needs a reply | Draft-first |
| Arriving as text, needed as data | Extraction |
| Scattered and needs consolidating | Digest |
| Thin and needs context | Enrichment |
| Rare but important to catch | Monitor |
Where the human goes
Every pattern needs a human somewhere. The question is where.
| Risk if AI is wrong | Human position |
|---|---|
| Someone outside sees it | Before output. Always. |
| Money moves | Before output, with a second check |
| An internal record is wrong | After output, with an audit trail |
| Nothing, it is a suggestion | Not required |
The test: if this runs wrong 100 times before anyone notices, what is the damage? If the answer is more than annoying, the human goes before the output.
The failure modes
No filter on the input. The automation runs on every email rather than the ten that matter. You discover this on the bill.
Silent failure. The AI step errors, the workflow continues, and the output is empty or wrong. Every workflow needs an error path that tells a human.
Temperature too high on a classification step. The same input classified differently on different runs. Covered in the temperature guide — keep judgement steps at 0–0.3.
No logging. Something went wrong three days ago and there is no record of what the AI actually returned. Log the input and the output of the AI step.
Automating a bad process. The most expensive one. If the underlying process is wrong, automating it produces wrong results faster. Fix the process first.
Before you build
- I have done this task manually enough times to know its edge cases
- The process is worth keeping, not just worth automating
- I know what the input filter is, precisely
- I know where the human sits, and why there
- I know what happens when the AI step fails
- I know what this costs per run, and per month at expected volume
- Someone other than me could understand it from the documentation
Your integration map
| Task | Pattern | Input | AI does | Output | Human where | Runs/month |
|---|---|---|---|---|---|---|