Adding AI steps
The four jobs AI does well inside workflows — classify, extract, draft, summarize — and the decision test for when an AI step earns its place.
Your Module 2 workflow has a known hole: the order number buried in a forwarded mess, the email whose meaning no pattern can parse. Deterministic steps read structure; the real world ships prose. AI steps are the bridge — and inside workflows, AI does four jobs superbly:
- Classify — 'which bucket?': complaint / question / refund request / spam; urgent / routine; English / Spanish. Classification is the workhorse, because buckets are what branches route on. AI classifies messy human text the way rules never could.
- Extract — 'pull out the...': order number wherever it hides, product name, requested action, customer sentiment. Extraction turns prose into fields — and fields are what every downstream step eats. This is the fix for test email #5.
- Draft — 'write the first version': the reply, the summary row, the reminder. Drafts go to humans or through approval gates (Module 4) — AI writes, something accountable sends.
- Summarize — 'compress this': the 30-message thread into four bullets, the week's tickets into themes for Riley's ops report. Summaries make the human steps of a workflow faster, which counts as automation too.
Where AI steps live in your platform
Every major platform now ships AI steps natively — an 'AI / LLM' action you drop in like any other: it takes your prompt plus mapped fields as input, returns its answer as output fields for later steps to map. (Platforms also offer an 'HTTP request' step for calling an AI API directly — an API being the doorway one piece of software uses to talk to another — more control, more setup; native steps are the right start.) One practical note before anything else: check the data-handling terms the same way you did for connected apps — an AI step is another vendor seeing your data, and your company's approved-tools list applies.
'Natively' hides real differences. What each platform actually requires:
- Zapier — the built-in AI action needs no key of its own; dedicated AI-app actions need a provider account plus an API key (a password-like code that identifies your account to the provider).
- Make — AI modules need a provider API key.
- n8n — AI nodes take provider credentials.
- Power Automate — AI Builder actions run on credits included in some plans.
- 1If your platform wants a provider key: create an account at the provider's developer console (pick a current mainstream model provider — your company's approved list decides which).
- 2Add a small amount of credit — a few dollars covers every lab in this course.
- 3Create a key, copy it, and paste it into the platform's AI connection screen. Treat it like a password.
Cost, in one sentence: each AI step run costs a fraction of a cent to a few cents depending on the model — check the provider's current pricing.
The earn-its-place test
- Try rules first, honestly. If a pattern or a keyword filter genuinely covers 95% of cases, the deterministic step wins: free, instant, and identical every run. AI steps are for where rules visibly fail — like meaning.
- Demand a checkable output. Good AI-step outputs are short and verifiable: a category from YOUR list, a yes/no, an extracted value that later steps validate. 'Do something smart with this email' is not a step; 'return exactly one of: complaint, question, refund, other' is.
- Design for 'usually'. AI steps are right usually. The workflow around them must be right always — which means every AI output gets either validated, branched on conservatively, or reviewed by a human before anything irreversible happens. That design discipline is the next two lessons.
Deterministic rails, AI judgment at marked points, humans at consequential ones — you're building the small, sturdy cousin of what AI engineers ship. When a workflow's AI needs to choose its own steps, you've left workflows and entered agents (the Agentic AI Systems course, several rungs up the ladder). For the overwhelming majority of office automation, the workflow pattern isn't the compromise — it's the correct tool.