AI Automation and Chatbots

In Zapier or Make, build the form-to-spreadsheet flow from the lesson: a trigger, one AI classification step, and a row write. Submit three real test entries and confirm each lands in the sheet with a sensible category. Done when all three rows are correctly sorted without you touching the sheet.

Build a Voiceflow bot with a knowledge base of your real FAQ (or a friend's small business) and a human-handoff intent. Publish it and embed it on any page, even a free one-pager. Done when a stranger can open the page and get a correct answer to a real question.

Away from the keyboard, write down by hand every repetitive task you did this week and how long each took. Circle the three that follow a clear trigger-then-action pattern. Done when you have a paper list of at least three candidates ranked by time wasted.

Sit a friend or coworker in front of your published chatbot without explaining it. Watch silently and note every moment they get confused or stuck. Done when you have three concrete observations and at least one fix you will make.

Create a flow that takes incoming messages (from a webhook or form), classifies them once with AI, then routes each category to a different destination, with an Other fallback branch. Done when test messages of three different types each land in the correct place and an unknown type hits the fallback.

Use a Webhook trigger and an HTTP Request step to wire together two services that have no ready-made connector. Store the API key in the credentials manager, not inline. Done when data flows from the source through your flow into the target via raw HTTP.

Upgrade your chatbot with variables (name, order or account reference) and conversation memory, then give it to one real user with a genuine task. Done when the user completes their task and the bot never re-asks something it already captured.

Away from building mode, feed your automation deliberately bad input: blank fields, an emoji-only message, a wrong data type, a duplicate submission. Write down each way it failed by hand. Done when you have a list of failure modes and a planned fix for each.

Pick a process you want to automate and draw the whole thing on paper first: trigger, every branch, error paths, and where AI fits. Done when a non-technical friend can follow your drawing and tell you what happens to a refund request.

Create an n8n AI Agent (or API-based agent) with at least three tools, including one that reads data and one that takes an action with a guard (limit or confirmation). Done when the agent correctly chooses different tools for different requests and refuses or escalates an action that exceeds its guard.

Build a chatbot that retrieves from a vector store of your real documents and answers only from retrieved context, citing the source and admitting when it does not know. Done when it answers an in-docs question with a citation and correctly declines an out-of-docs question instead of inventing an answer.

Add full tracing to an existing AI automation (input, output, model, tokens, cost, latency), then apply at least two cost optimizations from the lesson. Done when you can show a before-and-after cost per run and prove quality held using your eval set.

Build an eval set of at least fifteen real inputs with known-correct outputs. Make a prompt or model change, run the set before and after, and record the score delta. Done when your eval correctly flags whether the change helped or regressed.

Away from the screen, present your AI automation or agent to a business owner or manager who would actually use it. Explain what it does, what it costs to run, and how it fails safely. Done when they either commit to trying it or give you a specific reason they will not, which is just as valuable.