Automations: when AI works on its own

In module 7 you used AI to research, write and analyze, and every workflow had one thing in common: you were there, asking and reviewing. This module is about the next step: tasks that run without anyone starting them. An email that becomes a row in a sheet, a summary every morning, an assistant that answers on WhatsApp at eleven at night.

It is still a no-code module: what changes when AI works alone, how a flow is built, what is realistic on WhatsApp, how to pull data out of invoices and, above all, what to automate first and what to leave alone.

What is the difference between asking AI for something and automating it?

When you chat, you are both the trigger and the reviewer: you start the task, read the result, correct it. An automation removes both. Something external starts it (a time, an email, a form) and the result lands somewhere (a sheet, a draft, a message) without anyone reading it at that moment. That makes it valuable, and delicate.

The biggest practical change is in the prompt. In a chat you can be vague because you fix things in the next message. In an automation nobody fixes anything: the instruction has to say what a good result looks like, what to do with odd cases and what to do when something fails. Module 4 (role, context, format, examples) matters twice as much: the prompt is written once and runs a hundred times.

The simplest form is the scheduled task, and as of October 2026 it is built into the assistants themselves. In the Claude app you describe a task and a frequency in any conversation and schedule it. It is available on paid plans; according to Anthropic’s documentation, runs execute in the cloud even when your computer is off, except for tasks that need local files or programs, which run only on your machine. Claude Code, the tool in part 3, also has "routines": a prompt, one or more repositories and connectors packaged to run on Anthropic’s cloud on a schedule, from an external call or from a GitHub event. They are in research preview (details change); module 15 covers them in depth.

How does an automation work with n8n, Zapier or Make?

Automation platforms predate AI: they connect apps (email, spreadsheets, CRM, WhatsApp) through flows you draw by linking boxes on screen. What is new is the AI step in the middle. Almost every useful workflow has four pieces:

1. 1Trigger. What starts it: an email with an attachment, a new row, a form, a time of day.
2. 2AI step. The model receives the content plus a fixed instruction (extract, classify, draft) and returns something structured.
3. 3Action. What is done with that output: create the row, the draft, the CRM contact, the message.
4. 4Human review. A point where a person approves before anything reaches the world. At first, in every flow; later, only where the consequences call for it.

Three typical flows for a small business:

• Invoice lands in the inbox, row appears in the sheet. Trigger: supplier email with an attachment. AI: extracts tax ID, number, date, subtotal, VAT and total. Action: new row in the payables sheet. Review: the row stays "pending" until accounting confirms it.
• Website form, classified contact in the CRM. Trigger: someone submits the form. AI: classifies (service, city, urgency, budget) and drafts a first reply. Action: CRM contact with those tags and the reply saved as a draft. Review: the salesperson reads it and sends it.
• Daily summary. Trigger: every day at 7. AI: reads yesterday’s bookings, messages or tickets and writes ten lines on what needs a decision. Action: email to whoever runs the business. Review: none; a summary acts on nothing, and if it is wrong you notice.

As of October 2026, the three platforms you will hear about most all document that AI step. n8n has a schedule trigger and AI nodes (an agent node plus models from Anthropic, OpenAI and others, with your own API key); you can use its cloud or install it on your own server, which matters when you do not want data to leave. Zapier has "AI by Zapier", a built-in AI step that runs on its own models or on your key from OpenAI, Anthropic, Google and others. Make has an Anthropic Claude app with a simple text prompt module that needs no API account and another that uses your key. Prices and plans change often; check each site.

I want to automate this process in my business: [DESCRIBE THE PROCESS IN 3 TO 5 SENTENCES: what comes in, what is done with it, what goes out and who does it today].

Before proposing anything, ask me the questions you need about four things, one at a time: (1) the trigger (which event or time starts it), (2) the data (where it comes from, in what format, whether it includes personal data), (3) the tools I already use (email, spreadsheet, CRM, WhatsApp, accounting software) and (4) the risk (what happens if a result is wrong and nobody sees it).

With my answers, give me the flow step by step in this format: Trigger, AI step (with the fixed instruction the model would receive), Action, Human checkpoint (mandatory: tell me exactly where it stops to wait for approval and why there). Close with the three most likely odd cases and what the flow should do in each one.

Does a WhatsApp agent work for customer service?

Yes, for part of the conversations. A WhatsApp agent does well with questions that follow clear rules and have an answer in a source you approve: hours, location, payment methods, availability in your system, lead qualification, appointments, order status and the basics after hours. It does badly, and should not be alone, with complaints, money, legal or health matters, negotiations and anything it cannot verify against your data. We cover it in the guide to WhatsApp customer service with AI.

The technical route has two levels. With low volume, the WhatsApp Business app with quick replies may be enough. For an AI agent you need the WhatsApp Business Platform, whose Meta-hosted version is called the Cloud API: it requires a Meta business portfolio and a dedicated number, and it is almost always contracted through a provider. Meta charges per message, with rates by type and country that have changed several times in 2025 and 2026; check the official pricing page before signing.

On Meta’s rules, as of October 2026, this is what can be verified on its official pages. The WhatsApp Business Platform terms bar AI providers from the platform when AI is the primary function of what they offer (a general-purpose assistant), and Meta decides what falls under that; a business using AI to serve its own customers about its own business is not that case. The messaging policy requires that, if you automate, customers have a prompt, clear way to reach a person (in the same chat, by phone, email or web), and that you have their opt-in before messaging them on your own initiative. These rules change; the source is the official policy, linked at the end.

Two design decisions you do not delegate. The first is the handoff: define which topics go to a person from the first message, which words trigger it ("I want to talk to someone", "complaint", "refund") and how the person finds out. The agent introduces itself as an automated assistant and says how to ask for a human. The second is data: every chat carries a name, a number and sometimes an ID number, an address or health information. Apply the traffic light from module 5: the agent asks only for what it needs, never card numbers or passwords, the first message links to your privacy notice, and you know where conversations are stored and what the provider does with them. With sensitive data, talk to your lawyer before connecting anything.

How do I get the data out of an invoice with AI?

It is the most common automation in a small business. First, a Colombian rule: when a supplier issues electronic invoices, the invoice is the signed XML file and the PDF is only its graphic representation, under DIAN’s technical annex. That XML already carries the tax ID, number, dates, taxes and totals as fields; you do not need a model to "read" anything, you need your flow or your accounting software to import the XML. AI is for what does not arrive structured: delivery notes, bills from individuals, receipts as photos and scanned PDFs. More in the guide to invoice data extraction with AI.

For those documents, the instruction asks for structured output: you define the fields and the format (usually JSON) and the model fills exactly those fields. Current Claude models, for example, guarantee through the API that the response follows the schema. What no model guarantees is that the value is correct: a 7 that looks like a 1 in a blurry photo arrives perfectly formatted. Hence the second half of the flow, validation:

• Subtotal plus taxes equals the total, with a few pesos of tolerance for rounding.
• The tax ID exists in your supplier list and the date falls in a reasonable range.
• The document number is not already in your sheet.
• Every field carries a confidence level (high, medium, low) assigned by the model, and anything low goes to review.
• Anything that fails a rule is flagged and waits for a person; nothing is recorded as valid by default.

You are a bookkeeping assistant for [BUSINESS NAME]. You receive a document (text or image) and return only a JSON object, with no explanation or extra text, with exactly these fields:

{
  "document_type": "invoice | delivery_note | bill_from_individual | other",
  "issuer_name": "",
  "issuer_tax_id": "digits only, without the check digit",
  "document_number": "",
  "date": "YYYY-MM-DD",
  "subtotal": number without separators or symbols,
  "taxes": number,
  "total": number,
  "items": [ { "description": "", "quantity": number, "unit_price": number } ],
  "confidence": "high | medium | low",
  "notes": "what you could not read or what does not add up"
}

Rules: if a field does not appear in the document, use null; never infer it. If subtotal + taxes does not equal total, report it in notes and set confidence to low. If the image is blurry, cropped or has stamps over the data, confidence is low. Convert Colombian number formats (1.250.000) to a plain number (1250000). [ADD ONE OR TWO EXAMPLES OF A DOCUMENT AND ITS CORRECT JSON HERE]

What should I automate first?

The temptation is to start with the most impressive thing; the rule is to start with what hurts most and risks least. A simple score helps you decide without arguing. For each task, write down three numbers:

1. 1Frequency. How many times a week it happens. A daily task is worth more than a month-end one.
2. 2Minutes. How long each run takes, counting searching, copying and confirming. Multiply the two and you have the hours going out the door.
3. 3Human error risk. From 1 to 5: how easy it is to get wrong by hand and how expensive that is. Copying figures between systems is a 4 or a 5; a courtesy email is a 1.

Multiply the three. High-scoring tasks are candidates; among them, pick first the one whose inputs look alike every time and whose result can be checked in seconds. Then apply two delegation rules from module 1: nothing gets automated before you have done it by hand with AI several times, and the first version is always "draft for approval", never "fully automatic". When you have spent weeks approving without correcting, you decide which steps to release.

And log everything. Every run leaves a trace: when it ran, what came in, what went out, who approved and what was changed. Without that record you cannot tell whether the flow is getting better or worse, or explain a badly recorded row. It is the foundation of module 9.

What not to automate, at least for now:

• Decisions with consequences for money, contracts or people: approving a payment, rejecting a candidate, replying to the tax authority. AI prepares; a person decides.
• Processes with no written rules. If everyone does it their own way, the first job is writing the process down, not linking boxes. AI improvises in the gaps.
• Tasks with red-light data until you have a business plan and a real need to use that data.
• Things that happen twice a year. Building and maintaining the flow never pays back.

I am [YOUR ROLE] at a [WHAT IT DOES] business in [CITY]. Here are my repetitive tasks. For each one, estimate or use the data I give you: frequency per week, minutes per run and human error risk from 1 to 5 (how easy it is to get wrong by hand and how expensive that is). Compute the score = frequency × minutes × risk.

My tasks:
[LIST OF 5 TO 12 TASKS, ONE PER LINE, WITH WHAT YOU KNOW ABOUT FREQUENCY AND MINUTES]

Return a table sorted from highest to lowest score with these columns: task, frequency, minutes, risk, score, hours per month lost, and a "first version" column saying whether it would start as a draft for approval or as an informational summary. Mark with NO the tasks that should not be automated yet and say why in one line (decision with consequences, no written rules, personal data without controls, too infrequent). Close with the single task you would start with this week and what would need to be written down before starting.

The next module, "Trust, but verify", takes these flows and adds the checks: sources, human review in the right place, tests with real cases and the signs that something is failing quietly.