Skip to content
KAEDEDIGITAL
עב
Business systems

Build automations with AI: describe the workflow in chat, then check five things before it runs alone

17 November 2026 8 min read Kaede Digital

In short

  • You can now build automations with AI by describing a workflow in a chat tool, and platforms like Zapier set it up.
  • Brief the AI like a new employee: the trigger, each step, the decisions, the approvals and what happens on failure.
  • Before a flow runs alone, check approval points, failure handling, data access, cost per run and who owns it.
  • OpenAI's Agent Builder shuts down on 30 November, so keep each workflow documented outside the tool that runs it.
  • Flows that touch money, customer data or the sales pipeline are safer in a custom CRM and automation setup.

You can now build automations with AI by describing the workflow in plain words inside a chat tool, and the automation platform sets it up for you. Since September, Zapier lets you describe a workflow in Claude, ChatGPT or Cursor and have it go live on Zapier, and Google has added outside apps, webhooks and approval settings to Workspace Studio. The catch is that it has never been easier to build something nobody checks.

This guide covers how to brief the AI so it builds the right workflow, the five checks to run before a flow works on its own, how to avoid building on a tool that disappears, and when a custom setup is the safer choice.

What changed when you build automations with AI?

The chat window has become a place where automations are built, not just discussed. Several platforms moved this way within a few weeks, and one well known tool is going away.

Zapier’s Next Gen Zaps

At ZapConnect, Zapier launched Next Gen Zaps. You describe a workflow inside an AI tool that supports MCP (an open standard for connecting AI tools to other apps), such as Claude, ChatGPT or Cursor, and it is set up on Zapier. You can also build inside Zapier itself with plain language, code or a visual canvas.

Compared to classic Zaps, they process whole lists with no 500-item loop limit, can pause mid-run for a human reply and then carry on, and keep running when one step fails. They can also have a monitoring agent that finds failures and applies or suggests a fix.

Next Gen Zaps are in early access by request, on the Professional, Team and Enterprise plans only, and run through Zapier MCP. Runs are free for now. From 1 December, each run counts against your plan’s tasks.

Google Workspace Studio

From 17 September, Workspace Studio started rolling out custom triggers, custom steps, connections to third-party apps (in beta) and webhooks. New steps can move or copy a Drive file, send a Google Chat reply and reply to an email.

Admins can change the human approval requirements in the Admin Console and turn off individual steps. They can also require user approval when an action could share data outside the organization.

n8n 3.0 and OpenAI’s Agent Builder

n8n released version 3.0 on 27 October. It is a major release with breaking changes, and n8n says to read the breaking changes notes and run its migration tool before upgrading.

OpenAI is shutting down Agent Builder, its visual tool for building agents, together with its hosted Evals product. Both will be gone from 30 November. The named replacements are the Agents SDK and ChatGPT Workspace Agents.

The common thread: building a workflow is getting much faster, so keeping it under control is now the real job.

How do you describe a process so the AI builds the right workflow?

Brief it the way you would brief a new employee: what starts the process, every step in order, the decisions, the exceptions and what “done” looks like. The AI fills gaps with guesses, so fewer gaps mean fewer surprises.

  1. Start with the trigger. Name the exact event, such as “a new form submission on the contact page”, not “when a lead comes in”.
  2. Name every system and field. Say which CRM, which pipeline stage, which column and which mailbox.
  3. Spell out the decisions. Write the rules as if/then: if the enquiry mentions a rental, assign it to the rentals agent.
  4. Mark where a person must approve. Flag any step that messages a customer, moves money or shares a file.
  5. Say what happens when something goes wrong. Cover missing data, duplicates and an app that does not respond.
  6. Define the end. Say what the flow leaves behind: a task, a CRM note, a message to the owner.

Here is how a real estate office might brief it:

“When someone submits the property enquiry form, create a lead in our CRM with the source ‘website’ and the property they asked about. If it is a rental, assign it to the rentals agent; otherwise to sales. Draft a WhatsApp reply but do not send it until the agent approves. If the phone number is missing, create a task to reply by email.”

Before you switch the flow on, ask the AI to explain it back to you step by step, then test it with realistic data, including a messy entry. If the explanation does not match what you meant, fix the brief, not just the flow.

What should you check before an automation runs on its own?

Check five things: where a person approves, what happens on failure, who sees the data, what each run costs and who owns the flow. A flow built in five minutes can run for months, so this time is well spent.

1. Where does a person approve?

Find every step that acts outside your business: replying to a customer, sending a quote, issuing a refund, sharing a document. The new tools support this, since Next Gen Zaps can pause for a human reply and Workspace Studio admins can require approval when data could leave the organization. A beauty salon might let a flow book appointments freely but require approval before any cancellation message goes out.

2. What happens when a step fails?

Decide whether a failure should stop the flow or let it continue, and who hears about it. Carrying on is fine for a list of reminders. It is risky for an online store, where a failed payment check should never be followed by a “your order has shipped” message.

3. Who sees the data?

When you build through a chat tool, data can pass through the AI tool, the automation platform and every connected app. A clinic handling patient details should know exactly where each field travels and keep sensitive fields out of steps that do not need them.

4. What does each run cost?

Check how your plan counts runs and steps before going live. A flow that processes whole lists, or triggers on every incoming email, can use far more of your plan than you expect.

5. Who owns it?

Every automation needs a named owner who knows what it does, gets the failure alerts and approves changes. Without one, a flow becomes something everyone is afraid to touch.

A short pre-launch checklist:

  • The brief is saved in writing, outside the automation tool
  • Customer-facing and money-related steps have approval points
  • Failures alert a named person
  • You know which apps see which data
  • You know how runs are counted on your plan
  • One person owns the flow

How do you avoid building on a tool that gets shut down?

Keep your business logic and data in systems you control, and treat the automation tool as replaceable plumbing. Agent Builder is a clear reminder: anyone who built on it must rebuild before 30 November.

Keep the written brief for every workflow so you can rebuild it anywhere. Store customers, deals and history in your CRM, not inside the automation. Read the release notes of the tools you rely on, and before a major upgrade like n8n 3.0, test the migration on a copy first.

If you are still choosing a platform, our comparison of Make, Zapier and n8n covers the trade-offs. If you want AI to work with your customer records directly, see how to connect your CRM to ChatGPT.

When is a custom CRM and automation setup safer than a chat-built flow?

A custom setup is safer when the workflow touches money, customer data or your sales pipeline, runs at high volume or is something the whole team depends on. Flows built in chat are great for small, internal, low-risk tasks.

Good chat-built candidates: copying a file into a client folder, a daily summary for the owner, a restock reminder. For more ideas, see these business tasks you can stop doing by hand.

Better in a custom setup: a B2B service company whose quotes and invoices must follow the same rules every time, a restaurant whose bookings and WhatsApp confirmations depend on staff schedules, or an online store where refunds, stock and customer messages must stay in sync. Here a custom CRM holds the data and the rules, and the automations around it are documented, monitored and owned. When one platform changes, you swap a connection, not your whole operation.

Fast to build, so build the checks in too

AI now lets small teams build automations just by describing them. What remains is the part the chat window does not do for you: clear briefs, approval points, failure alerts, data boundaries, cost awareness and an owner. At Kaede Digital we help businesses decide which flows they can safely build themselves and which belong in a proper system. When the stakes are higher, our business process automation service builds and maintains the workflows around your CRM, so they keep running when tools change.

FAQ

Can I build automations with AI without technical skills?

Yes, for simple flows. Tools like Zapier’s Next Gen Zaps let you describe a workflow in a chat tool and have it set up for you. You still need to describe the process clearly and test it.

Is a workflow built in chat less reliable than one built by hand?

Not necessarily, but it is easier to launch without review. The AI may fill gaps in your description with guesses, so ask it to explain the flow back and test it with realistic data before it runs alone.

What should I do if I built something in OpenAI’s Agent Builder?

Plan the move now, since Agent Builder will be gone from 30 November. OpenAI names the Agents SDK and ChatGPT Workspace Agents as replacements. Write down what each agent does so you can rebuild it on any platform.

Do I need a CRM to use AI-built automations?

Not for small internal tasks. For anything involving leads, customers or sales, a CRM gives automations one reliable place to read and write data, and your records stay with you if you change tools.

Business systems

Want us to build this for you?

See the service