How to automate a workflow that spans several apps
To automate a workflow across apps, first map its trigger, inputs, steps and result. Then pick the approach that fits: an integration platform such as Zapier, Make or n8n when every app has an integration and the rules never change; a desktop macro when the steps are identical every time; or an AI agent with MCP tools when a step needs judgment. Ditto AI builds that kind of agent from a recording of you doing the workflow once.
Map the workflow first
Most workflows that are worth automating look simple until you write them down. Take five minutes and fill in this map:
| Part | Question | Example: the weekly invoice |
|---|---|---|
| Trigger | What starts it? | A vendor emails an invoice |
| Inputs | What changes each time? | Vendor, invoice number, amount, date |
| Steps | What do you do, in which app? | Open the email, save the PDF to Drive, add a row to the expenses sheet |
| Judgment | Where do you decide something? | Which folder, whether it’s a duplicate |
| Result | How do you know it’s done? | The file is in the folder and the row is in the sheet |
The judgment row decides which approach you need. If it’s empty, rules can do the whole job. If it isn’t, you need something that can look at the situation and decide.
Three ways to connect apps
Integration platforms
Zapier, Make and n8n connect web apps through their APIs. You choose a trigger in one app and actions in others, and the platform runs it in the cloud whenever the trigger fires, even when your laptop is closed.
- Fits: every app in the workflow has an integration, and the rules are fixed.
- Breaks: when an app has no integration, when the data isn’t where the rule expects it, or when a step needs a decision you didn’t encode as a branch.
Desktop macros and RPA
Macro tools and robotic process automation replay the clicks and keystrokes you’d make in each app. They reach desktop apps and websites without integrations, and they run exactly what you recorded.
- Fits: steps that are identical every time, in a user interface that doesn’t change.
- Breaks: when a window moves, a button is renamed, a dialog appears, or the item you need is in a different place in a list.
AI agents with tools
An AI agent looks at the task, picks the next step and calls a tool to do it. Tools usually come from MCP servers, small programs that give a model safe, typed access to one app: list emails, upload a file, add a row. Where no integration exists, an agent can work the app the way you do.
- Fits: workflows with judgment in them, and inputs that vary from run to run.
- Needs: clear permissions, so the agent only writes where you expect, and a way to review what it will do. See Is it safe to let an AI agent use your accounts?
How MCP makes cross-app agents practical
The Model Context Protocol (MCP) is an open standard for connecting AI models to tools. Instead of every agent needing custom code for every app, an app publishes, or the community builds, one MCP server, and any MCP client can use it: Claude, Kiro, ChatGPT and others.
For a cross-app workflow, that means one agent can hold a mail server, a Drive server and a spreadsheet server at once, each limited to what it needs. There’s more in What is an MCP server?
Build a cross-app agent from one recording
Choosing and configuring MCP servers is the tedious part of building an agent by hand. Ditto AI does it for you from a recording.
- Record the workflow once on your Mac, across all the apps it touches, with real data.
- Ditto reads the recording and lists the process, the steps with the app for each, and the inputs.
- It finds a tool for every step. Public web pages use a headless browser. Apps that need your account get an MCP server from the official registry or npm, checked with one read-only call.
- You review it and fix any step, input or tool before it runs.
- You run it with new inputs, in Ditto, or in Claude Code or Kiro, and approve each write while you’re building trust.
How Ditto AI works covers each stage in detail.
What to automate first
Score each workflow you have in mind from 1 to 3 on each of these, and start with the highest total:
- Frequency: daily is a 3, weekly a 2, monthly a 1.
- Time: more than 20 minutes is a 3.
- Low risk: easy-to-undo results (drafts, rows, files) are a 3; payments and deletions are a 1.
- Stability: a workflow you do the same way each time is a 3.
Keep a person in the loop where it matters
Automating a workflow doesn’t mean removing yourself from it. Good places to stay involved:
- Before anything is sent to someone outside your team.
- Before money moves, or anything is deleted.
- When the input is unusual: a new vendor, a very large amount, a missing field.
In Ditto AI, the Ask before it creates, sends or changes anything option turns every write into a draft you approve. With it off, an agent still asks before deleting, sending somewhere new, or doing anything you didn’t record.