A lot of office work consists of moving information between apps: reading an email, copying details into a spreadsheet, drafting a reply, updating a tracker, notifying a colleague. Automation platforms have handled simple versions of this for years. Adding AI lets automations deal with messy, unstructured information such as emails, forms and documents.
Updated September 2026: we added sources on prompt injection risk and on tools that enforce structured AI output.
The main tools
- Zapier, Make and n8n connect thousands of apps and include AI steps.
- Microsoft Power Automate suits organizations using Microsoft 365.
- Built-in automation in tools such as Google Workspace, Notion, Airtable and CRMs.
- AI agents that can operate across apps, discussed in our agents explainer, for more flexible but less predictable workflows.
Good candidates for automation
Look for tasks that are frequent, follow a pattern and involve reading or writing text:
- Sorting incoming emails or form submissions by type and urgency
- Extracting details from invoices, receipts or applications into a spreadsheet
- Summarizing new support tickets and routing them to the right person
- Drafting replies for human approval
- Turning meeting transcripts into action items in a task tracker
- Creating weekly digests from several sources
A simple example
- Trigger: a new message arrives in the sales inbox.
- AI step: extract company name, request type and urgency, and write a two-sentence summary.
- Action: add a row to the leads spreadsheet.
- Condition: if urgency is high, notify the sales channel.
- AI step: draft a reply and save it as a draft, not sent.
A person reviews the draft and sends it. The tedious parts are automated; the judgment stays human.
Design rule Automate the reading and drafting. Keep a human approval step before anything is sent to a customer, money moves or records are deleted.
Making AI steps reliable
- Ask for structured output, such as specific fields or categories from a fixed list, rather than free text. Some AI platforms can enforce this: OpenAI’s Structured Outputs, for example, constrains responses to a schema you define.
- Give examples in the AI step’s instructions. Our prompting guide applies here.
- Handle uncertainty: include an “unclear” category that routes to a person.
- Test with real data, including odd cases, before switching it on.
- Log everything, so you can see what the automation did and why.
Common failures
- Automations that silently break when an app changes
- AI misclassifying unusual inputs without anyone noticing
- Runaway costs from high-volume AI steps
- Sensitive data flowing into tools that have not been approved, which your AI policy should address
- Prompt injection, where text inside an email or document tries to instruct the AI to do something else; OWASP ranks prompt injection as the top security risk for applications built on language models
Start small and measure
Automate one workflow, run it alongside the manual process for a week or two, check the results and track time saved, as described in our ROI guide. Then expand.
Start with the repetitive work
AI-powered automation is most valuable for the repetitive reading, sorting and drafting that fills everyone’s day. Build with structured outputs, human approval for important actions and good logging, and you can reclaim hours each week without losing control.
Sources
- LLM01:2025 Prompt Injection, OWASP Gen AI Security Project
- Introducing Structured Outputs in the API, OpenAI, August 2024
