What to automate first with AI

When a team decides to implement AI, the first question is usually "where do we start?" The answer most people give is some version of "find a repetitive task and automate it." That's not wrong, but it's not specific enough to be useful.

Here's the framework we'd actually use.

The four criteria that matter

A good first automation target scores well on all four:

1. High volume. You want something that happens many times a week, not occasionally. If the task only comes up twice a month, the impact of automating it — even if you do it perfectly — is small. Volume is what turns small efficiency gains into real savings.

2. Consistent input. AI works best when what goes in is predictable. "Summarise this customer support ticket" is a good prompt. "Handle this situation however seems right" is a bad prompt. Start with tasks where the input is structured enough that you can write clear instructions.

3. Low consequence of error. For your first automation, pick something where a wrong output is annoying rather than catastrophic. First drafts that a human reviews before sending are good. Automated emails that go directly to customers without review are not a good first automation.

4. Measurable outcome. You should be able to tell whether the automation is working. "Faster" is measurable. "Better" is not. "Tickets responded to in under 2 hours" is measurable. "Customer satisfaction" is not — at least not in the short term.

The tasks that usually score well

• First-draft responses to common inbound queries (support tickets, sales enquiries, partner requests)
• Summarisation — meeting notes, long documents, email threads
• Research briefs — competitive summaries, background on a prospect, topic overviews
• Content reformatting — turning bullet points into prose, adapting copy for different channels
• Internal documentation — turning a Loom recording or meeting notes into a structured doc

Notice what these have in common: they're all tasks where a human is in the loop. Claude produces something; a human checks and uses it. That's the right structure for early automations.

The tasks that usually score poorly

• Autonomous customer communication — anything where Claude acts without review
• Data entry into critical systems — accounting, legal, compliance
• Anything with unpredictable edge cases — situations where "it depends" is the real answer
• Tasks that require judgment about sensitive situations — personnel decisions, legal risk assessments, medical advice

These aren't permanent no-go zones. They're wrong starting points because failure modes are harder to catch and more costly.

The question to ask your team

Get your team together and ask: "What do we spend the most time on that follows a predictable pattern?"

The person who's been doing the job longest usually knows the answer immediately. They've built mental shortcuts for the repetitive parts. Those mental shortcuts are what you're trying to encode into a system prompt or workflow.

Start there. Score it against the four criteria. If it passes, that's your first automation.