What to automate first: a practical guide to agentic automation

Most companies I talk to want to "use AI agents" but aren't sure where to start. The usual mistake is starting with the most impressive idea instead of the most suitable one. A good first automation is boring, frequent, and easy to check.

Score your workflows

List the repetitive work your team does each week, then score each item on five questions:

Volume: does it happen many times a week?
Clarity: could you write down the rules for a new hire in a page?
Checkability: can a person verify the result in seconds?
Reversibility: if it goes wrong, can you undo it cheaply?
Data access: is the information it needs available in a system, not in someone's head?

The best first candidates score high on all five. Anything that's rare, vague or irreversible can wait.

Patterns that work

A few shapes come up again and again:

Triage and draft. Agents sort incoming requests (support tickets, leads, invoices) and draft a reply or action for a person to approve.
Extract and structure. Pull fields out of documents, emails or PDFs into a system of record.
Research and summarise. Gather information from several sources into a short, consistent brief.
Report and monitor. Produce the weekly report nobody enjoys writing, and flag what changed.

All of these keep a person in the loop for the final decision, which is exactly where you want them at the start.

Start with a human in the loop

The first version of any automation should draft, not act. Let the agent prepare the work and a person approve it. Track how often the person changes the draft. When that number gets low and stays low, you've earned the right to let the agent act on its own for that task.

Keep sensitive data private

If the workflow touches customer data, contracts or anything confidential, you don't have to send it to a third-party API. Open-weight models can run on your own servers, which keeps the data in-house and removes the per-request bill. For many internal tasks, a well-chosen local model is more than good enough.

Measure it

Before you automate, write down how long the task takes today and how often it goes wrong. Afterwards, measure the same two numbers. An automation that saves time but adds errors isn't a win, and you'll only know if you measured first.

Common mistakes

Automating a process that was already broken. Fix the process first.
Skipping the human review step to look impressive.
Building one giant agent instead of several small ones with clear jobs.
No logging, so nobody can tell why the agent did what it did.

If you want help mapping your workflows and choosing what to automate first, that's exactly what my agentic automation consulting covers.