What to Automate First

Automate first the tasks that are frequent, rule-based, and cheap to get wrong. Not the task that annoys you most, and definitely not the one the newest tool happens to demo well. The method is simple: map the workflow you already run, score each step on a few honest criteria, define what “done” looks like, and then pick the simplest layer of automation that actually does the job.

Why start with the workflow instead of the tool?

Tools change monthly; your workflow is the stable thing. When you start from a tool, you end up hunting for problems that fit it, and you usually automate something that did not matter. When you start from the workflow, the right tool becomes obvious, and it is often simpler and cheaper than expected.

The most common failure pattern we see: a team signs up for an automation platform, builds something impressive, and quietly abandons it within a month because it never mapped to how work actually flows. The fix is unglamorous: before touching any tool, write down the process as it exists today, messy parts included.

A useful process map answers five questions for each step:

• What triggers this step?
• What information does it need, and where does that live?
• What decision, if any, gets made here?
• What does the output look like when it is done well?
• Who checks it, and what happens next?

Every “it depends” you write down is a judgment point. Those are the steps that keep a human. Everything between them is a candidate.

Which tasks make good first automations?

Good first automations are frequent enough to repay setup time, rule-based enough to describe completely, and low-stakes enough that a single error is an annoyance rather than an incident. They also have a crisp definition of done, so you can tell at a glance whether the output is right.

We score candidates on five criteria, 0 to 2 points each:

Criterion	Ask yourself	0 points	1 point	2 points
Frequency	How often does this happen?	Few times a year	Weekly	Daily or more
Rule-based	Could you write complete instructions?	Mostly judgment	Rules with exceptions	Clear rules
Stakes per error	What does one mistake cost?	Expensive or public	Annoying but fixable	Trivial to fix
Definition of done	Can you verify the output fast?	Hard to judge	Takes some review	Obvious at a glance
Energy drain	How much do you dread it?	Fine, honestly	Mildly tedious	Actively draining

Score 8 or higher: strong first candidate. Score 5 to 7: viable later, once you have one win behind you. Under 5: leave it manual for now. The energy criterion is deliberate; automating something people hate builds momentum, and momentum is worth points.

Note that “stakes per error” is scored so low stakes earn more points. Your first automation should be one where a bad output wastes five minutes, not one that emails the wrong customer.

Define done before you build anything

Write one or two sentences describing what a correct output looks like, concretely: “a draft reply in the right tone with the order number included,” or “a spreadsheet with no duplicate rows and every date in the same format.” If you cannot write that sentence, the task is not ready to automate.

This step feels skippable and is not. The definition of done becomes your review checklist, your test for whether the automation is working, and eventually the standard an agent gets held to. Fuzzy definitions produce automations that are technically running and practically useless.

Pick the simplest layer that works

Automation is a ladder, and most tasks do not need the top rung:

1. A chat workflow. A saved prompt plus your documented steps, run manually in a chat where the AI has the context it needs. Zero infrastructure, surprisingly powerful. Many “automation” wins are just this.
2. A template. A reusable document, checklist, or prompt structure that makes the manual task fast and consistent.
3. A scheduled or triggered automation. A tool like n8n runs the fixed steps on a schedule or when something arrives, and a human reviews the output.
4. An agent. Software that makes small decisions across steps, with logs and approval gates. Powerful, and last for a reason.

Start at the lowest rung that solves the problem. You can always climb; climbing down after over-building is much more painful. We walk through the full progression, including when agency is actually warranted, in our pillar guide to agent workflows for real people. And whatever layer you choose, design it to run quietly; the argument for that is in Calm AI Systems.

What should you not automate yet?

Hold off on anything judgment-heavy, relationship-heavy, or rare. Judgment-heavy tasks need a human because the rules do not exist yet. Relationship-heavy tasks carry trust you cannot delegate. Rare tasks never repay the setup cost, and the automation breaks silently between uses.

Some honest examples of “not yet”:

• Client-facing negotiation or sensitive replies. The words matter less than the relationship behind them.
• Hiring, performance, or people decisions. Assistance with drafting, yes. Deciding, no.
• Anything you have done fewer than five times. You do not understand it well enough to describe it, and description is the whole game.
• Quarterly or annual processes. By the next run, the automation is stale and nobody remembers how it works.
• Tasks where errors are invisible. If a mistake would not surface for months, you have no feedback loop to improve with.

“Not yet” is not “never.” As you run the manual version and document it, judgment points turn into rules and tasks migrate into automation range. You can follow the rest of this series in our agents and automation topic hub.

Key takeaways

• Start with the workflow, not the tool; map the manual process before evaluating anything.
• Score candidates on frequency, rule-clarity, stakes per error, definition of done, and energy drain.
• Write the definition of done before building; it becomes your review checklist and quality bar.
• Choose the simplest layer that works: chat workflow, template, scheduled automation, then agent.
• Do not automate judgment-heavy, relationship-heavy, or rare tasks yet; document them until the rules emerge.
• One small win that sticks beats five impressive builds that get abandoned.

Common questions

How many things should I automate at once?

One. A single automation that runs for weeks and earns trust teaches you more than a batch launched together, and it builds the review habits everything later depends on. Once the first one is boring, add the next.

What if my highest-scoring task still feels risky?

Add a human approval step and shrink the scope. Automate the gathering and drafting, keep the sending and deciding manual. Most risk in automation lives in the last step, where output leaves your control, so keep that step human until the track record says otherwise.

Do I need an agent, or is a simple automation enough?

Usually a simple automation is enough, and often a template is. Agents earn their complexity only when a task requires small decisions mid-process that fixed rules cannot cover. If you can write the steps as “always do X, then Y,” you do not need agency yet.

How do I know when an automation is actually working?

Compare its output against your written definition of done, every run at first. Track how often you have to correct it. When corrections become rare and minor, the automation is working; when you stop reading the output entirely, you have stopped supervising, which is a different thing.