What to Automate First: A Simple Framework

There’s a predictable pattern when businesses start with AI: they automate the exciting thing — a chatbot, an image generator — before the boring thing that actually drains their week. The flashy automation gets shown off once and quietly abandoned. The boring one would have paid for itself every day.

Choosing what to automate first matters more than choosing the tool. Here’s a simple framework I use with clients, with real examples and a short list of what to leave alone for now.

Score each task on five questions

Take the tasks that eat your team’s time and score each from 1 (low) to 5 (high):

1. Frequency — how often does it happen? Daily beats monthly.
2. Time per instance — how long does one take a person?
3. Error cost — how bad is it if it’s done wrong or late? (This one is a trap — see below.)
4. Data readiness — is the information the task needs already available and reasonably tidy?
5. Reversibility — if the AI gets it wrong, is it easy to notice and undo before any harm?

The first two tell you how much a task is costing you. The last two tell you how safe and feasible it is to hand over. Start where both are high: a frequent, time-consuming task that uses data you already have, where mistakes are easy to catch.

Here’s how a few common tasks tend to score:

Task	Frequency	Time	Data ready	Reversible	Do first?
Answering common enquiries	5	3	4	5	Yes
Follow-up emails after a lead	5	3	4	5	Yes
Weekly performance report	4	4	4	5	Yes
Drafting proposals / quotes	3	5	3	3	Assist only
Posting final replies to reviews	4	2	3	2	Keep human
Approving refunds / pricing	2	2	2	1	Keep human

The pattern is clear: the best first automations are high-frequency, medium-effort tasks where the data exists and a mistake is obvious and cheap to fix.

Drain = frequency × time. Feasibility = data readiness + reversibility.

The error-cost trap

“Error cost” is where most people go wrong. A high error cost doesn’t mean “don’t automate” — it means “don’t fully automate.” There are two safe modes:

• Assist — the AI does the first draft; a person checks and sends. Right for anything customer-facing or high-stakes: proposals, important emails, anything with your name on it.
• Full auto — the AI does the whole thing with no human in the loop. Only for tasks where an error is cheap and easy to catch: internal summaries, tagging, first-line enquiry replies that clearly say a human will confirm.

When in doubt, start in assist mode. You keep almost all of the speed and lose almost none of the safety. Once you’ve watched it get things right for a few weeks, you can widen the gate.

Real first automations that tend to pay off

For most small businesses, the same handful come up again and again:

• First-line enquiry handling — answering the same twenty questions people ask, capturing their details, and flagging anything unusual to a human.
• Follow-up sequences — the polite “just checking in” emails that never get sent because everyone’s busy.
• Reporting — turning your numbers into a plain-English weekly summary instead of a spreadsheet no one reads.
• Repetitive writing — product descriptions, listings, standard replies: drafted by AI, tidied by a person.
• Meeting and call notes — turning a recording into a summary and a task list.

None of these are glamorous. All of them give you time back every single week — which is the entire point.

A worked example: scoring “follow-up emails”

Say your team is slow to follow up on enquiries. Score it 1–5:

• Frequency: 5 — it should happen after every enquiry, several times a day.
• Time: 3 — a few minutes each, but they add up and quietly get skipped.
• Error cost: 2 — a slightly-off follow-up is easy to fix; the real cost is not sending one.
• Data readiness: 4 — you already have the enquiry details and your standard answers.
• Reversibility: 5 — you read the draft before it goes out.

High drain, high feasibility, low risk — a textbook first automation. Let AI draft a tailored follow-up from each enquiry, have a person skim it, and send. You stop losing warm leads to a busy inbox, and nothing goes out that a human hasn’t seen.

Agentic AI changes the boundary, not the method

In 2026, more tools can browse, use software and take actions rather than only draft an answer. That can remove additional hand-offs, but it also increases the cost of a misunderstood instruction. The NCSC’s May 2026 guidance on adopting agentic AI carefully recommends starting small, choosing low-risk tasks, restricting access and planning for failure.

Before an AI system can act, add four questions to the score:

1. Permission: what is the narrowest data and system access it needs?
2. Approval: which actions must pause for a named person?
3. Evidence: what inputs, outputs and actions will be logged so an error can be understood?
4. Recovery: can the action be reversed, and who knows how?

Drafting a follow-up and placing it in a review queue is one risk level. Sending it, changing a customer record and scheduling another action is a different system. Do not treat “the draft looked good ten times” as evidence that every connected action is safe.

Automation level	What AI does	Appropriate starting control
Suggest	Recommends a next step	Person decides and acts
Draft	Prepares the output	Person checks every output
Execute with approval	Prepares an action in another system	Named approver confirms before execution
Limited autonomy	Acts inside an allowlisted, reversible boundary	Logging, monitoring, spending/action limits and tested rollback

For most first projects, “draft” or “execute with approval” captures much of the benefit without pretending uncertainty has disappeared.

What to leave alone (for now)

Hold off on automating anything where all three are true at once: the data is messy or missing, an error is expensive, and a mistake is hard to spot. And keep a person on the moments that build or break trust — the apology, the negotiation, the difficult refund. AI can help you draft those; it shouldn’t send them.

Start with one, keep the human in

Pick the single highest-scoring task from your list. Automate just the first draft of it, keep a person checking the output, and watch the time it saves for two weeks. If it holds up, widen it and move to the next. Automating one boring task properly beats half-automating five exciting ones.

If you’d like to score your own tasks and set the first one up properly, that’s what we work through in AI automation training — bring your real workflow and we’ll map it together — or book a private lecture to start.