What to Automate First: Ranking Work by Payback, Not by Annoyance

Ask an owner what they would automate first and you will almost always get the same kind of answer. Expense reports. The monthly board pack. That one recurring spreadsheet.

These are the tasks that irritate them personally. They are rarely the tasks with the biggest payback.

Annoyance and payback are different measurements. Annoyance measures how much a task grates on the person describing it. Payback measures hours, multiplied by frequency, weighted by what it costs when the work goes wrong. The two overlap by accident, not by design. The month-end pack that ruins one afternoon a month is a fraction of the payback of the support queue nobody complains about, because that one has been absorbed into everyone’s day.

Automate by payback. Here is how to measure it.

Four factors, scored 1 to 5

Take each candidate process and score it on four things.

Volume, meaning how often this happens. 1 is monthly or less. 3 is a few times a week. 5 is many times a day.

Time per instance, meaning how long one takes end to end. 1 is under two minutes. 3 is around fifteen minutes. 5 is an hour or more.

Error cost, meaning what one mistake costs you. 1 is trivial and caught internally. 3 is an awkward conversation or a small rework. 5 is a lost client, a missed deadline, or a compliance problem.

Documentation, meaning how well the rules are written down. 1 means it exists only in someone’s head. 3 means a rough checklist and a lot of tacit knowledge. 5 means a new hire could follow the written process without asking anyone.

Now do the arithmetic in two steps, which matters more than it sounds.

Step one: volume times time equals hours. This is the size of the prize. It is the only part of the score that tells you how much work there is to reclaim, so it does the heavy lifting. A 5 by 4 process has twenty points of raw opportunity; a 2 by 2 process has four, and no amount of scoring elsewhere will rescue it.

Step two: adjust for error cost and documentation. These do not create the prize. They decide whether you can collect it. A high error cost raises the value of removing human inconsistency, but only if documentation is high, because a poorly documented, high-consequence process is the one place an agent will do real damage. Treat documentation as a multiplier on your confidence, not as another additive point.

A useful rule of thumb: hours score, plus error cost, all multiplied by documentation divided by 5.

Five typical processes, ranked

Illustrative scores for a 20-person B2B service firm. Your numbers will differ; the shape of the answer usually does not.

Support email. Volume 5, time 3, so 15 hours points. Error cost 3, documented 4. Score 14.4. Rank 1.

Inbound call handling. Volume 5, time 2, so 10. Error cost 4, documented 4. Score 11.2. Rank 2.

Invoice chasing. Volume 3, time 2, so 6. Error cost 4, documented 5. Score 10.0. Rank 3.

Outbound prospecting. Volume 4, time 3, so 12. Error cost 2, documented 2. Score 5.6. Rank 4.

Month-end reporting. Volume 1, time 5, so 5. Error cost 5, documented 2. Score 4.0. Rank 5.

Support email usually wins. High volume, meaningful time per instance, and, critically, a large archive of past replies that functions as documentation whether or not anyone wrote a policy. You already have thousands of worked examples. That is why it loads fast and performs well.

Inbound call handling comes close. Slightly less time per call, but a higher error cost, because a missed call is an unrecovered lead. Firms consistently underprice this one.

Invoice chasing is the quiet winner. Lower volume, but the rules are genuinely written down, and the error cost is direct cash. It is also unglamorous, which is why nobody nominates it.

Outbound prospecting scores badly and gets picked anyway. It looks like the growth lever. But the rules are rarely written down, quality judgment is high, and getting it wrong damages your name in your own market. Automate it after you have proven the pattern elsewhere.

Month-end reporting scores worst despite being the most annoying. Once a month is once a month: five points of raw hours cannot beat fifteen, however much the afternoon hurts. It also tends to be poorly documented, because it lives inside one person’s spreadsheet and one person’s habits.

Is it written down? is the best predictor there is

Of everything on that list, documentation is the factor that most reliably separates automations that work from automations that get switched off in week six.

The reason is simple. An agent does not infer your intent. It executes a described process. If the description is vague, it will act confidently on the vague version and you will discover the gaps in front of a client.

There is a second reason, less obvious. Undocumented processes are usually undocumented because they were never agreed. Three people do it three ways. Writing it down forces the argument that has been quietly deferred for years, and the argument, not the agent, is where a lot of the value actually is.

When nothing is written down, do this. Take the last twenty real instances of the task. Not the ideal version, the actual twenty. Write down what was decided in each and why. Patterns emerge fast: usually four or five decision rules cover seventeen of the twenty, and the remaining three are the genuine exceptions. Those exceptions become your escalation rule. That exercise takes a couple of hours and it is the highest-return work in the whole project. Do it before you buy anything.

The sequencing rule

Three principles, in order.

Start where the work is repetitive, high-volume and low-judgment. Your first agent is not there to be impressive. It is there to prove the pattern and give your team a reason to trust the next one. Pick something boring and frequent.

Add governance before you add the second agent. Limits on what each agent may do unsupervised, a clear escalation path, and an audit trail you can actually read. Doing this at one agent is a morning’s work. Retrofitting it at four is a project, and by then you will have already had the incident that made you want it.

Only automate customer-facing work after you have watched it run silently. Run the agent against real traffic in draft mode. It produces the response; a human sends the one they would have sent anyway. Compare. Two weeks of that will tell you more than any evaluation, and it costs you nothing but attention. Turn it live only when the drafts stop needing edits.

Where to go next

The free Workflow Automation Audit workbook on our resources page walks you through scoring your own processes on these four factors, across 45 common small-business tasks, with the ranking table already built.

Once you have picked one, the 90-Day AI Rollout Tracker gives you the sequence, from loading through silent running and governance to go-live, with the checkpoints that stop a good first automation from stalling in month two.

Both are free at www.agentic-ai-adoption-now.com/resources.