To decide what to automate first, score each candidate task on five questions from 1 to 5, weight them, and then reject anything that fails one of four hard gates. The highest score that passes every gate goes first. Typical first picks for a service business are acknowledging enquiries, collecting basic details and sending reminders; pricing exceptions, complaints and messages to people you have no consent to contact are never first.
Step 1: score each task on five questions
Pick a task, count honestly, and score it. The bands are our suggested ones; adjust them to your size, but keep them fixed across candidates.
| Question (weight) | Score 1 | Score 3 | Score 5 |
|---|---|---|---|
| Volume: how often does it happen? (×2) | Under 20 a month | 100 to 300 a month | Over 500 a month |
| Time: effort per item (×1) | Under a minute | About 5 minutes | 15 minutes or more |
| Rules: can you state the rule? (×2) | Depends on who does it | A rule with exceptions | Two sentences, no exceptions |
| Data: are the inputs in one place? (×1) | In people’s heads or chats | Sometimes in a sheet | In one system, in fields |
| Speed: does delay lose business? (×2) | A day’s wait is fine | Same day | Minutes matter |
Total = 2×Volume + Time + 2×Rules + Data + 2×Speed, out of 45. Volume, rules and speed carry double weight because they decide whether automation pays back, whether it will behave, and whether anyone notices it working.
Step 2: four hard gates that overrule the score
A single “yes” to any of these means the task is not a first candidate, however high it scores:
- 01Judgement gate. Does it need someone to decide a price, an exception or a complaint? Automation can fetch the context; a person decides.
- 02Consent gate. Would it send a message to someone without a basis to contact them on that channel, in that country? Rules differ by country; see consent rules by country.
- 03Undo gate. If it goes wrong, can you reverse it? Money moved, a legal statement sent or a record deleted fails this gate.
- 04Owner gate. Is the process written down, with one named person who owns it? If not, automating it locks in the confusion.
A worked example
The scores below are illustrative, for an invented quote-based service business that gets enquiries on WhatsApp. They are not client data. The point is the method: score your own tasks with your own counts.
| Task | V | T | R | D | S | Score | Gate |
|---|---|---|---|---|---|---|---|
| Acknowledge a new enquiry | 5 | 2 | 5 | 4 | 5 | 36 | None: first |
| Collect name, service and city | 5 | 3 | 4 | 4 | 4 | 33 | None: second |
| Appointment reminders | 4 | 2 | 5 | 5 | 3 | 31 | None: third |
| Chase unanswered quotes | 3 | 3 | 3 | 3 | 4 | 26 | Consent: check first |
| Calculate a quote | 3 | 4 | 2 | 3 | 3 | 23 | Judgement: a person |
| Handle a complaint | 2 | 4 | 1 | 2 | 2 | 16 | Judgement: never |
Read it as a rule of thumb: 30 or more with no gate is a good first build; 20 to 29 is worth automating partly, so a person is helped rather than replaced; under 20, or any gate, stays human. Here the first three tasks would be built first, quote chasing waits for a consent check, and the last two stay with people. For the first one, see how to count a reply deadline in working hours.
Three things never to automate
- 01A final price or an exception to it. Let software gather the facts; a person commits.
- 02A complaint or a dispute. Acknowledge it automatically if you must, then hand it to a named person at once.
- 03A message to someone you have no consent to contact. An automation that scales a consent mistake scales the fine.
What to do with the score
Start with the top task and write down how you will know it works: the count it handles, how often a person had to step in, and what went wrong. Review after two weeks, then take the next. Scoping a larger project follows the same logic; see what AI automation services include and what drives their cost.
What we do and do not claim: we design and build automations for service businesses, and this sheet is how we would rank your candidates. We do not publish automation results, because we have no documented ones to show yet. See AI automation for what an engagement covers.