Automate first the work that is frequent, already written down, low risk, and easy to check. Lead routing, copying invoice fields, sorting support tickets, and a weekly numbers email are typical starters. Keep a person on anything that moves money, signs a contract, or sends a final answer to a customer.
An AI Implementation Specialist can score the list with you and build the winners. The scoring method below is a framework you can copy onto your own processes.
Which processes are best for AI automation, and which are worst?
The best first processes are boring, repeated, and already done the same way twice a week. The worst are rare, political, or impossible to undo.
Good starters have a trigger you can name (“a form is submitted”, “an invoice PDF arrives”), a finish you can see, and a person who can tell right from wrong in under a minute. Bad starters need judgment that is not written down, or they touch a customer in a way you cannot pull back. “Reply to every angry email in our brand voice, with no review” is a bad starter. “Draft the reply and leave it in a folder for the owner” is a reasonable one.
If you cannot describe the process without pointing at one employee’s memory, automate the documentation first. A time audit shows where the hours go. Use how to run a time audit before you buy a tool. Then read how to delegate and systemise the work. Automation and delegation are neighbours. Some tasks should move to a person, not a model.
How do you score and prioritize automation opportunities?
Score each candidate from 1 to 5 on five factors. Call it VRDRH: Volume, Rules-clarity, Data readiness, Risk, and Human-review need. High scores on the first three, and low scores on risk, are what you build first. Human review is not a penalty. It is a design choice.
Write one row per workflow. Be honest. A 5 is “this happens every day and we can show the steps”. A 1 is “this happened once and only Sam knows”.
| Factor | 1 | 3 | 5 |
|---|---|---|---|
| Volume | A few times a month | A few times a week | Every day, many times |
| Rules-clarity | Nobody agrees the steps | A rough SOP exists | Exceptions are written down |
| Data readiness | The inputs are in someone’s inbox | Most inputs are in one tool | The fields are clean and exportable |
| Risk | A mistake is public, legal, or financial | A mistake wastes an hour | A mistake is easy to spot and undo |
| Human-review need | You want zero review | A sample is checked | Every output is checked before it counts |
Risk is scored so that 5 means safer, matching the other factors. If you prefer “5 means dangerous”, flip it, but do not mix the two directions in one sheet.
Add the five numbers. Do not automate a row just because the total is high if Risk is 1. A daily process that can move money is still a bad first build. Start with a high total and a Risk of 4 or 5, and decide the human-review column on purpose.
The person who fills this sheet should be the internal owner, not the vendor. A vendor will score their favorite tool highly. You are scoring your process.
Should you use rules-based automation or AI for a given task?
Use rules when the decision is a formula. Use a model when the input is messy language or a document, and a person will still check the result. Use a person when the cost of a wrong answer is higher than the time you save.
Rules-based work is “if the order is tagged high risk, email the team and do not capture payment” or “if stock is under ten, create a task”. Shopify’s own Flow templates include examples of that kind, which is covered in AI automations for Shopify stores. No model is required, and you should not add one to look modern.
AI is justified when a person currently reads unstructured text: a PDF invoice, a long email, a pile of reviews. The model extracts or drafts. A rule then routes the draft. The human accepts or edits. If you cannot show a wrong draft to a human before it leaves the building, do not put a model in the path yet.
A simple test: can you write the decision as a table a new hire could follow without asking you? If yes, build a rule. If the table would be “use your judgment” in every row, either keep the person or use a model as a draft only.
What are good first AI projects for a 5 to 50 person business?
Good first projects are internal, reversible, and finished in a week of part-time effort. The fifteen below are examples to score, not a menu Rekruuto has delivered.
Which starter workflows should you score first?
Score these with VRDRH before you build any of them. They are examples, not case studies.
- Route a new lead to the right owner based on fields the form already collects.
- Copy invoice header fields from a PDF into a sheet a person approves.
- Tag a support ticket by topic and leave the reply for a human.
- Draft a weekly numbers email from a sheet you already trust.
- Turn a meeting note into tasks in the tool your team uses, for a person to edit.
- Flag CRM records that are missing the one field sales promised to fill.
- Summarise new reviews into a weekly note. Do not auto-reply.
- Prepare a first-draft product description a merchant edits before publish.
- Remind an owner when a proposal has had no reply for seven days.
- File a receipt email into the right folder and mark the ones that look odd.
- Build a daily list of orders a fraud rule already flagged, for a person to decide.
- Notify a buyer when a bestseller crosses a stock threshold you set.
- Draft a customer “where is my order” reply with the tracking link filled in, and keep it unsent.
- Check a new hire’s access list against a template and show what is missing.
- Produce a Monday backlog of workflows that failed, with the last error attached.
Notice how many of them stop before a customer sees the output. That is deliberate. You can loosen the review after a month of clean samples, not on day one.
A specialist from $2,500 a month is the right builder when several of these will be live and watched. A general assistant at the $1,800 starting price can run the human side, such as approving drafts, if the workflow is simple. The 90-day plan shows how the first build should be sequenced.
What should never be fully automated without human review?
Do not fully automate payments, contracts, legal replies, medical or financial advice, hiring decisions, or the last word to an upset customer.
A model will sound sure when it is wrong. A rule will do exactly what you wrote, including the mistake you wrote. Either way, money leaving an account, a contract going out, or a promise to a customer needs a named person. “The bot sent it” is not a control.
Also keep a human on anything you cannot undo: a deleted record, a public post, a price change across a catalogue. You can automate the preparation. You should not automate the commit until you have watched a sample and written down what “stop” looks like.
If a vendor says their agent is safe to leave unsupervised on day one, ask them to name the log, the cap on spend, and the person who gets the alert. If those three are vague, keep the human in the loop. Maintenance of that loop is the subject of who maintains AI automations.
Frequently asked questions
Which process should a small company automate first?
The one that happens often, has written steps, uses clean fields, and can be undone. Drafts and tags beat unsupervised customer replies.
What is VRDRH?
A five-part score: Volume, Rules-clarity, Data readiness, Risk, and Human-review need. Score each from 1 to 5. Do not start with a high-risk row even if the other scores look good.
When is a rule better than a model?
When you can write the decision as a table. Use a model for messy text, and keep a person on the result until the sample is clean.
What should stay human?
Payments, contracts, legal or medical answers, hiring decisions, and the final message to an upset customer. Automate the prep, not the commit.
How should I use the fifteen examples?
Score them with VRDRH against your own volumes. Start with the highest-scoring row that a person can still undo.
Get matched with your hire
Bring a scored sheet when you get matched with your hire. An AI Implementation Specialist can build the top two rows and leave the rest in a backlog. The 1-week trial with replacement or refund is enough to see whether the first workflow was the right one.