In most cases, what to automate with AI in a small business comes down to three questions. Does the task repeat weekly? Does its answer already live in a system? Does waiting cost you a sale? Inquiries, bookings, client onboarding and orders retyped between tools usually pass. Complaints, negotiated prices and anything nobody has written down stay with a person.
Updated
What to automate with AI in a small business: score it in five questions
The three questions above do most of the work. Two more catch the tasks that look easy and aren't: can you write the steps on one page, and what happens the day the automation gets one wrong.
Score each candidate from 0 to 2 on every question in the table, then add it up. 8 to 10: automate it now. 5 to 7: automate the preparation (the draft, the form, the record) and let a person approve it. 0 to 4: keep it human for now, and score it again once the process is written down.
Do the scoring with the person who does the task every day, not the one who supervises it. They know how often it really happens and where it gets stuck.
| Question | 2 points | 1 point | 0 points |
|---|---|---|---|
| How often does it happen? | Every day | Every week | Once a month or less |
| Where does the answer live? | In a system: your booking panel, CRM or stock software | In a spreadsheet someone keeps up to date | In one person's head |
| What does waiting cost? | A sale or a booking | An annoyed customer | Not much |
| Can you write the steps on one page? | Yes, and they rarely change | Mostly, with a few exceptions | No, every case is different |
| What if it gets one wrong? | Someone catches it before the customer does | You send an apology | Money, health or legal trouble |
AI automation ideas for small business, scored
Here is where the usual candidates tend to land. The scores are our own rule of thumb, not a benchmark: yours will move with your volume and with where your data lives.
In the table, look at the last column. Several high scorers don't need AI at all. Moving an order from your store to your invoicing tool is a rule, and a rule is cheaper to run and easier to check than a model.
| Task | Typical score | What to do | AI or plain workflow? |
|---|---|---|---|
| Answering the same questions on WhatsApp or web chat | 9 | Automate, with a clear way to reach a person | AI: every customer writes differently |
| Checking availability and taking a booking | 8 to 10 | Automate, if availability lives in a system | AI for the chat, workflow for the booking |
| Moving orders, stock and invoices between tools | 9 | Automate | Workflow, no model needed |
| Onboarding a new client or distributor | 7 | Automate the form, the access and the records; a person approves | Workflow |
| Qualifying a lead before a sales call | 7 | AI asks and sorts; you decide who to call | AI |
| Drafting email replies | 6 | AI drafts, a person sends | AI |
| Turning call notes into CRM updates | 6 | AI summarizes, a person checks | AI |
| A weekly report from your own numbers | 6 | Automate the data pull; read it yourself | Workflow, AI optional for the summary |
| Complaints and refunds | 3 | Keep human; automate only the routing | Neither, for now |
| Custom quotes and discounts | 3 | Keep human | Neither, for now |
| Medical, legal or tax advice | 1 | Always a person | Neither |
AI automation examples from projects we've shipped
Three projects by Kivo Marketing, each one automating a single job. None of them tries to automate the whole company.
The conversation. For Al Teu Aire, a WhatsApp agent books events for more than 40 venues in Barcelona. It reads availability and prices live from the team's own panel and answers in Spanish and Catalan. It fits the first-candidate profile: the questions repeat, and the answer already lives in the team's panel.
The paperwork. In the Nuderma Partner Portal, distributors apply in five steps, sign in with a one-time code instead of a password and see the catalog and prices for their account type. Every application lands on an admin desk where it gets approved. The repetitive part runs by itself; deciding who becomes a partner stays with the team.
The plumbing. JD Yuversity's academy and shop in Miami are wired to Zoho for stock, orders and invoices, with n8n in between. A job like this doesn't need a language model: the point is that an order placed on the site reaches Zoho without anyone typing it in again.
And one of our own: Kivo CRM, our product, comes with an AI copilot that replies with the real data in the CRM. That's the last rung of the ladder. First the data lives in one place, then you can ask it questions.
Does the task need AI, or just a workflow?
A workflow follows rules that fit on a sticky note. When an order is paid, create the invoice. When a form is complete, send it for approval. It does the same thing every time, and when it breaks, you notice: a missing field, a run that stopped.
A language model earns its place when the input is free text nobody controls: a WhatsApp message with the date buried mid-sentence, a question with a typo, a request that mixes two things. The model works out what the customer wants. The workflow writes it into your systems. Many useful automations need both.
One condition comes before any of it. If the price or the availability isn't in a system, the AI will have to guess it. In that case your first project isn't AI at all: it's getting that data into a system.
- Workflow: the input is a form, a payment or a status change, and the output always has the same shape.
- AI: the input is written by a person, in their own words, maybe in two languages.
- Both: the AI reads the message, a workflow checks the system and records the result.
- What we build with: n8n for workflows between tools, OpenAI or the Claude API as the model, and TypeScript on Node.js when an agent holds a conversation and reads live data. Weighing workflow tools? See n8n vs Make vs Zapier.
What a small business shouldn't automate yet
A low score isn't a failure. It usually means the task needs a person's judgment, or the process isn't ready to be handed to software.
- Complaints and refunds. The customer is already upset. Automate the routing so the message reaches the right person fast, and let that person write the reply.
- Prices you negotiate. If the price changes per customer and isn't stored anywhere, an agent would have to make one up.
- Medical, legal or tax advice. In a clinic, an agent can book the appointment. It shouldn't tell anyone which treatment they need.
- Tasks that happen once a month. Building and maintaining the automation will likely take longer than doing the task by hand.
- Processes nobody has written down. Write them first. If two people do the task two different ways, the automation will copy one of them, maybe the wrong one.
- Decisions about people. Who you hire, who gets credit, who becomes a distributor. In the Nuderma Partner Portal, every application is still approved on the admin desk.
How to scope your first AI automation
Take the task with the highest score and shrink it until it fits in one sentence. “Answer price and availability questions on WhatsApp and take the booking” can be built. “Automate customer service” can't.
That sentence is where our brief starts. Kivo Marketing runs every project in four steps (brief, full design, real build and improvement after launch), with the same team designing and coding it in Miami and Barcelona, in English and Spanish. The whole service is on our AI automation page.
Before anything gets built, write down five things:
- One task, one channel. WhatsApp or email, not both at once.
- One source of truth. The system every price, date and stock level will come from.
- A handover rule. Which cases go to a person, and to which person.
- A before picture. How the task runs today and who does it, so you can compare after launch.
- A review date. Read what the automation did in its first weeks and tune it.
Work we have shipped
- Al Teu Aire
A WhatsApp agent that books events for more than 40 venues in Barcelona. It reads availability and prices live from the team's own panel and answers in Spanish and Catalan.
See the project - Nuderma Partner Portal
Five-step onboarding for distributors, sign-in by one-time code, catalog and pricing by account type, and an admin desk that approves every application. Next.js, n8n, Zoho CRM and Stripe.
See the project - JD Yuversity
Korean aesthetics academy and shop in Miami, wired to Zoho for stock, orders and invoices. PHP, Stripe, n8n, Zoho Inventory, Zoho CRM and Meta CAPI.
See the project
Frequently asked questions
What is AI automation for a small business?
It's software that does a repeated task for you, with an AI model handling the part that needs to read or write language. The rest is ordinary automation: rules that move data between your tools. In a small business that usually means answering customers, taking bookings, keeping orders in sync and preparing work for a person to approve.
Can I automate my business with ChatGPT or Claude myself?
Your own work, yes: drafting, summarizing, sorting notes. Putting a model in front of customers is a different job. It needs your live data, a rule for when to hand over to a person and a log you can review, which means connecting it to your systems through the API (OpenAI or the Claude API), not a chat window.
Are there free AI automation tools for small businesses?
Plenty of automation tools have a free plan or a trial, and they're a fine way to test an idea. What's never free: model usage, which AI providers bill by consumption, the time it takes to connect your systems, and the upkeep when one of them changes. Test for free; pay for a proper build when customers depend on it.
Should I hire an AI automation agency or do it myself?
If the task lives inside one tool and only your team sees the result, try it yourself first. If customers see it, it touches two or more systems or it has to work in two languages, an agency saves you the trial and error. Whoever you hire, ask to see one of their systems running in production before you sign.
How much does a first AI automation cost?
There's a one-off build and a running cost: hosting, model usage and, on WhatsApp, Meta's messaging fees. A plain workflow with no model is usually the cheapest place to start. Kivo Marketing doesn't publish prices yet and quotes after the brief. For a WhatsApp agent, the running side is broken down in our WhatsApp AI agent cost guide.
Will AI automation replace my staff?
It replaces tasks, not the people who know the business. A sensible split: the automation handles the routine, prepares the paperwork and keeps your systems in sync, while your team takes the exceptions, the upset customers and the decisions. Someone also has to own the automation: read what it did, fix the data it reads and decide what changes.