Five AI automations worth building first in a small business

Every small business has a handful of chores that eat a few hours a week each: reading enquiries, retyping receipts, writing up meetings, chasing reviews. None of them is hard. They are just constant, and they land on whoever is least able to say no.

These five are where we suggest most small teams start. Each one deals with messy language, which is where a model earns its place, and each one keeps a person approving anything that leaves the building or touches money. If you want the broader picture first, our post on what AI can do for a ten-person business covers where it helps and where it fails.

1. Inbox triage for new enquiries

Today, someone reads every message that comes through the contact form or the general inbox, decides what it is, forwards it, and retypes the useful parts into the CRM. Urgent requests that arrive on a Friday evening wait until Monday.

The automation passes each new message to a model with a short, specific instruction. It classifies the message (new work, existing client, supplier, job applicant, or spam) and pulls out the sender’s name, what they need, how soon, and any budget signals. It writes that record to the CRM, puts a draft reply in the shared inbox, and sends a text to whoever is on call if the request looks urgent.

Nobody receives a reply until a person reads the draft and presses send.

What goes wrong: odd enquiries get filed in the wrong bucket, so give the model an “unsure” category that always routes to a person. Models also like to fill blanks, so tell it to leave the budget empty when none is stated, and spot-check that it does. And because email is untrusted input, the step that reads messages should have no power to send, delete, or forward anything on its own.

2. Quote drafting from past quotes

Today, the estimator opens three old quotes for similar jobs, copies the line items into a new one, and adjusts. It works, but it takes an hour, and old mistakes get copied along with the good parts.

The automation takes the job description or site notes, your current price list, and a library of past quotes, then produces a draft in your template: scope wording, line items, quantities, and exclusions. Prices come from a lookup against the price list, not from the model’s memory of what you charged last spring. Totals and tax are calculated by the quoting tool or spreadsheet, never by the model.

A person approves every price before the quote goes out. The model drafts; it does not set prices.

The usual failure is an exclusion or line item carried over from an unrelated old job. The price list goes stale, and the drafts quietly inherit last year’s numbers. Put a “last updated” date on the price list and have the automation refuse to run when it is more than a quarter old.

3. Receipt and invoice capture

Today, receipts sit in a jacket pocket, a phone camera roll, and an email folder until month end, when someone types them into the bookkeeping system.

QuickBooks Online and Xero both include receipt and bill capture of their own, typically a forwarding email address plus a mobile app that reads photos. Check what your plan includes and start there before building anything custom. The AI layer on top suggests a category based on how you have coded that vendor before, splits out HST, and flags anything unusual, such as a new vendor, a foreign currency, or an amount far above the usual.

The bookkeeper confirms each category before anything posts. For the first few months, nothing posts automatically.

Expect ordinary errors. A crumpled receipt gets read as $81.00 instead of $18.00. The same receipt arrives twice, once photographed and once emailed. Tax is split wrongly on a mixed purchase. These are the same mistakes a rushed person makes, and the confirmation step exists to catch them.

4. Meeting and call summaries into your task tool

Today, the action items from a meeting live in one person’s notebook, or nowhere.

The automation takes the transcript, writes a short summary of decisions, and turns each action item into a draft task with a suggested owner and due date in Planner, Asana, or whatever your team already uses.

The person who ran the meeting reviews the list before tasks are assigned. That takes two minutes and catches most of the problems.

The common mistakes are small: the wrong person gets assigned, or the model invents a due date nobody agreed to. Recording also raises consent and confidentiality questions of its own, which we cover in our guide to setting up AI meeting notes safely. Some meetings should not be recorded at all.

5. Review requests and follow-ups after a job

Today, the owner means to ask happy clients for a review and rarely does.

When a job is marked complete or the invoice is paid, the automation waits a few days, then sends a short thank-you with a review link. A model writes one or two personal lines from the job notes, so the email mentions the actual work. If there is no response, one follow-up goes out a week later, and then it stops.

A person approves the templates once and sees each day’s send list.

Two rules keep this one out of trouble. Keep the trigger uniform, such as “invoice paid”, and avoid filtering for clients you expect to be pleased, because Google’s review policies prohibit selectively soliciting positive reviews. Canada’s anti-spam law (CASL) also applies to commercial email. An existing client relationship usually gives implied consent for a limited period, but confirm how it applies to your situation.

Notice that most of this one needs no model at all. That leads to an important point.

When plain automation beats AI

If the rule can be written down, write it down and skip the model. “When an invoice is paid, wait three days, then send this email” is a fixed rule, and a scheduled workflow does it perfectly every time. A model adds cost, a little delay, and the chance of a different answer on a different day. Use one only for the parts where the input is messy human language. Our post on when not to use an LLM has a fuller checklist.

Choosing the tool

Zapier is the easiest place to start and connects to almost everything. It charges by tasks, though, and every step of a multi-step workflow counts. A triage flow on a busy inbox can climb through pricing tiers faster than you expect, so watch your usage in the first month.

Make is more visual and usually cheaper per operation. It is also fiddlier, and it bills by operations, so the same creep applies.

n8n can be self-hosted, which keeps data on infrastructure you control and removes per-task pricing. In exchange, you own the updates, backups, and security of another server. There is a hosted version if you want the tool without the upkeep.

Power Automate is the natural fit for Microsoft 365 shops, and basic use is included with many business plans. Premium connectors and AI features are licensed separately, so check before you design around them.

Custom code makes sense when volume is high, the logic is specific to your business, or the data has to stay in your own systems. It costs engineering time up front, and someone has to maintain it. When the data is sensitive enough that it should not reach a third-party model, our private AI work covers the hosting side.

How to measure it

Before you switch anything on, time the manual version for a week. Then track three numbers.

Minutes saved per week, multiplied by the loaded hourly cost of the person doing the work. If triage saves six minutes on each of 15 enquiries a week, that is 90 minutes. At a loaded cost of $45 an hour, it is about $67 a week, or roughly $3,500 a year, before you subtract the tool’s cost.

Error rate. Pull 20 outputs a week and count how many were wrong.

Override rate, meaning how often the person reviewing materially changes the draft. A high rate means the automation is not saving much. A rate near zero is worth a second look too, because it can mean the reviewer has stopped reading.

One at a time

Pick one automation, preferably the one with the most volume and the lowest stakes. That is usually inbox triage or meeting summaries. Run it for a month, measure it honestly, and fix what breaks. Then build the next one.

If you would like help choosing the first automation and building it properly, that is what our AI automation work does. Write us a short brief about the chore that eats the most time in your week, and we will reply in writing within one business day.