The safest way into AI automation for a small business is one workflow at a time: pick a task that happens often and where a mistake is cheap to catch, automate the repetitive part with plain rules where you can, use AI only for the part that needs judgement on text, and keep a person approving anything that sends, pays or deletes. That is how we run our own small company. Most of what runs by itself here is ordinary automation with no AI in it at all, and the one AI workflow we run, the one that drafts our blog posts, stops for a person before anything goes public. Below is what we automate, why each piece is built the way it is, and how to choose your first AI workflow.
What small business owners are actually searching for
Before writing this, we looked at a year of Google search interest. Searches for "ai for business" were large, though they fell over the year, and searches for "ai for small business" were smaller but carried telling related queries: "ai tools for small business", "best ai tools for small business" and "ai for small business marketing". In the automation searches, "small business automation" and "whatsapp business automation" were rising, and so was "workflow automation" alongside it.
Those are people looking for tools and starting points, not for theory. So this post starts with practice: the automations that run our own business, including the ones that do not use AI.
What we automate in our own business, and what we do not
Here is everything that runs on its own around our website and publishing today.
| Workflow | What runs automatically | AI involved? | Where a person comes in |
|---|---|---|---|
| Project enquiries | The contact form checks the request comes from our own site, limits repeat submissions from the same email address, runs a bot check, saves the enquiry to a database and copies it to the Notion workspace the owner reads | No | The owner reads and answers every enquiry |
| Job applications | The same checks, saved and copied to a separate Notion database with status tracking | No | The owner reviews every application |
| Uptime | Every 30 minutes, a scheduled job loads the homepage and the contact page; if either fails three times, it emails the owner | No | Only when something is down |
| Backups | Every morning, the enquiries database is exported and kept for 90 days | No | Only for a restore |
| Publishing on schedule | Each morning, if a blog post is due that day, the live site is rebuilt and search engines are notified | No | The owner approved the post before it was queued |
| Blog drafts | A daily run picks an approved topic, researches it, drafts a post and checks it against a quality gate | Yes | The owner reviews and approves every draft before it can reach the site |
Five of the six use no AI. They are rules: check this, save that, alert if this fails. They are cheaper than AI, they behave the same way every time, and when one breaks, it breaks loudly. For most small businesses, the first useful automations look like these, and an AI tool would add cost and unpredictability to them without adding anything.
The sixth is where AI earns its place, because writing a post from research is a job with no fixed steps. We described that workflow in detail in our post on the blog pipeline. The point here is its shape: the AI does the open-ended work, scripts check everything a script can check, and nothing reaches the public until a person has read it.
Choose the first AI workflow by volume and by the cost of a mistake
Plot your candidate tasks on two questions. How often does this task happen? And what does it cost if the AI gets it wrong once?
- High volume, cheap mistakes: start here. Sorting incoming emails into folders, drafting replies a person sends, summarising long documents, tagging enquiries by type, extracting details from forms. A wrong answer is caught before it matters.
- High volume, costly mistakes: automate the draft, not the action. A refund, a price quote or a message to a customer can be prepared by AI and approved by a person in one click.
- Low volume, anything: usually not worth automating yet. The setup and checking cost more than the time saved.
- Anything involving money, legal commitments or deleting records: keep a person in the loop, always.
The second category is where most of the value sits for a small business and where most of the risk does too. Our post on designing the approval step covers what a person needs to see in order to approve a drafted action properly, and how approval goes wrong when people start clicking yes without reading.
Give the AI the least access that does the job
An AI tool connected to your email, your files and your accounts can do everything those accounts can do. OWASP, the security community's standards body, lists excessive agency among the top risks for AI applications, with three root causes: too much functionality, too many permissions and too much autonomy.
For a small business, that becomes a short list of rules:
- Give the AI tool its own account, never the owner's login.
- Separate reading from sending. A tool that drafts replies does not need permission to send them.
- Limit it to one mailbox, one folder or one spreadsheet, not the whole drive.
- Keep payment details, passwords and customer records out of what you paste into AI tools.
- Review what it has access to every time you add a tool or a connection.
The UK's National Cyber Security Centre warns in its guidance on AI and cyber security that language models are prone to prompt injection, where text crafted by someone else makes the model behave in unintended ways, including revealing confidential information. An AI tool that reads incoming emails is reading text written by strangers. That is why its permissions matter more than its instructions. We mapped every permission our own blog agent holds, and the gaps we found, in AI agent permissions.
Measure it before you trust it
Run any new AI workflow in the background for a week or two before relying on it: let it produce its output, have a person do the job as before, and compare. Count how often it was right, how often it was wrong, and how often a mistake would have been costly. The US National Institute of Standards and Technology frames this kind of ongoing measurement as one of the core functions of managing AI risk in its AI Risk Management Framework, alongside governing, mapping and managing.
Keep the comparison going after launch with a handful of real cases, especially the ones it got wrong. Our guide to building an evaluation set from your own failure cases shows how to do that without a data science team.
A starting plan for the next month
- List ten tasks your team repeats every week, with how long each takes and what a mistake would cost.
- Automate the ones that need no judgement with plain rules: forms, alerts, backups, reminders.
- Pick one high-volume task with cheap mistakes for AI, and set it up to draft, not act.
- Give the tool its own account with the narrowest access that works.
- Run it alongside a person for two weeks and count its mistakes.
- Only then let it act on its own, and only for the cases where its record is clean.
If you would like help choosing that first workflow or building it with the approvals and limits in place, start a conversation with us.
