Exploring AI Automation For Small Business Workflows

We have been putting AI tools through our own workflows for a while now, partly because clients keep asking and partly because we would rather answer from experience than from a vendor's landing page. This is an honest interim report: what has held up, what has not, and how we now decide before committing.

The headline finding is unglamorous. The tools work well on a narrower set of tasks than the marketing suggests, and on that narrower set they work better than we expected.

Where it has genuinely saved time

A pattern emerged quickly. Automation earns its place where the task is high-volume, low-stakes, and easy to check. All three conditions, not two.

• First drafts of repetitive copy. Product descriptions across a large catalogue, alt text, meta descriptions. The output needs editing, but editing a draft is faster than facing an empty page eighty times.
• Reformatting and extraction. Pulling structured fields out of messy input — supplier lists, enquiry text, inconsistent spreadsheets. Genuinely reliable, and the errors are obvious when they happen.
• Summarising for triage. Not to replace reading something, but to decide what to read first.
• Rubber-duck review. Asking a model what is unclear about a page or what a draft fails to answer surfaces real gaps, because it has no idea what you meant.

The common thread is that a wrong answer is cheap and immediately visible. That is what makes the speed worth having.

Where it has cost more than it saved

The failures were more instructive. Three kinds:

Anything requiring facts we have not supplied. Asked for something specific it does not know, a model produces something plausible instead — a confident, well-written answer that is wrong. In a client-facing context that is not a time saving, it is a liability, and verifying every claim costs more than writing it yourself.

Work where checking is as hard as doing. Anything requiring judgement about a specific business, a real relationship, or a decision with consequences. If verifying the output means reconstructing the reasoning, automation has moved the work rather than removed it.

Tools that need managing. This one surprised us most. Several automations technically worked and still lost on net, because each one added a thing to configure, monitor, and fix when it silently stopped. A small team can absorb a limited number of moving parts; every automation spends some of that budget.

The question is not "can this be automated" — increasingly the answer is yes. It is "will the automation cost less attention than the task did", and that answer is often no.

How we decide now

Four questions, in order. A no anywhere stops it.

• How often does this actually happen? Measured, not estimated. Automating something monthly rarely repays setting it up.
• What does a wrong answer cost? If it is embarrassment in front of a customer or a bad number in a report someone acts on, the bar rises steeply.
• Can it be checked in seconds? If not, expect the checking to become the new task.
• Who fixes it when it breaks? Not if. An automation with no named owner becomes an outage nobody notices for a week.

The part worth saying out loud

There is real pressure to present AI capability as further along than it is, and we would rather not add to it. What we can say honestly is that we are using these tools daily on internal work, we have a clearer picture than we did six months ago of where the line sits, and we are not going to recommend automating something because it is currently interesting.

The useful version of this for a small business is almost never "replace a role". It is finding the two or three repetitive, checkable tasks that quietly consume hours a week, and taking those back. That is a smaller claim than most of what you will read, and it is the one we can stand behind.

If you have a workflow you suspect falls in that category, it is worth a conversation before it is worth a tool. Our contact form reaches us, or the packages page explains how ongoing work is structured.