Why most automation pilots never get a second look
Most Hong Kong SME owners we speak to have already tried some form of automation. A chatbot on the site, a form that pipes leads into a spreadsheet, maybe an AI tool that drafts replies. The pattern repeats: it gets switched on with some enthusiasm, then quietly stops being used within a month. Nobody decides to abandon it. It just fades.
The reason is rarely that the technology is bad. It is that nobody defined what "working" would look like before the tool went live. Without that definition, there is no way to tell a slow start from a bad fit, and no way to defend the spend when the next invoice arrives.
This article gives you a concrete test you can run in a week, using work you already do. It is not about whether AI is good or bad. It is about how to find out, cheaply, whether a specific automation earns its place in your business.
Step 1: Pick one task, not one department
"Automate customer service" is not a task. "Draft the first reply to the five most common WhatsApp enquiries" is a task. The narrower you go, the faster you get a real answer.
Choose a task that meets all three of these conditions:
- It repeats. At least several times a week, ideally daily.
- It has a clear finish line. You can say whether a given output was acceptable or not.
- Someone currently does it manually. This matters more than it sounds — if the task is not being done today, automating it does not save anything.
Good candidates for a Hong Kong SME: replying to stock availability questions, converting enquiry emails into a structured log, drafting quotation cover notes, tagging incoming messages by product line. Poor candidates: "handling complaints", "managing social media", "improving customer experience". Those are outcomes, not tasks.
Write down the manual baseline first
Before you switch anything on, record how the task is done now. For one week, note:
- How many times it happens
- How long each instance takes, start to finish
- Who does it and what else they could be doing instead
- How often the manual version gets it wrong or needs rework
A rough count is fine. The point is to have a number you can compare against later, not a perfect audit. If you skip this, you will end up judging the automation on how it feels, and feelings are easily swayed by a good demo.
Step 2: Set a pass mark before you see the results
This is the step almost everyone skips, and it is the one that decides whether the test tells you anything. Decide in advance what result would make you keep the tool, and what result would make you switch it off.
A workable pass mark has three parts:
- Quality: for example, at least 8 out of 10 drafts are usable with light editing.
- Time: the total human minutes per instance drop by at least a third.
- Cost: the monthly tool cost stays below the value of the hours saved, with margin for the editing time you did not expect.
Write these down and share them with whoever will be using the tool. It prevents the quiet drift where a tool that saves nothing is kept because switching it off feels like admitting a mistake.
Include the editing time honestly
The most common accounting error in automation is pretending the output is finished. It rarely is. If a staff member spends eight minutes reviewing and correcting each AI draft, that time is part of the cost. Count it. A tool that saves twenty minutes of writing but costs eight minutes of editing has saved twelve, not twenty — and that difference decides whether the numbers work at your volumes.
Step 3: Run it for one week against the baseline
Keep the manual process available in parallel. Do not switch off the old way on day one. Let the automation handle the task where it can, and let people fall back to the manual route whenever they are unsure. Log both.
At the end of the week, compare four numbers: instances handled, minutes per instance, rework rate, and total tool cost. That is the whole test. It does not need a consultant or a dashboard.
Watch for the three failure signals
Certain patterns show up again and again when an automation is not going to stick:
- People route around it. Staff quietly go back to the old method because the new one is slower in practice, even if it looks faster on paper.
- Rework creeps up. The first week looks fine, then the corrections pile up as the tool meets cases it was not set up for.
- Nobody owns it. When something goes wrong, no one is responsible for fixing the prompt, the rules, or the data feeding it. Automations without an owner decay within weeks.
If you see two of these in the first fortnight, the tool is not the problem — the fit is. Stop, and pick a narrower task.
Step 4: Decide, and write the decision down
At the end of the test, make one of three calls: keep, adjust, or stop. Write a short note recording which one, and why, with the numbers attached. This sounds bureaucratic for a small team, but it does two useful things. It stops the same failed experiment being repeated in six months, and it gives you a defensible answer when someone asks why you are or are not using AI.
If the answer is "keep", set a review date. Automations drift as your products, prices and customer questions change. A quarterly fifteen-minute check on the same four numbers is usually enough.
Where this fits with your website
Most of these tasks start on your website or your messaging channels — an enquiry form, a WhatsApp button, a product page. That means the test is easier to run if the underlying site is set up to capture enquiries cleanly in the first place. If you are unsure whether your current setup is giving you the data you need to measure any of this, it is worth a short conversation before you spend on tools. You can see how we scope that kind of work on our pricing page, or get in touch and describe the task you want to test.
Summary
Whether an automation pays off is not a question about the technology. It is a question about fit, and fit can be measured. Pick one repeating task with a clear finish line, record the manual baseline for a week, set a pass mark in advance covering quality, time and cost, then run the automation alongside the old method and compare. Count editing time honestly, watch for staff routing around the tool, and write down the keep-or-stop decision with its numbers. A week of disciplined measurement will tell you more than any demo — and it costs far less than a year of paying for something nobody uses.