AI & Automation
What this article covers
Every small team has now been told that AI will save it time. Most of them have tried, and most of those attempts produced a tool nobody opens, a chatbot that annoys customers, or a workflow that breaks quietly and is discovered three weeks later.
The failure is rarely the technology. It is the choice of task.
The test: frequent, rule-shaped, and cheap to be wrong about
A task is worth automating when all three of these are true.
Frequent. Something done twenty times a week repays the setup. Something done twice a year never will, no matter how tedious it is each time.
Rule-shaped. You can describe how the decision is made without saying "it depends on the situation" more than once. If the rule lives only in one person's head and changes with context, writing it down is the actual project — and it is worth doing whether or not you automate afterwards.
Cheap to be wrong about. Ask what happens when the automation makes a mistake at two in the morning. If the answer is "a draft looks odd and someone fixes it", proceed. If the answer is "a customer is charged twice" or "a contract goes out unreviewed", the automation needs a human gate, and that gate usually eats the savings.
What usually passes the test
- First-draft anything. Product descriptions, meeting notes, ad variations, FAQ answers. The draft is fast and the review is quick, which is exactly the right shape.
- Sorting and routing. Reading an inbox and deciding which message is a sales enquiry, which is support, and which is a supplier. Low stakes, high frequency, genuinely tedious.
- Extracting fields from messy input. Pulling a name, a company and a request out of a paragraph someone typed into a form at midnight.
- Answering the same twenty questions. Delivery times, opening hours, what is included, whether you work with a certain kind of client. Worth doing only if the answers are already published somewhere — see the warning below.
- Translating and adapting copy. Especially between English and Arabic, where the useful output is a good first pass a bilingual person then tightens.
What usually fails it
- Anything that quotes a price or commits you. A confident wrong number is worse than a slow right one.
- The first contact with an unhappy customer. People can tell, and being handled by a machine while annoyed makes them more annoyed.
- Strategy and judgement. A model will happily produce a plan. It has no idea which of your constraints are real.
- Work you do rarely. The setup cost is the same; the payback is not.
The mistake almost everybody makes first
Automating a process that was already broken. If your intake form asks the wrong questions, automating the follow-up means you now send the wrong follow-up faster, to more people, without anyone reading it. Automation removes the friction that used to make bad processes visible.
So before you wire anything up, do the boring version by hand for two weeks and write down what you actually did. That document is both the specification and the proof that the task is worth the effort.
Where the answers should come from
If you put an assistant in front of customers, it must answer from your published material — your service pages, your prices where you publish them, your policies — and it must be allowed to say "I don't know, here is a human". An assistant that invents a delivery time to be helpful creates an obligation you did not agree to, and a support ticket you now have to apologise for.
This is the single design decision that separates an assistant customers trust from one they learn to skip. Everything else is plumbing.
Start with one task, measured
Pick the single most frequent, most rule-shaped, least dangerous task on your list. Time it by hand this week. Automate it next week. Time it again.
If the saving is not obvious in the second measurement, the task was wrong — not the idea. Move to the next one on the list.
If you want a second opinion on which of your processes are actually worth automating, that conversation is what our AI solutions and automation work starts with.
