When I was leading acquisition and retention teams at an agency, I had ideas about how we could bring in clients, onboard them and help them stay. The CEO heard me out, but the company was going in a different direction. It was a good place to learn. I also realised I wanted more freedom to try things.
Later, I worked inside a company to understand how decisions affected the product as well as the marketing around it. That connection still matters to me. An advertisement can bring someone in, but what happens next involves the rest of the business.
It is also where I would start with AI business automation: choose a problem you can observe, understand why it happens, and test whether AI helps you solve it. Buying a tool is only one part of that work.
Find the delay before choosing the tool
Imagine a business receiving enquiries that take two days to answer. An AI assistant could draft replies quickly. But suppose the delay happens because nobody knows who should reply, or because every quotation needs the owner’s approval. Faster writing may leave most of the waiting untouched.
This is an illustrative example, but you can check for the same problem in your own business. Follow five recent enquiries from arrival to the next meaningful step. Record who handled each one, what information was missing, and where it waited. Five cases will not establish a reliable benchmark. They can show you where to investigate.
Separate the time spent doing the work from the time spent waiting. A ten-minute task can sit untouched for two days. Those two problems need different changes.
When does AI actually improve efficiency?
There is evidence that AI assistance can help. A study of 5,172 customer-support agents found an average 15% increase in issues resolved per hour after an AI assistant was introduced. The largest benefits went to less experienced and lower-skilled workers. This was one company’s support operation, so it is evidence of a useful application, not a promise for every business. Read the published study, Generative AI at Work.
A reasonable first candidate is a repeated task with clear inputs, an outcome you can check and mistakes you can catch before they reach a customer. Drafting a reply from approved service information might fit. Making an exception to a contract or promising a delivery date the business cannot meet needs a different level of control.
Some tasks are better served by a simpler change. A shared owner, a shorter form or a standard reply may solve the problem without adding another system to maintain.
Run one small test with a clear owner
Write a short test plan before connecting anything:
- The problem: describe what is happening, with a recent example.
- The proposed change: say exactly what the AI will do and what a person will check.
- The outcome: choose a measure such as time to a useful reply, correction rate or successful customer onboarding.
- The owner: name the person who reviews exceptions and can stop the test.
Use information the business is authorised to process, and begin with the minimum access required. A drafting assistant does not need permission to send messages or change customer records just to prove that its drafts are useful.
Include the time spent checking and correcting the output. If drafting saves twenty minutes but reviewing adds twenty-five, you have learned something worth knowing. Faster generation is not enough to establish a saving.
Check what happens to the next team
In Alchemy Avenue’s SmartHeart case study, a change to the buying journey helped more customers complete their purchases. The increased demand then put pressure on the teams delivering the service. This was a customer-journey improvement, not an AI experiment.
It illustrates why an automation test should include the next step. If more enquiries become customers, can onboarding handle them? If replies become faster, are they accurate enough that customers do not have to contact you again?
I wanted to understand these connections when I moved closer to the product. They are still the questions worth asking when the new tool happens to use AI.
