By Ani Nandi  |  Founder, EcomChief

A customer wants 20 chairs delivered before Friday’s opening. A retailer’s new AI assistant checks a stock sheet and confirms the order in seconds. There’s just one problem: the warehouse has eight chairs, not 20.

The customer has already started making plans. Now someone has to make an uncomfortable phone call. The AI answered quickly, but it was working with information the business should have checked first.

Look at the work before the software

Small firms have good reasons to be curious about AI. The same enquiries arrive every day, newsletters take longer than they should, and somebody always seems to be compiling the weekly figures. Handing over a few of those jobs sounds sensible.

But spend an hour following one task through the business before choosing a tool. Where does the information come from? Who checks it? What happens when the usual answer doesn’t apply? Ask the person doing the job which parts they quietly fix as they go along. Those improvised repairs are easy to miss until software starts repeating the mistakes.

In ecommerce, even the word “available” needs care. An item can appear on a website while it is being picked for another order, awaiting a stock update or too late for that day’s courier collection. A quicker reply is no use if it promises what the team cannot deliver.

Three jobs worth testing first

Content preparation. Give AI approved product information and let it draft an email, product description or social post. A colleague should still check the price, claims and tone. Producing a draft is a modest risk; publishing unchecked promises is not.

Routine customer questions. Opening hours, booking procedures and standard delivery policies are reasonable starting points when the answers come from a current, agreed source. Complaints, order disputes and unusual requests should go to a person.

Weekly reporting. AI can help turn a sales export into a short summary or flag figures worth investigating. First make sure the totals match the underlying records. A confident explanation of the wrong numbers is worse than a spreadsheet that takes ten more minutes to read.

Five questions to settle before switching it on

1. Which information is correct? Two employees may both have a file called “latest stock list”, but one is using a copy saved last Friday. Decide which record is authoritative and who updates it. If the source is unreliable, the output will be too.
2. Who is responsible? Name the person who approves the workflow, checks its results and fixes problems. A software subscription cannot take responsibility for a disappointed customer.
3. What can the system promise? Put boundaries in writing. May it suggest delivery dates, confirm a booking or offer a refund? At the beginning, let it draft these replies rather than make commitments without a human check.
4. What does the whole task cost? Count fees, setup time, checking and rework. Saving five minutes on the first draft means little if someone spends six minutes correcting it.
5. When does a person step in? Set clear triggers for unclear requests, upset customers, sensitive details and exceptions. The handover should include the conversation so customers do not have to explain everything twice.

Data protection belongs in the same discussion. Before putting customer or employee information into a tool, check the provider’s terms, your organisation’s obligations and how the information is handled. The Information Commissioner’s Office publishes ICO guidance on AI and data protection for UK organisations.

Keep a scorecard that tells the truth

Choose one repetitive task and measure a normal week before the pilot. Then record the same information with AI assistance: how many tasks were completed, average time per task, mistakes or corrections, and total cost including review. Keep a note of exceptions that needed a person.

Suppose 50 customer enquiries take six minutes each. That is five hours. If drafting with AI and checking the replies brings the average down to four minutes, the team saves 100 minutes. Useful, certainly. But it is only a business improvement if accuracy holds up, the software costs make sense and those minutes can be put to better use. These are illustrative figures, not a promised result.

Try it for a month

Week one: map the task, remove outdated information and agree the rules. Week two: test with fictional or appropriately protected examples, including the awkward ones. Week three: run a small, supervised pilot and log the corrections. Week four: compare the results with your original scorecard. Keep it, change it or stop.

The people handling the work should have a say in that decision. They know which exceptions customers actually raise and which shortcuts create problems later.

Tomorrow morning, choose one task your team repeats every week. Sit beside the person who does it and ask what they have to double-check. You might find an opportunity for AI. You might find that fixing a spreadsheet or updating a policy would help just as much.

Either way, you will be solving the problem before buying the solution.

About the author

Ani Nandi is the founder of EcomChief, with more than 20 years of experience in ecommerce and web development. He writes about practical technology and the everyday decisions facing small businesses.

Website: ecomchief.com