Most first AI projects do not fail because the model is not good enough. They fail because the project was too broad to finish, too vague to measure, or too far from the systems where the work actually happens. Here is how we help brands choose a first project that avoids those traps.
Start where the work repeats
Look for work that happens many times a week, follows a recognisable pattern and has a clear right answer most of the time. In e-commerce, the usual candidates are:
- answering order-status, shipping and returns questions;
- checking stock and price before replying to a reseller;
- moving orders and customer records between the storefront, ERP and CRM;
- writing the weekly summary of what changed in sales and advertising.
Rare, high-judgement work is a poor first project. It is hard to test, and mistakes are expensive.
Make sure the data is reachable
An agent is only as good as what it can look up. Before committing, check that the answers live somewhere a system can read: an order database, a product catalogue, a policy document, an API. If the answer only exists in someone's head, the first project is writing it down.
Define "done" before you build
Agree on a few measures you can read from records, not from opinion:
- the share of conversations or tasks handled without a person;
- how long customers wait for a first answer;
- how often a person has to correct the agent.
If you cannot measure it, you cannot tell whether the agent helped.
Keep a person in the loop
Design the hand-off before the automation. When the agent is unsure, or the stakes are high, it should pass the case to a person with the context attached. A good first project makes your team faster; it does not remove their judgement.
Ship small, then widen
Launch on one channel, one product line or one type of question. Watch real conversations for a few weeks, fix what goes wrong, then widen the scope. A narrow agent that works earns the trust a broad one needs.
A simple test
If you can describe the task in one sentence, point to where the data lives, and name the number you expect to move, you have a good first project.
If you would like a second opinion on yours, email [email protected].