Every software company says it uses AI now. That sentence tells you nothing. What tells you something is a real feature in a real product: what the AI is allowed to do, what it isn’t, and who is responsible when it gets something wrong. So here is one, from a platform we run, and the decisions we deliberately keep away from it.
The platform is Lasso Media, a call tracking and lead generation business we have been the development team for since December 2016.
A Sign-Up Form That Fills Itself In
When a business signs up with Lasso, one step asks them to choose the services they offer from a fixed list for their trade, at least six of them. It is a tedious step. But by then the business has already given us their website, and their website already says what they do. So we read it, and the step opens with those services already ticked.
Why Simple Matching Wasn’t Enough
We tried plain word matching first, and it failed in three instructive ways:
- The words don’t match, but the meaning does. A website says “AI automation”; the list calls it “Artificial Intelligence Solutions”. No shared words, same service.
- The words match, but the meaning doesn’t. A business that only does data analysis was offered Data Backup and Data Security, on the strength of the word “data”. An electrician was offered Generator Installation because their site mentioned “installation”.
- One phrase is really two services. “Web Application Development & Maintenance” is two rows on the list, and matching one phrase to one row can’t express that.
On one test, word matching found three of the six services the step needs. With an AI model choosing from the same list, on the same input, it found ten.
The Rules Around It
That number is the least interesting part. What makes the feature safe to put in front of customers is what the AI is not allowed to do:
- It can only choose from the list. Anything it returns that isn’t on the list is thrown away before a customer ever sees it. It cannot invent a service.
- It doesn’t pad. A six-service minimum is exactly the kind of pressure that tempts a system to guess. If it is confident about four, it offers four, and the customer adds the rest.
- It never overrides the customer. It only fills an empty step. If the customer has already made their own choices, those stand.
- New services are the customer’s decision. If a business offers something the list doesn’t have, adding it is a deliberate act by a person, never something the AI does on their behalf.
- The customer is told where the suggestions came from, so a pre-ticked box is never mistaken for something they chose.
- It isn’t on the critical path. The matching runs in the background, and if the AI service is unavailable, the old word matching takes over. Sign-up never waits on it.
Before it shipped, it was checked end to end in a browser against a real electrician’s website, and each of those rules has an automated test that we proved by deliberately breaking the rule and watching the test fail.
Where a Person Decides
The same platform records calls, turns them into transcripts, and uses AI to write a short summary of each one. That saves real time. But whether a call counts as a lead is decided by a person, with the transcript in front of them as a reading aid. And what a client is charged depends only on the length of the call, never on what an AI made of the conversation.
That is the rule we work to: AI can prepare a decision. When money or a customer’s trust depends on it, a person makes it.
How We Use It in Our Own Work
We use AI every day: to draft code, write tests, review changes, find our way around large and old codebases, and produce first drafts of documentation. It makes us noticeably faster, and we’re not shy about saying so.
What we don’t hand to it: what we tell you about cost and risk, whether something needs rebuilding or just fixing, and whether a feature is finished. Finished means we ran it and watched it work, not that a model said it looked right.
If You’re Adding AI to Your Own Product
Whoever builds it for you, ask these four questions before anything goes live:
- What is the AI allowed to choose from, and what happens to an answer outside that?
- When it’s wrong, who sees it first: your team or your customer?
- What happens when the AI service is down?
- Which decisions stay with a person?
If the answers are vague, the feature isn’t ready. If you’d like AI inside your own application with those answers settled up front, see what we build or tell us about your project.