Business · Artificial intelligence
Generic AI or AI that knows your business
· 4 min read ·
A customer writes: "When is this arriving?"
A generic tool answers something like "the usual lead time is three to five working days; please confirm with your supplier." It's correct, it's polite, and it's useless — the customer already knew that.
A tool connected to your own systems answers that the order shipped yesterday, that this particular customer has a Saturday delivery arrangement from an old agreement, and that the address on file is the old one. The question was the same. The model might even be the same. What changed was what it could look up.
Where the generic one is genuinely good
It's worth being fair to it, because there's work it's unbeatable at, and cheaply so.
Writing a first draft, summarising a long document, translating, rewriting an email to make it shorter, explaining a technical term to someone outside the field, tidying up meeting notes. None of this requires the tool to know anything about your company — it requires it to know how to write.
If that's where your pain point is, you don't need a project. You need a subscription and half an hour showing the team how to use it.
Where it fails, and it's always the same spot
It fails the moment the right answer depends on a fact that only exists inside your company.
That a particular customer has special terms. That an item's been discontinued since March. That the job stalled waiting for approval. That the factory closes in the second week of August. These are ordinary facts that anyone in the building knows, and that no model trained on the internet could ever guess.
The worst part isn't getting it wrong. It's getting it wrong with the same confidence it gets things right. A tool that answers "three to five working days" with total confidence, when the order has actually been stuck for two weeks, does more damage than not answering at all.
What it takes for it to know your business
Less than you'd think, and different from what you'd think. You don't need to train your own model — that's rare and expensive. You need to give it access to the right information at the moment it answers.
- The information has to be accessible. A system it can query, not a file someone exports on Fridays.
- It has to be up to date. An answer built on yesterday's state is a wrong answer, with good grammar.
- It needs an owner. If the delivery date exists in two places with different values, someone has to decide which one counts. It's the same problem we wrote about in why nobody trusts the numbers.
In practice, this is systems integration work. The artificial intelligence part is usually the shortest bit of the project.
Ten years of data isn't worth more than two
There's an idea floating around that more history is always better. It isn't quite true.
Ten years of records where the criteria changed three times, where half the fields were filled in however each person felt like it, and where two migrations ate the attachments, are worth less than two consistent years. The tool will learn that there are three ways of saying the same thing and treat them as three different things.
If you have to choose between cleaning up old history and making sure the next two years go in correctly, choose the second. It's cheaper and the results show up sooner.
Want answers that actually know your business?
Tell us what questions your customers ask most often, and where the answers live today. We'll see what needs connecting.