AI predictive analytics is the phrase that closes a demo the fastest, and the one that gets explained the least afterward. It sounds like something serious: a model that learns from your business and tells you what's going to happen. When you ask what's actually inside, the answer is almost always much simpler than that, and that's fine. The problem isn't that it's simple, it's that you get billed as if it weren't.
What you're actually being sold when someone says "predictive AI"
Two very different things live under that same label, and almost no provider tells you which one you're getting:
- A pace-based projection. You take a value that changes over time (hours consumed, stock, visits), look at how much it changed over a recent period, and extend that line forward. It's a rule of three with a nice interface on top. Nothing needs training, and it works well when behavior is reasonably stable.
- A model that actually learns patterns. It detects seasonality, weighs data differently depending on how old it is, combines several variables at once (not just one), and shifts behavior when the pattern shifts, without anyone reprogramming the rule by hand. Here there's real training, labeled historical data, and an actual validation of how accurate it is.
Neither one is better in the abstract. The pace-based projection is cheaper, faster to build and, above all, easier to explain to whoever is using it. The model that learns patterns costs more, takes longer, and only pays off when the problem genuinely needs it. The problem shows up when you're sold the first one at the price and pitch of the second.
The real case: when will an hour bank run out
At AutoBoost we manage the hour banks of the projects we run with our clients through our own system: every client has a dashboard where they can see how many hours are left in their active bank. Until recently, that dashboard stopped there: a number going down. Useful, but incomplete, because a number that's left forces the client to do the math themselves to know if it will last until next month.
We added a depletion forecast: the date at which, at the current pace, that bank will hit zero. The calculation runs on the consumption pace of the last 4 weeks, not the entire history since day one, because the recent pace predicts what happens next week better than an average from half a year ago that no longer reflects how the service is being used now.
Here's the uncomfortable part to admit, and we admit it the same way: this isn't a model that learns. It's hours remaining divided by recent weekly pace, the same old rule of three, well chosen in the time window it uses and placed where the client sees it before it's too late. We didn't build any AI to make it, and we didn't need to.
What did genuinely change was its usefulness: a number that's left ("you have 40 hours left") doesn't move anyone to act. A date ("at the current pace, it runs out on November 14th") does, because it turns a passive figure into a decision with a deadline. That's the lesson that carries over to any business that sells by bank, subscription or usage, whether AI is involved or not: showing the pace, not just the balance, moves the renewal conversation up to a point where there's still room to have it calmly.
When you actually need a model that learns, not a rule of three
A rule of three breaks down the moment the behavior you want to predict depends on more than one thing at once, or changes in a way a straight line doesn't explain. That's where you genuinely need a model trained on real data, not a projection.
That's the difference with what we do, for example, in the data and AI platform that unifies data for a pharmacy group: there, the AI doesn't project a line, it reasons over more than a million sales line items to reconstruct what was actually sold when the ERP itself was calculating it wrong, cross-checking several sources at once and validating the result down to the cent. No rule of three solves that, because there isn't a single variable changing over time: there are multiple sources that need to be reconciled against each other.
| Question | If yes, a rule of three probably covers it | If yes, you probably need a real model |
|---|---|---|
| Does it depend on a single variable changing over time? | Yes (hours, stock, visits) | No, it depends on several at once |
| Does recent behavior predict the next period well? | Yes, it's reasonably stable | No, it shifts with seasonality or external factors |
| Do you need to cross-check or reconcile several distinct data sources? | No, one source is enough | Yes, several sources need combining and validating |
| Is the error from a simple projection cheap to correct? | Yes, it adjusts by hand without drama | No, an error propagates into an expensive decision |
Before you pay for "predictive AI," ask this
- Does the prediction use more than one variable at once, or just the recent pace of one?
- Is there an actual training process and a validation of how accurate it is, with concrete figures?
- Does it shift behavior when the pattern shifts, or does it just extend the last line?
- Can they explain in one sentence what the number they're giving you is fed by?
- Does the price match building and maintaining a model, or a formula that gets written in an afternoon?
If most answers point to the simple column, there's nothing wrong with sticking to a rule of three. Just make sure the price you're paying matches a rule of three too.
Why we tell it this way, even when it would be easier to just sell "AI"
At AutoBoost we build custom software and AI integrated into each client's real system, and part of that work is deciding, for every single piece, whether it needs a model that learns or whether a well-designed calculation solves the same problem with less cost and less risk of failure. Adding AI where it isn't needed doesn't impress anyone who uses it for more than two weeks: it just adds a black box where a verifiable calculation used to be. Order the data and pick the right tool first, then apply AI where it genuinely adds value, not before.
If you're evaluating a project being sold to you as "predictive AI" and want to know which of the two things it actually is, let's talk at AutoBoost.