Most small businesses use AI to write. Far fewer use it to forecast. Yet your orders, work orders and purchase invoices already hold answers: when a machine is likely to fail, how much you will sell in March, when to buy to pay the right price. A predictive model is the tool that finds those answers.
The use of AI nobody talks about
Among Canadian businesses that use AI, data analytics ranks first (36.6%), ahead of text analytics and chatbots, according to Statistics Canada. It gets less attention because it cannot be shown off in ten seconds like a chatbot. Yet that is often where the dollar gains are.
What a predictive model does for you
A predictive model learns from your history to estimate what will probably happen: a quantity, a price, a date, a risk. It does not chat and it does not write. It is a different family of AI from ChatGPT, which we explain in our article on automation, generative AI and agents.
- It gives an estimate with its margin of error. A good model also says how sure it is.
- It has to beat your current method. The average of the last three months, or the planner’s gut feeling. If it does not do better, we keep your method, and we tell you.
- It is worth the decision it changes. A forecast sitting in a report earns nothing.
Five useful forecasts for a small business
| We predict | With | What it changes |
|---|---|---|
| Demand | Order history, over at least two seasonal cycles | You buy and produce at the right time, with less stock sitting idle. |
| Breakdowns | Work orders, running hours, sensors if there are any | The part gets changed during a planned stop, not in the middle of an order. |
| Purchase prices | Supplier invoices, dates, quantities | You know when to buy and what price is reasonable. |
| Supplier delays | Promised dates and actual receiving dates | Planning adjusts before the delay hits. |
| The real time a job takes | Estimated and actual hours on past jobs | More accurate quotes and schedules. |
- Demand. How many parts or products your clients will order next month, from order history. The buyer orders at the right time, and idle stock shrinks.
- Breakdowns. Which machine is likely to fail, from work orders and running hours. Maintenance happens during a planned stop, not in the middle of a rush order.
- Purchase prices. When a material or service is cheapest, from your invoices. You know when to buy, and whether a quoted price is reasonable.
- Supplier delays. Which supplier is likely to deliver late, by comparing promised and actual dates. Planning adjusts before the delay hits.
- The real time a job takes. How many hours an order will really take, from past estimated and actual hours. Your quotes and promised dates get more accurate.
For maintenance, the order of magnitude has been known for a long time. According to the U.S. Department of Energy’s best practices guide, a predictive maintenance program saves 8% to 12% over preventive maintenance alone, and sometimes more than 30% when a business mostly runs in repair mode.
Prediction models found where the savings hide, in 8 million airfares
It is one of our most advanced projects, done for the Pulse Incubateur HES in Geneva (HES-SO). The question: what really drives the price of an airline ticket, and when should you book to pay the right price?
It started with 8 million airfares. After cleaning, 163,000 reliable tickets were kept. Prediction models built in Python then measured the effect of each factor. The result: up to $151 difference per ticket, and a window of 22 to 45 days before departure when the price is fairest.
What made the difference was the deliverable: an app that shows the fair price before booking. For a large company buying 10,000 tickets a year, the savings are estimated at up to $1.5M a year. That is an estimate, not a result measured at a client. The full story of the project.
What it looks like in your week
A predictive model does not live in a report opened once a month. It shows up where the decision is made. On Monday morning, the buyer sees the suggested quantity for each material on the ordering screen, with its margin of error. They still have the last word, but they no longer start from scratch.
The maintenance lead gets the three machines to watch this week, and why. The estimator sees, next to their estimate, how long similar jobs really took. Nobody learned new software: the forecast moved into the tools they already opened.
And each time reality contradicts the forecast, the model takes it into account at the next training. It gets better the more your business uses it.
What you need before you start
- A consistent history. Not necessarily huge, but with the same codes and units from one year to the next. Why ERP data quality matters so much.
- A baseline number. How you forecast today, and how wrong you are. That is the number the model has to beat, as we explain in the real test of a successful AI project.
- A person who decides with it. The buyer, the planner or the maintenance lead. The model is built with them.
This is the kind of project we build in our Intelligent Production line: models built on your data, delivered in a tool your team actually uses, connected to the ERP when possible.
Where to start
Think of a decision you make every week with a bit of instinct: how much to order, when to service a machine, what lead time to promise. If you have two or three years of history on that decision, there is probably a model to build. We look at your data first, tell you honestly whether it is worth it, and then we build.
Frequently asked questions
How much data does a predictive model need?
It depends on what you forecast. For seasonal demand, ideally at least two full cycles. For breakdowns, enough past events for the model to see the warning signs.
Is a predictive model the same thing as ChatGPT?
No. ChatGPT produces text. A predictive model learns from your numbers to estimate a quantity, a price, a date or a risk. The two can work together: generative AI can explain in words what the model forecasts.
Do we need sensors for predictive maintenance?
Not always. Work orders and running hours are often enough for a first model. Sensors come later, on the machines where the stakes justify them.
What do we get at the end of the project?
A tool, not a report: an internal app, a dashboard or a forecast that appears directly in the ERP, with its margin of error.
What happens when the model is wrong?
It will sometimes be wrong, like any forecast. We measure its error continuously, compare it with your previous method and retrain it when conditions change.
This work falls under our intelligent production service: applications, agents and predictive models built for your operations.
A decision you make on instinct?
Tell us about it. We check whether your data can forecast it, and tell you honestly if it is worth it.
What this article argues comes from a project carried out for Pulse Incubateur HES. The case study walks through what was built, how, and what a company can take from it.
