What machine learning can and can't do
A machine learning model handles tasks well when there are many similar examples from the past and a clear correct outcome. For example, thousands of orders with a known result (bought / didn't buy), or several years of sales.
At the same time, a model can't predict a future that never appeared in the data. It doesn't know about a new competitor, a change in regulations, or a product that has never been sold. That's why a forecast is guidance with a certain level of accuracy, not a final answer that needs no checking.
Typical business use cases
- forecasting sales, demand, and warehouse or call center workload;
- probability scoring: a lead will convert, a customer will churn to a competitor, an order will be returned;
- classification: automatically routing requests, products, or documents into categories;
- recommendations for products, services, or content on your website and in email campaigns;
- segmentation of customers by behavior;
- anomaly detection in sales, payments, or technical metrics.
What it looks like in practice
Take a hypothetical online store with several thousand products. Purchasing is currently planned like this: a manager looks at last year's sales and adds some extra stock just in case. As a result, some products sit in the warehouse while others run out during peak season.
A demand forecasting model learns from sales history, prices, promotions, seasonality, and stock levels. Every week it calculates expected sales for each product, and the manager sees the forecast in a familiar spreadsheet next to current inventory. The purchasing decision stays with a person, but now it's based on a calculation, not just memory.
When you don't need an ML model
We'll tell you honestly if the problem can be solved more simply. Machine learning is unnecessary when:
- the rules are obvious and stable: "wholesale customers get a discount" doesn't need a model;
- there's too little data: a few hundred records usually aren't enough for a model to learn anything useful;
- the output has nowhere to go: a forecast nobody sees in their daily work won't change a single decision;
- you need to explain causes rather than predict outcomes: solid data analysis is often enough here.
How much data, and what kind
There's no exact number, but there are some guidelines. A seasonal forecast needs history covering at least several full seasons. Lead scoring needs not only the leads themselves but also the outcome: whether the customer bought or not. Without that, the model has nothing to compare against.
Quality matters too: if sales reps don't always close deals in the CRM or statuses are set at random, the model will learn from those mistakes. That's why we always audit the data before training and tell you what you can realistically expect from it.
External data is often useful as well: weather, holidays, exchange rates, ad campaign schedules. It helps the model explain spikes that aren't visible in sales history alone.
How we verify that a model works
We evaluate the model on data it didn't see during training. For forecasts, we use the most recent period, as if we had made the forecast in advance, and then compare it with what actually happened.
The key comparison is against the simple method you use today: "same as last year," a recent average, or a manager's experience. If the model isn't noticeably better than that, we'll say so, and it isn't worth deploying.
How the model gets into production
A model that lives only on an analyst's laptop delivers no value. Depending on the task, we:
- write forecasts to a spreadsheet or data warehouse where they show up in reports;
- add a lead score to the CRM so reps see it on the customer record;
- connect recommendations to your website or app via API;
- set up scheduled forecast updates and anomaly alerts.
If a language model is a better fit than a classic model, for example for processing the text of support requests, that falls under AI integration for products.
Benefits and risks
Benefits: decisions are made faster and more consistently, sales reps focus on the most promising customers, purchasing is planned on more than gut feeling, and repetitive work is partly automated.
Risks: models make mistakes, especially in unusual situations; their accuracy degrades over time as the market changes; a complex model is hard to explain, and people may not trust it. That's why we build in accuracy monitoring and, where possible, choose models whose decisions can be explained.
What to prepare on your side
- a description of the decision you want to improve and the people who make it today;
- access to historical data: sales, leads, customers, inventory, for the longest period available;
- context on anything unusual: when prices, product range, or business rules changed, and any gaps in record-keeping;
- a person on your team who can help check whether the model's results make sense.
What affects the cost
- Data readiness: a ready-to-use dataset, or data that needs to be collected and cleaned from several systems.
- Task complexity: a single forecast of total sales, or forecasts for thousands of products across several regions.
- Integration: a report with results, or integration into your CRM, website, or inventory system.
- Update frequency of forecasts and response-time requirements.
- Support: accuracy monitoring and retraining after launch.
We often suggest starting with a pilot on your data. It shows whether the required accuracy is achievable before you invest in a full rollout.
Frequently asked questions
What accuracy do you promise? None in particular until we've reviewed the data. Accuracy depends on whether patterns actually exist in your history. After the pilot, you see real numbers on your own data.
Can we use off-the-shelf services instead of a custom model? Sometimes, yes, and it's cheaper. For example, some ecommerce platforms already have built-in recommendations. A custom model makes sense when ready-made solutions don't account for the specifics of your business.
What do you build models with? Mostly Python and SQL, with tools chosen to fit the task and your infrastructure.
Who will own the model? You will. The code, the trained model, and the documentation are handed over to you.
Why Promodex
Since 2015, Promodex has delivered 750+ projects in development, design, and marketing. That matters for machine learning for a practical reason: a model only delivers value when it's built into your website, CRM, or reports, and our in-house team builds and supports those systems. We start by asking whether you need a model at all, compare it against simple methods, and never push you to deploy something that doesn't make a real difference.
