Guide · Predictive Analytics
Predictive Analytics: A Business Buyer's Guide
Predictive analytics uses your historical data to forecast what is likely to happen next, demand, failures, churn, so you can decide with foresight rather than hindsight. This guide is for the person deciding whether to buy it: what predictive analytics can genuinely do, what data it needs, how to evaluate a supplier or a platform, what it realistically costs, and how to avoid the most common outcome of all, a model that works technically and gets used by nobody.
Key takeaways
- Buy the decision, not the model. If a forecast would not change what someone does, it is not worth paying for.
- Your data almost certainly matters more than the algorithm. Ask any supplier to assess it honestly before you sign anything.
- Accuracy is not the goal. A slightly less accurate forecast that reaches the decision-maker in time beats a better one that arrives too late.
- The most common failure is not a bad model. It is a good model that lands in a dashboard nobody opens.
Table of Contents
Realistic capability
What predictive analytics can genuinely do
Predictive analytics works by finding patterns in what has already happened and projecting them forward. That sounds modest, and it is: there is nothing mystical about it. But applied to the right decision, it is genuinely valuable, because most businesses currently make forward-looking decisions using backward-looking reports and a lot of gut feel.
The uses that reliably repay the investment share a shape. They involve a decision made repeatedly, where being roughly right in advance is worth more than being exactly right afterwards, and where enough history exists for a pattern to be real rather than imagined.
Demand forecasting
How much you will need, and when. Usually the highest-return use, because getting demand wrong costs twice: in lost sales and in tied-up capital.
Predictive maintenance
Which equipment is likely to fail before it does, turning unplanned downtime into a scheduled task.
Churn prediction
Which customers are likely to leave, surfaced while there is still time to do something about it.
Revenue and cash forecasting
What is likely to come in and when, based on real patterns rather than a spreadsheet extrapolation.
Risk signals
Problems flagged while they are still cheap to fix, rather than discovered in a post-mortem.
Capacity and staffing
Planning people and resources around what is coming rather than what happened this time last year.
The part suppliers skip
What it cannot do, honestly
Predictive analytics projects patterns forward. That means it is structurally poor at predicting things that have no precedent in your data. A model trained on three years of steady trading will not anticipate a shock it has never seen, and any supplier implying otherwise is overselling.
It also cannot fix a decision nobody is empowered to make. If your team already knows demand will spike and still cannot act, because approvals take three weeks or the supplier needs six weeks’ notice, a better forecast changes nothing. That is a process problem wearing a data problem’s clothing, and it is worth identifying before you commission a model rather than after.
And it cannot manufacture signal that is not in the data. If the thing you want to predict genuinely depends on factors you have never recorded, no algorithm recovers that. A good supplier will tell you this at the scoping stage. A poor one will build you something anyway and let the accuracy disappoint you later.
The real prerequisite
What data you actually need
The honest headline: your data matters more than the algorithm. Modern modelling techniques are widely available and largely commoditised. What differentiates a forecast that works from one that does not is almost always the quality, consistency and relevance of the history behind it.
What you generally need is enough history to contain the pattern you care about. For anything seasonal, that means multiple cycles, one year of sales data cannot tell a model what December looks like relative to July with any confidence. You need it recorded reasonably consistently, because a change in how you categorised products two years ago can quietly break a model. And you need it to actually contain the thing being predicted: to forecast equipment failures, you need a record of past failures, not just maintenance visits.
What you do not need is perfection. Working with real, messy business data is the job, not an obstacle to it. Any supplier who requires pristine data before starting either has not worked with real businesses or is managing their own risk at your expense.
Comparing options
How to evaluate a supplier or platform
Broadly there are three routes, and the right one depends far more on your situation than on which is objectively best.
| Off-the-shelf platform | Commissioned build | In-house team | |
|---|---|---|---|
| Best when | Your need is common and well-defined: standard demand forecasting, standard churn scoring | Your decision is specific to how your business works, or the data is awkward | Predictive work will be continuous and central, not a one-off |
| Speed | Fastest, often weeks | Moderate, scoped per project | Slowest by far from a standing start |
| Fit to your process | You adapt to the tool | Built around your decision | Fully tailored, over time |
| Cost shape | Ongoing subscription | Fixed project cost plus running cost | Salaries, plus a long ramp before value |
| Main risk | Generic forecast that does not fit your reality | Choosing a supplier who builds the model but not the adoption | Hiring specialists with nobody experienced to direct them |
Most businesses buying predictive analytics for the first time are better served by a platform or a commissioned build than by hiring, because a single data scientist without experienced direction is an expensive way to discover what you needed.
Due diligence
Questions to ask before you sign
These separate suppliers who will deliver value from suppliers who will deliver a model.
| What a weak answer sounds like | What a good answer sounds like | |
|---|---|---|
| Which decision will this change? | "It will give you visibility into demand" | A named decision, a named person, and when they make it |
| Is our data good enough? | "We will work with whatever you have" | An honest assessment first, including "not yet" if that is true |
| How will the forecast reach the decision-maker? | "It will be available in the dashboard" | A specific answer about where it appears in their actual workflow |
| How accurate will it be? | A confident percentage before seeing your data | "We cannot know until we look, and here is how we will measure it" |
| What does it cost to run, not just build? | Only the build price is quoted | Build cost and ongoing running cost, estimated up front |
| What if it does not work? | Avoidance, or blaming data quality in advance | A clear success measure agreed before starting |
Budget reality
What it costs, and where budgets go wrong
Cost varies enormously with scope, and anyone quoting a figure before understanding your decision and your data is guessing. What is consistent is the shape: a focused, single-decision forecasting project is a contained, scopeable piece of work. A broader system feeding multiple decisions across an operation is a project, priced accordingly.
Two things reliably blow budgets. The first is scope creep from starting too broad, “let’s forecast everything” has no natural end point, whereas “let’s forecast demand for these forty product lines” does. The second is the ongoing running cost, which is easy to overlook at purchase: models need maintaining, data pipelines need running, and where AI models are involved they charge per use. Ask for that figure before you commit, not after.
The failure mode
Why most predictive projects quietly fail
They do not usually fail loudly. There is no moment where the model breaks and everyone notices. What happens instead is that the project completes, the model is technically sound, the output goes into a dashboard, and then nothing changes. Six months later nobody is opening it, the decisions are still being made on gut feel, and the conclusion drawn is that predictive analytics did not work for this business.
The cause is almost always the same: the project started with the data rather than the decision. Nobody established, before building, which decision the forecast was meant to improve, who makes it, when, and what would have to reach them for it to change what they do. Without that, the model has nowhere to land.
The single most useful thing you can do as a buyer is refuse to commission anything until that question has a specific answer. Not “it will improve our planning” but “it will tell Sarah, every Monday morning, how much of each line to order, in the system she already uses.” That level of specificity is the difference between a model that changes the business and one that decorates it.
Frequently asked
Buyer's questions about predictive analytics
Predictive analytics uses your historical data to forecast what is likely to happen next, such as future demand, likely equipment failures, or customers likely to leave. Unlike standard reporting, which describes what already happened, its purpose is foresight: seeing what is coming in time to act on it.
You need enough history to contain the pattern you care about, recorded reasonably consistently, and it must actually include the thing you want to predict. For anything seasonal that means multiple cycles. It does not need to be perfect. Any credible supplier will assess this honestly before you commit, and tell you if the answer is "not yet".
A platform suits common, well-defined needs and is fastest. A commissioned build suits decisions specific to how your business works. Hiring in-house makes sense once predictive work will be continuous and central. Most first-time buyers are better served by the first two, because a lone data scientist without experienced direction is a costly way to learn what you needed.
Be wary of anyone quoting a confident accuracy figure before seeing your data. More useful than raw accuracy is whether the forecast is good enough to beat the current decision-making method and arrives in time to be acted on. A slightly less accurate forecast delivered at the right moment beats a better one delivered too late.
Predictive analytics produces the forecast. Decision intelligence turns that forecast into a specific recommendation, delivered where and when the decision is actually made. The forecast is the input; the recommendation is where the value is realised.
Commissioning a model before establishing which decision it is meant to change. That single omission causes most quiet failures: the model works, the output goes into a dashboard, and nothing about how the business operates actually changes. Insist on a specific answer, naming the decision, the person and the moment, before signing anything.
Work out whether it is worth buying at all
Book a free consultation. Tell us the decision you would want a forecast to improve. We will tell you honestly whether your data can support it, what it would take, and whether it is worth doing, before you spend anything.