Two vendors can look at the same problem and propose opposite solutions. One offers a system that scores every customer's likelihood of cancelling. The other offers a tool that writes personalised win-back emails at scale. Both get called AI. One estimates, the other produces, and swapping them is one of the quietest ways a promising pilot dies.
The distinction decides what data you need, how you measure success, what a wrong answer costs, and whether your team ships in six weeks or six months.
The shortest possible distinction
Predictive systems answer a question with a number, a label, or a ranking. Will this invoice be paid late? Which of these 400 leads should a rep call first? Is this transaction unusual for this account? The output is small and checkable.
Generative systems produce something a person would otherwise write, draw, or assemble: a summary, a draft reply, a piece of code, a translated paragraph, a structured list pulled out of a messy transcript. The output is larger, more open-ended, and usually reviewed by a human before it goes anywhere.
One line to remember: predictive AI chooses, generative AI creates. Neither is more advanced. They simply have different job descriptions.
What predictive AI is genuinely good at
These models learn from records where the outcome is already known. Thousands of past customers, each labelled as stayed or left. Thousands of past invoices, each marked paid or overdue. If the future resembles the past in the ways that matter, the model can estimate what happens next.
Typical uses: churn scoring, demand forecasting, fraud flags, predictive maintenance, credit risk, ranking search results, routing support tickets, spotting anomalies in logs.
Three properties make this work manageable. Accuracy can be measured against real outcomes. Running costs are usually low and predictable. And answers are stable — a churn model that performed well last quarter does not suddenly change its tone.
The precondition nobody mentions
You need two things before a predictive model is worth building: historical data with trustworthy outcomes, and a decision that actually changes based on the score. If nobody alters what they do when the number moves, you have built a dashboard, not a system.
What generative AI is genuinely good at
Generative tools work on unstructured input, which is most of what a business actually has. Contracts, call transcripts, tickets, emails, policy documents, incident notes. They need far less labelled data, and they respond to plain instructions, so a prototype can exist in an afternoon.
Typical uses: summarising long documents, drafting first versions, extracting fields from messy text, answering questions against an internal knowledge base, translating, writing boilerplate code, turning a meeting transcript into tasks with owners.
The trade-offs are equally real. Output varies between runs, so the same prompt can produce two different paragraphs. Errors arrive in fluent, confident prose, which makes them easy to miss. And quality is judged by people, so your evaluation is slower and more subjective than accuracy on a spreadsheet.
Where generative tools quietly fail
Anything needing an exact figure, a guaranteed calculation, or identical wording across thousands of requests. Do not ask a language model to total a column of numbers or apply a rate table. Retrieve the value or compute it in ordinary code, then let the model phrase it. Arithmetic is not a language problem.
Three questions that usually settle the choice
- Is the output a decision or a draft? A yes/no, a score, or a ranked list points to predictive. Something a person would edit before sending points to generative.
- Do you have history with outcomes? Years of labelled results make predictive viable. If your knowledge sits mostly in documents and conversations, generative gets you further, faster.
- What does a wrong answer cost? If mistakes are expensive and must be traceable, lean predictive, or keep generative on language while facts come from a controlled source.
If two answers both seem right, you are probably looking at a workflow rather than a model. Predict who is at risk of leaving, then generate the draft outreach. That is a chain, and chains are fine.
Four mismatches that waste real budgets
- Generative tools doing arithmetic. A model asked to compute a total will produce a plausible number. Plausible is not correct. Compute separately.
- Predictive models doing prose. A model trained on labels has no concept of a good paragraph. You get classification, not composition.
- Expecting repeatable generative output. If every customer must receive identical wording for a compliance reason, that is a template or a rule, not a prompt.
- Treating a prediction as a decision. A churn score is not a retention plan. Someone still chooses the offer, the timing, and the spend.
A predictive model tells you what is likely. A generative model makes something new. Most expensive mistakes come from asking one to do the other's job.
When the answer is both
Useful systems are usually chains. A support queue: a predictive model classifies the ticket and routes it, a generative model drafts a reply, an agent edits and sends, and a second predictive model watches for escalation risk. A finance team: forecasting produces the numbers, generative writing turns them into commentary the board will actually read.
Keep the parts separate in your architecture. Score in one place, write in another, and log them independently. Then you can measure the prediction against reality and the writing against your team's standards, without one blurring the other.
Before your next pilot, write three sentences
Naming the tool is not a plan. Write down the decision, the output, and the test.
Every Monday we decide which fifty accounts get a call. The system returns a ranked list with one line of reasoning per account. We will know it works if the call team reaches more of the right accounts in the same number of hours.
That is a pilot you can run in two weeks with real users. Start narrow: one prediction feeding one decision, or one bounded writing task with a human approving everything before it leaves the building. Expand only after the measurement holds.
If the decision touches credit, hiring, housing, health, or legal matters, get qualified professional advice before automating any part of it. Model quality does not settle questions of fairness or compliance.
Choose the tool from the decision backwards, not from the demo forwards. The technology is rarely the hard part. Knowing what you are asking it to do is.
Photo: Pexels / Pixabay
