Predictive AI vs generative AI
Predictive AI calculates a probability or a number, generative AI writes text. The difference, when to use which and why revenue control needs both.
Predictive AI forecasts an outcome and returns a number: a probability, an amount, a category. Generative AI creates new content, usually text, in response to an instruction. For revenue control this means predictive AI calculates where the risk lies, while generative AI explains what is going on or prepares an action. They solve different problems, and most mistakes happen when you use one for the other's job.
What is the difference between predictive and generative AI?
The difference is not how clever they are, but what they return.
| Predictive AI | Generative AI | |
|---|---|---|
| Output | Number, probability, label | Text, code, summary |
| Learns from | Your historical data, with known outcomes | Vast amounts of general text |
| Typical question | How likely is this customer to leave? | Summarise this contract and state the indexation clause |
| Checked by | Comparing the outcome with what happened | Setting the output against the source |
| Typical failure | Wrong prediction from poor or outdated data | Convincing-sounding text that is wrong |
| Examples | Churn score, deal probability, payment prediction, anomaly detection | Language models such as those behind ChatGPT, Claude or Gemini |
A predictive model is usually trained on your own data. It knows nothing about the world, but a great deal about your customers. A generative language model is trained on general text. It knows a lot about the world, but nothing about your customers, unless you supply that information at the moment you ask it a question.
What does predictive AI do?
Predictive AI is the workhorse behind predictive analytics. It receives structured data, such as a table of customers, orders and invoices, and learns which combinations of attributes are associated with an outcome.
Applications in a B2B revenue context:
- The probability that a deal closes this quarter, based on the deal's own attributes rather than the standard probability per stage.
- The probability that a customer will not renew their contract.
- Expected revenue next month, with a range.
- A deviation in a series, for example a customer who was invoiced EUR 8,000 every month and is now at EUR 5,200. That is called anomaly detection.
Its strength: the outcome is measurable. After three months you can see exactly how often the model was right. Its weakness: the model can only recognise patterns that are in the data. A new arrangement that sits in an email and not in any field is invisible to it.
What does generative AI do?
Generative AI, in practice a large language model, is strong at anything to do with language. And that is exactly where a lot of B2B information lives that never made it into a field:
- A thirty-page contract with an indexation clause on page 22.
- An email in which an account manager promises a discount "for this year".
- A quote with additional work that was never entered as an order line.
- Notes in the CRM about an expansion the customer wants.
A language model can read such text, extract the relevant arrangement and return it in a fixed format: customer, arrangement, amount, start date. Ordinary software can then compare that arrangement with what has been invoiced.
Its weakness is well known: a language model produces probable text, not verified facts. If it does not know something or the source is unclear, it can give an answer that sounds precise and is still wrong. That is why AI output always has to be checked against the source.
Why is a language model not a calculator?
A common mistake is to have a language model do calculations on business data. You paste an export of 2,000 invoice lines into a chat and ask: which customers are paying too little? You get an answer. It looks good. It cannot be trusted.
A language model does not add up like a database. It reads the text and writes what the answer probably is. With a few lines that often goes well; with thousands of lines it goes wrong invisibly: lines are skipped, amounts swapped, totals rounded. Modern systems solve this by not letting the language model calculate at all, but having it call a database query or a calculation function instead. That principle is called tool calling.
The rule of thumb: let a language model decide what needs to be calculated, and let software do the calculating.
Worked example: checking indexation with both
Worked example: suppose you have 180 service contracts. Some of them have an indexation clause, but exactly which ones, and at what rate, is only recorded in the PDFs.
Step 1, generative AI. A language model reads the 180 contracts and extracts for each one: is there an indexation clause, which index or percentage, from what date. Output: a table. An employee checks a sample of twenty contracts against the PDFs.
Step 2, ordinary software. A query compares that table with the invoices in the accounting system. Suppose 45 contracts should have been indexed by 3 percent, and for 11 of them the rate was not adjusted.
Step 3, calculation. Those 11 contracts have a combined annual value of EUR 660,000. Missed indexation: 3 percent of that is EUR 19,800 a year, and the amount compounds for as long as it is not corrected.
Step 4, predictive AI. A model learns which attributes are associated with missed indexation: contracts closed by a particular team, contracts with an unusual start date, contracts invoiced manually. It flags new contracts with that profile, so you check them first in the next indexation round.
The numbers are an example. The point is the division of labour: the language model reads what is in text, software calculates, and the predictive model points to where you should look first next time. More on this particular leak in revenue leakage from missed price indexation.
When should you use predictive AI and when generative AI?
Ask yourself three questions.
1. Is the output a number or text? If you want a probability, an amount or a ranking, predictive AI is the basis. If you want a summary, an explanation or a draft, generative AI.
2. Is the information in fields or in text? If it sits in tables, such as orders, invoices and payments, ordinary software or predictive AI can work with it. If it sits in contracts, emails or notes, you need a language model to read it first.
3. Can you measure the outcome afterwards? With predictive AI that must be possible. If you cannot establish whether a prediction was right, you cannot improve the model. With generative AI you check the output beforehand, against the source.
Where does it go wrong in practice?
- A chatbot as an analysis tool. Asking a language model questions about a pasted spreadsheet produces answers that are not reproducible. Ask the same question twice and you may get different numbers.
- A predictive model without an explanation. A churn score of 0.82 without a reason cannot be checked and cannot be used in a conversation with the customer.
- Generated text passed on as fact. A contract summary that goes straight into the CRM without anyone having looked at the source.
- Calling everything AI. Many revenue checks are plain comparisons: contract price against invoice price. They need no AI, and a simple rule is more reliable.
How do they work together in one system?
In a mature set-up both kinds of AI sit side by side, with ordinary software in between. That pattern recurs in the AI architecture for Revenue Intelligence and in the overview article on how AI works within Revenue Intelligence:
- Data is pulled from CRM, billing and accounting.
- A language model reads unstructured sources and turns arrangements into fields.
- Rules and queries compare what was agreed with what was invoiced.
- Predictive models and anomaly detection point to deviations and risks.
- A language model phrases the finding in plain language, with the source attached.
- A person reviews and decides.
Frequently asked questions
Is ChatGPT predictive AI or generative AI?
Generative AI. Technically a language model keeps predicting the next word, but its purpose and its output are text. By predictive AI we mean models that predict a measurable outcome, such as a probability or an amount.
Which is more reliable?
That depends on the task. Predictive AI is easier to test, because you can measure the outcome later. Generative AI is stronger at reading and writing, but you have to check its output against the source every time.
Can generative AI forecast my revenue?
Not reliably by estimating it itself. It can do so by calling a predictive model or a calculation function and explaining the result. The language model is then the interpreter, not the calculator.
Do I need both for revenue control?
If important arrangements exist only in contracts and emails, yes. If everything is neatly in fields, plain comparisons and predictive models will get you a long way.
More in this cluster
- How does AI work within Revenue Intelligence?Start here
- AI architecture for Revenue Intelligence
- What is anomaly detection?
- What is predictive analytics?
- What is an AI agent?
- What are AI agents in RevOps?
- What is tool calling?
- How do you give AI access to business data?