AI revenue forecasting explained

AI revenue forecasting is predicting revenue with models that learn from your own historical deals, invoices and customer behaviour, instead of from probabilities someone filled in per stage. The model estimates, per deal and per customer, how likely revenue is and when it will arrive, and adds that up into a forecast with a range. It only works when there is enough history and the data is right, and it does not replace the judgement of the people who know the customers.

What does AI do differently from an ordinary forecast?

A traditional pipeline forecast works with fixed rules. A deal in the stage "Quote sent" counts for 50 percent, a deal in "Negotiation" for 75 percent. Someone once came up with those percentages, and they apply to every deal in that stage, whether it was created yesterday or has been stalled for eight months.

An AI model turns this around. It looks at every deal you have won and lost in the past, and learns which characteristics went together with winning, losing and delay. For example:

• how long a deal has been in its current stage, compared with your normal cycle length;
• how often the close date has already been moved;
• the number of touchpoints in recent weeks, from email and calendar;
• the size of the deal relative to what that customer or segment usually buys;
• whether it is a new customer or an existing one;
• how the salesperson has historically estimated their own deals.

For recurring revenue, a model looks at each customer's buying pattern, payment behaviour, support tickets and contract data. The result is an expectation per customer, instead of one growth percentage applied to everything.

The difference from the classic approach is worked out further in AI forecasting vs traditional forecasting. This article is about how it works.

How does a forecasting model learn?

In plain language, it goes like this.

1. Collect. The model receives a table of all closed deals from recent years: for each deal, its characteristics as they were while it was still open, and the outcome (won or lost, and when).
2. Train. The model looks for the relationship between the characteristics and the outcome. Common techniques are logistic regression and gradient boosting for the probability of winning, and time series models for recurring revenue.
3. Test. You have the model predict a period it has not seen, for example the last quarter, and compare it with what really happened. Only if it does better there than your current method does it add value.
4. Apply. The model assesses every open deal and every customer, and adds up the expectations into a forecast.
5. Keep learning. Every closed deal becomes new training data. A well set-up system retrains periodically and tracks whether accuracy stays stable.

Step 1 has a trap that undermines many home-built models. If training uses fields that are only filled in after a deal is won, such as a project number or a contract date, the model learns that those fields predict winning. On historical data it then scores perfectly. On open deals it is worthless, because those fields are still empty. A good model only uses what was known at the moment of prediction.

Where does generative AI fit, and where does it not?

The word AI now mostly brings to mind language models, such as the assistants you chat with. They are not built to predict numbers. The forecast figure itself comes from a predictive model: statistics and machine learning on structured data. The difference is explained in predictive AI vs generative AI.

Language models do have a role around it. They can read notes and emails and pick up signals, such as "customer mentions budget freeze" or "decision-maker has left". And they can explain in plain language why the forecast dropped this week. But a language model that makes up a revenue figure by itself gives you a number that sounds plausible and rests on nothing.

Worked example: what the difference can be

Worked example: suppose you have 40 open deals in the "Quote" stage, together worth EUR 2,000,000. The CRM counts them at 50 percent, so the weighted pipeline is EUR 1,000,000.

A model looks at each deal. Twelve deals, together EUR 700,000, have been in that stage for more than 90 days. Of comparable deals in the past, 10 percent were won. The other 28 deals, together EUR 1,300,000, are recent and active. Historically, 45 percent of those were won.

• Old deals: EUR 700,000 x 0.10 = EUR 70,000
• Recent deals: EUR 1,300,000 x 0.45 = EUR 585,000
• Total: EUR 655,000

The difference from the CRM weighting is EUR 345,000. That is no proof the model is right. It is, however, a question someone needs to answer before the number goes to the bank or the shareholders. The figures here are illustrative; in your data the difference may be larger or smaller.

When does AI forecasting work?

AI forecasting delivers most under three conditions.

Enough history. A model learns from patterns, and patterns need repetition. A company that closes 60 deals a year has 180 examples after three years. There is something to learn from that, but not much per segment. A company with 2,000 orders a year has far more to go on. There is no hard minimum, but the less data, the wider the range should be and the more often a simple model beats a complex one.

Data that is correct. If deals are not marked as lost, close dates are not updated and won deals are not linked to invoices, the model learns the wrong lesson. That is no reason to wait, because you can work with gaps too, see forecasting with incomplete CRM data. It is a reason to have the model show you where the data is shaky.

A stable market. A model learns from the past. If your market changes abruptly, through a new product, a price change or an economic shock, the past has less to say. A good system notices this through falling accuracy and reports it.

Where does AI forecasting go wrong?

• Too much trust in one number. An AI forecast is an expectation with uncertainty. A system that shows a single amount without a range hides the most important part. How to read that uncertainty is covered in forecast confidence explained.
• A black box. If nobody can see why a deal sits at 20 percent, sales will ignore the number or work around it. A usable model shows, per deal, which characteristics pulled the probability down or up.
• Only pipeline, no invoices. A forecast that stops at "won" misses what happens next: deals that are never invoiced, discounts after the fact, orders that turn out smaller. The revenue that counts is invoiced revenue.
• Gaming the model. If salespeople know which characteristics raise the score, they will fill in those characteristics. Logging activity to push up the score makes the model worse. Use the score to have conversations, not to hold people to account.

How do you assess an AI forecast?

Whether you build one yourself, switch on a module in your CRM or choose a separate platform, these questions separate the useful from the merely attractive:

1. What data does it train on? Only the CRM, or also invoicing and contracts? A forecast that only sees the CRM predicts pipeline, not revenue.
2. How was it tested? Ask for its accuracy on a period the model had not seen, compared with your current method.
3. Does it show a range? And does that range hold afterwards, so reality falls inside it roughly as often as promised?
4. Can you see why, per deal? Which characteristics lowered the probability, and can you explain that to a salesperson?
5. What happens with deviations? Does the system report when its own accuracy falls, or does it keep predicting in silence?
6. Who decides? The model makes a proposal. A person signs off the forecast that goes outside the company.

RiOS includes forecasting as a module that gives a revenue forecast with a likely range, fed by the systems you connect. What that looks like is shown on the page about the system. The principles above apply to any tool.

For the place of all this within the broader discipline, see what revenue forecasting is.

Frequently asked questions

Is AI forecasting more accurate than a spreadsheet?

Often, with many deals and good data, but not automatically. With little data, a simple method can be just as good or better. The only fair test is to have both methods predict a past period and compare.

How much data do you need?

There is no fixed number. It depends on how many deals or customers you have, how similar they are and how stable your market is. With a few hundred closed deals you can start; with fewer, you are better off with simple rules and a wide margin.

Can an AI model predict a deal that is not yet in the CRM?

Not per deal. It can, however, estimate from history how much revenue on average comes from deals that do not yet exist within the period. That part belongs in the forecast as a separate, uncertain item.

Does AI replace the forecast meeting?

No. It changes the conversation. Instead of going through every deal, you discuss the deals where the model and the salesperson disagree. That is where the information is.

Should you show sales the model?

Yes. A salesperson who sees why a deal scores low can refute it with information the model did not have, or close the deal out honestly. Both make the forecast better.