AI forecasting vs traditional forecasting

Traditional forecasting uses fixed rules that people have devised: a percentage per pipeline stage, the salesperson's estimate, or last year plus a growth figure. AI forecasting learns those rules itself from your historical data and applies them per deal and per customer. AI usually wins with plenty of data, many signals and a forecast that has to be right every week; the traditional approach wins with little data, many large one-off deals and when explainability matters more than the last few percent of accuracy.

What counts as traditional forecasting?

"Traditional" is not a single method. In B2B you come across four, often mixed together.

• The sales commit. Each salesperson states which deals they will bring in this quarter. The sales manager adds them up and takes something off by feel.
• Stage-weighted pipeline. Each stage in the CRM has a fixed percentage. Value times percentage, added up, is the forecast. Salesforce, HubSpot and Pipedrive have this built in as standard.
• Top-down trend. Last year plus a percentage, possibly with a seasonal pattern. Usually built by finance in Excel, based on the accounts.
• Driver-based. A spreadsheet model with assumptions: so many leads, such a conversion rate, such an average deal value. Useful for budgets, weak as a prediction, because every assumption is an estimate.

What these methods share: a person sets the rule, and the rule applies to everything in the same category.

What does AI forecasting do differently?

An AI model is given the outcomes of the past and works out for itself which characteristics went together with winning, losing and delay. It applies them per deal or per customer. Two deals in the same stage can then get very different probabilities, because one has been stalled for four months and the other had three meetings last week. How such a model learns and is tested is explained in AI revenue forecasting explained.

AI vs traditional forecasting at a glance

Where does traditional forecasting go structurally wrong?

The weakness of the classic method is rarely the arithmetic. It is that the input comes from people who have a stake in it.

A salesperson who has to hit a target keeps deals open longer. A manager who wants to rescue the quarter moves close dates out by a month instead of closing deals out. A stage percentage of 50 percent that was once right is no longer right once the sales process has changed, and nobody adjusts it. These mechanisms are described in more detail in why are sales forecasts so often wrong.

The result is not just inaccuracy, but bias: an error that keeps pointing the same way. A forecast that is 10 percent too high every quarter is actually quite predictable. It is just that nobody corrects for it.

Where does AI forecasting go structurally wrong?

AI has no interest in hitting a target, but it has other weaknesses.

• It learns from the past. New products, a new market or a price change are not in the history. The model then does not know what it does not know.
• It also learns your bad habits. If deals have not been marked as lost for years, the model learns that old deals sometimes still close. That is true in your data, but not in reality.
• Little data gives wild outcomes. With 50 deals a year, a complex model finds patterns that are coincidence.
• Hard to explain. If a salesperson does not understand why their deal is at 15 percent, they will not trust the model, and then it gets ignored.

How do you test them fairly against each other?

Debates about which method is better are usually based on feel. You can also measure it, with a backtest: have both methods predict a past period, using only the information available at the time, and compare with what really happened.

Worked example: suppose you compare four quarters, in millions of euros.

The traditional forecast was off by 10.3 percent on average, and too high every time. The AI forecast was off by 4.9 percent on average, sometimes too high and sometimes too low.

Two lessons. First, in this example the AI forecast is more accurate. Second, the traditional forecast is systematically too high. That bias is valuable information in itself: even without AI, a 10 percent correction on the commit would have closed most of the gap. A backtest therefore tells you not only which method wins, but also what is wrong with the method you have now.

Watch three things in a backtest:

1. Use only the data from the time. Today's CRM contains updated close dates and amounts. If your system does not keep a history of changes, a fair backtest is difficult. In that case, start saving a weekly snapshot now.
2. Measure at the level you steer on. A method that is good per quarter but poor per month is fine for the board and useless for cash planning.
3. Look at bias and spread. The average deviation tells you how far off you are. The direction tells you whether you are structurally too optimistic.

When should you use which approach?

Traditional is enough, or wins, when you close few deals a year, when a handful of large deals decides the year, when your CRM history is short or unreliable, or when your business has just changed something fundamental. For many smaller companies, an honest top-down trend plus a critically reviewed deal list is the best forecast that can be made. More on this in revenue forecasting for SMEs.

AI wins when you have hundreds of deals or thousands of orders a year, when besides the CRM you can also connect invoicing, contracts and customer behaviour, and when the forecast has to be updated often without someone spending a day on it.

A combination works best in most companies. The model gives a probability per deal, the salesperson gives their estimate. Where they agree, there is nothing to discuss. Where they differ sharply, that is where the conversation is. The model can see things the salesperson misses (the deal has been stalled twice as long as normal), the salesperson knows things that are not in the system (the decision-maker called yesterday). Final responsibility stays with a person, the principle behind human-in-the-loop AI.

For the recurring base, a model on invoice data often works best, see forecasting from historical revenue. For new deals, the combined model above. How those layers come together into one forecast is set out in the complete explanation of revenue forecasting.

What does switching cost?

The largest cost is rarely the software. It is the data. An AI forecast needs a CRM in which deals are closed out on time, close dates are kept up to date and won deals can be linked to invoices. That clean-up takes time, but it produces a better forecast even without AI.

The second cost is trust. Run the model alongside your current method for a few quarters first, with nothing depending on it. Only when it is demonstrably better do you let it count.

Checklist for your choice

1. How many closed deals or orders do you have per year, and how many years of history?
2. Are lost deals actually marked as lost, and within what time?
3. Does your CRM keep the history of stage and date changes?
4. Can you link won deals to invoices?
5. How far off was your current forecast over the past four quarters, and in which direction?
6. Who needs to be able to explain the forecast, and to whom?

Frequently asked questions

Is AI forecasting always more accurate?

No. With plenty of data and a stable market it usually is, with little data or after a major change it is not automatically. Only a backtest on your own data gives the answer.

Can you improve traditional forecasting without AI?

Yes. Measure your bias and correct for it, consistently mark old deals as lost, and put the sales forecast next to a top-down trend. That often already delivers a large part of the gain.

What is the biggest risk of AI forecasting?

That people believe the number without understanding where it comes from. A model without explanation and without a range suggests more certainty than it has.

Should salespeople still submit a forecast if there is a model?

Yes. Their estimate is information the model does not have. The difference between salesperson and model is exactly what you should be looking at.