AI sales forecasting: how the models work, and how to backtest one on last year's numbers

This guide is for sales leaders, revenue operations managers and finance partners deciding whether to use AI for sales forecasting, or whether to trust the AI forecast their CRM already shows. It explains the two main approaches, what three major CRMs say their models use, how to measure forecast accuracy, and how to run a backtest on your own closed quarters. For the full set of forecasting methods, including ones that don't use AI, see the guide to sales forecasting methods.

Four ways to forecast, and where AI fits

AI for sales forecasting covers the last two rows. Most teams end up with a blend: an AI baseline, rep judgment on the large deals, and a manager who asks why the two differ.

What the major CRMs say their AI forecasts use

Vendors' own documentation, checked on 26 October 2026:

Two things stand out. First, these models answer different questions: HubSpot's extends recent revenue, while Salesforce's and Microsoft's work from the deals in front of you. Second, the published minimums are tiny. Microsoft notes that fewer opportunities means less training data and lower accuracy; ten closed deals will build a model, not a reliable one.

What AI forecasting does well, and what it doesn't

Does well: applies the same rules every week, without optimism or sandbagging; spots deals whose activity pattern looks like past losses; updates as soon as the data changes; and gives a range, which is more honest than a single number.

Does badly: new products or segments with no history; a quarter that depends on two or three large deals; changes it can't see, such as a price rise, a new competitor or a reorganized team; and deals that live in a rep's head rather than the CRM. It's also only as good as close dates and amounts. If reps push close dates every month, the model learns that close dates mean nothing.

How to measure forecast accuracy

Pick a fixed point in each period (for example, the end of week two of the quarter) and compare the forecast then with what actually closed.

• Error: forecast minus actual, in dollars.
• Percentage error: error ÷ actual. A forecast of $1.6 million against $1.42 million actual is +12.7%.
• Mean absolute percentage error (MAPE): the average of the percentage errors ignoring sign. It tells you how far off you typically are. It breaks down when actuals are near zero, so use dollar error for small or new segments.
• Bias: the average of the percentage errors keeping the sign. Consistently positive means the method runs high. Bias is the easiest problem to fix, because you can correct for it.

HubSpot reports accuracy as 1 minus the forecast error percentage, floored at zero, so a 12.7% error shows as 87.3% accuracy. Whatever tool you use, compare methods on the same periods and the same snapshot week.

Worked example: backtesting four quarters at a cybersecurity vendor

Illustrative example. The company and numbers are invented to show the method.

The company: Norhaven Security (norhaven.example) sells endpoint and email security subscriptions to companies with 200 to 2,000 employees. It has 9 account executives and closes 60 to 80 new deals a quarter.

The test: for each of the last four quarters, the operations manager records three forecasts as they stood at the end of week two: the reps' commit, the weighted pipeline from the CRM's stage probabilities, and the AI deal-level prediction rebuilt from the CRM's history as of that date. Then she compares each with actual new bookings.

What it shows:

• The weighted pipeline ran 31% high every quarter. Its stage probabilities were set years ago and had never been checked against actual win rates by stage. That's a fix worth making whatever forecasting tool you use.
• The reps ran about 11% high, mostly in Q1, after the holiday slowdown they forecast every year and never saw.
• The AI had the lowest error, but its worst miss was Q3, when a $240,000 deal from a new segment (a hospital group) closed. The model had no history for that segment and gave it a low probability. The one rep who knew it was coming was right.

The decision: Norhaven uses the AI prediction as its baseline, asks reps to explain any gap of more than 10% between their commit and the AI figure deal by deal, and reviews deals over $150,000 by hand. It also reset its stage probabilities to the last eight quarters' actual win rates.

Run a backtest on your own data

1. Choose the snapshot point, such as the end of week two of each quarter, and keep it fixed.
2. Collect four to eight closed periods. Fewer than four can't separate skill from luck.
3. Record every method's forecast as it stood at the snapshot: reps' commit, weighted pipeline and the AI prediction rebuilt from data available then. A model that has seen the outcome will look perfect.
4. Calculate error, MAPE and bias for each method.
5. Read the biggest misses one by one. They tell you which deals or segments each method can't handle.
6. Decide the rule: which number is the baseline, when reps must justify a gap, and which deals get reviewed by hand.
7. Repeat each quarter, because a model's accuracy can change after a retrain or a change in your market.

Get the inputs right first

• Close dates. Track how often each deal's close date moves. A pipeline where most deals slip at least once can't be forecast accurately by anyone.
• Stage definitions. Each stage should have an exit test a manager could check, such as "buyer confirmed budget owner", not "had a good call".
• Amounts. Record the amount when the deal is created and track changes, so you can see discounting.
• Automation. If AI fills in CRM fields from call notes, keep close date and amount out of reach; the guide to AI sales automation explains why.
• Coverage. A forecast can't be better than the pipeline under it. If the honest forecast is short, the answer is more qualified pipeline, which is a prospecting problem; Bebop finds companies and contacts that match a plain-language description and explains why each fits.

What to ask a vendor

• What data does the model use: closed revenue only, or the open pipeline and activity too?
• Can I see its accuracy on my own past quarters, at a fixed snapshot week, before I buy?
• Does it show a range, and the factors behind each prediction?
• How does it treat deals from segments or products with no history?
• How often does it retrain, and will I see if accuracy changes?

What it costs

AI forecasting is usually part of a CRM's upper tiers or a revenue platform priced on quote. As of 26 October 2026, HubSpot's AI projections (in beta) come with Sales Hub Professional and Enterprise, Microsoft requires a specific Dynamics 365 Sales license for premium forecasting, and Salesforce lists Predictive Forecasting among the built-in AI in its Sales Cloud editions, priced from $195 to $550 a user a month billed annually. Before paying for a dedicated tool, backtest the forecast your CRM already gives you.

Next step: set up a forecast you can measure with the sales forecast template.