Most AI forecast demos end on the same slide: a clean range, a midpoint, and a sales leader quietly wondering whether that number can go in front of the board.
Whether it can depends less on the model than on what the model learned from. An AI sales forecast is trained on your own CRM history, so it inherits whatever your closed-won data gets wrong. You can find out before the board does. Score the AI number and your team’s submitted forecast against what closed, the same way, for three months, and check the inputs the model reads while you’re at it. HubSpot already builds most of that test into a table that’s easy to miss.
What does an AI sales forecast learn from your CRM?
HubSpot’s Breeze AI forecast learns from closed won deals in the past three months, specifically the date a deal moved to Closed won and the deal’s amount, according to HubSpot’s knowledge base (Improve forecasting with AI projections, last updated May 29, 2026). It returns a range with a most-likely midpoint, plus upper and lower bounds, and refreshes on days 1, 7, 14, 21 and 28 of each month.
HubSpot is plain about the limits on that same page. “The accuracy of the projection depends on the accuracy of your data.” Projections “are estimates and are not guaranteed,” and they “should only provide perspective and not be solely relied upon to make business decisions.” The feature is in beta, it needs Sales Hub or Service Hub Professional or Enterprise, and the accuracy table needs a forecast cadence set to monthly.
So the model reads two fields, on deals that already closed. It doesn’t care what your reps think about open pipeline, which is the appeal. It also can’t tell when those two fields are wrong, which is the risk. Other AI forecast tools use their own inputs, so find out what yours reads before you trust what it says.
The CRM habits that mislead a model trained on closed-won history
If the model reads Closed won dates and amounts, anything that bends those two fields bends the forecast. These aren’t HubSpot statements. They follow from the inputs HubSpot documents, and you can check each one in your own portal.
- Deals marked won late. A deal signed on the 29th and marked Closed won on the 3rd lands in the wrong month. Do that often enough and the model learns a calendar your buyers never kept.
- Amounts edited at close. When the amount on a won deal gets rewritten to match the signed order, or one rep enters first-year value while another enters the full contract, the model learns from numbers that don’t mean the same thing.
- Deals never closed out. A deal that signed but still sits open isn’t in closed-won history at all. The model undercounts the month it belonged to.
- An unusual month in the window. Three months is a short memory. A one-off large deal or a holiday lull inside that window gets carried forward into the projection.
Close dates are a shaky input to begin with. InsightSquared and RevOps Squared’s 2021 State of Sales Forecasting research found less than half (47%) of opportunities close within the original forecast date, and of those that push, they move five times at median (press release, May 25, 2021). A model that learns when deals close is learning from a field your team moves a lot. (For the rep side of that story, see why reps forecast wrong.)
How does HubSpot score its own AI forecast?
HubSpot’s Forecast accuracy over time table compares each month’s Closed won total with the AI projections and with your forecast submission, using the team submission when one exists and otherwise the sum of rep submission averages (HubSpot knowledge base, May 29, 2026). Accuracy comes from the absolute percentage difference between forecast and actual, floored at zero, and the monthly figure averages the five projections.
In HubSpot’s own words, “Forecast accuracy for each day projection is measured by computing the absolute percentage difference between the forecasted and actual sales numbers,” and “When the forecast error is greater than 1, accuracy will always be 0.” The five projections are the 1-, 7-, 14-, 21- and 28-day ones.
The part that’s easy to skip is the comparison. That table already puts the AI number next to the number your team submitted, scored against the same actual. That’s a backtest, built in, and it only needs a monthly cadence to run. If your HubSpot forecast already drifts from what closes, why a HubSpot forecast comes out wrong covers the usual causes, and how to measure sales forecast accuracy covers the scoring outside HubSpot.
A three-month backtest you can run
The test is simple to describe and tedious to do, which is why it rarely gets done. Here’s what three months of it can look like.
The steps, in order:
- Set the forecast cadence to monthly and opt into AI forecasting. HubSpot’s accuracy table needs the monthly cadence, per the same knowledge base page.
- Record the AI range and the submission on day 1 and day 14 each month. Two snapshots tell you whether the AI number gets closer as the month goes, or the submission does.
- At month end, score both against Closed won the same way. Use HubSpot’s table, or apply its method to your snapshots by hand. Same actual, same math, both numbers.
- Before blaming the model, audit the month’s closed-won deals. Check Closed won dates against the real signature date, and look for amounts that changed at close. Why HubSpot reports aren’t accurate walks through where those fields drift.
The decision rule: if the AI number beats the submission in most of the months you test, show it beside the roll-up as a check. If it doesn’t, keep it out of the board deck and work on the inputs first. Resist picking an accuracy percentage as a pass mark. The question is relative, AI against your team, on your deals.
Does the AI forecast replace the rep roll-up?
No. It’s a second opinion built from a different input. HubSpot’s forecast tool sorts deals into forecast categories by deal stage, lets users override the category on a single deal, and uses Weighted amount (Amount times Deal probability) by default (Set up the forecast tool, September 9, 2026). The AI projection reads deals that already closed. When the two disagree, the gap is the conversation.
Human forecasts miss too, which is why the test is relative. InsightSquared and RevOps Squared’s 2021 State of Sales Forecasting research, a survey of nearly 400 B2B companies run in April 2021, found 91% of participants said their predicted forecast was six percent or more off from actual results, 68% saw more than 10 percent deltas between forecast and actuals, and only 15% of revenue leaders were very satisfied with their forecast process (press release, May 25, 2021). In that 2021 survey, most teams missed their own forecast and few liked how they built it.
Most teams in a 2021 survey missed their own forecast.
Share of participants whose forecast was off from actual results
Source: InsightSquared and RevOps Squared, 2021 · nearly 400 B2B companies, April 2021 · six percent or more off; more than 10 percent off
| Group | Share |
|---|---|
| Six percent+ off | 91% |
| Over ten percent | 68% |
So the bar for the AI number is the roll-up you already present, scored on your data. When they split, ask one of two questions. If the AI number is lower, which Commit deals look nothing like what closed in the last three months? If it’s higher, did last quarter carry a one-off, or is someone holding deals back? (If your team uses the categories loosely, the forecast categories glossary is a good place to align them.)
What the backtest tells you about your data
A model that scores badly on your history is often a data finding before it’s a model finding. If the audit in step four turned up late close-outs or rewritten amounts, the model learned those habits as if they were the market, and so did every report built on the same fields.
Close date behavior moves outcomes, not only forecasts. Ebsta and Pavilion’s B2B Sales Benchmark Report 2024, which analyzed 4.2 million opportunities from 530 companies, found that close dates moved more than 3 times see win rates reduce by 77% (B2B Sales Benchmark Report 2024). The same field that confuses the model is telling you something about the deal.
If you want to know whether your CRM can carry AI at all, start with what AI-ready data means and the AI readiness assessment.
Earn the number before it goes in the deck
You can put the AI forecast in front of the board on the vendor’s word, or you can earn it with three months of scoring against what closed. One of those holds up when a director asks how you know.
This gets harder when the history itself is the drag: when close dates and amounts mean different things across reps, the model and the roll-up will both miss, and the table can’t tell you why. AeolusGTM is CRM, RevOps and GTM advisory, with Atlas connecting to your CRM to find gaps the CRM can’t show natively. If you want a read on whether your data is ready for an AI forecast, start with the AI Readiness report.
This week, open the Forecast accuracy over time table in HubSpot. If your cadence isn’t monthly yet, write down today’s AI range and your team’s submission instead. Those two numbers are month one.