Yes: AI can improve sales forecast accuracy by reading calls, emails and CRM activity to spot deals that look likely to slip, and it improves accuracy further when each flag leads to a conversation that gets the buyer evidence the deal is missing.
The short version
A better prediction makes the number more accurate because doubtful deals get marked down earlier. The deal itself is unchanged. If a committed deal lacks a signer who has heard the case, the model can say it looks risky, but the rep still has to have the conversation that gets the signer in the room. That conversation is where the deal earns its place in commit or leaves it. The longer argument is in why a risk score is not a fix.
The tools are good at the prediction half. They read far more of each deal than a manager can, and they compare it with deals that closed or slipped before. The vendors publish strong numbers: Gong says its Deal Likelihood Score, built on more than 300 buying signals, predicts which forecast deals will close with 20% more accuracy than CRM data alone and 22% more than reps' own predictions, and Outreach says its deal health model's precision was 81% against 34% for a rules-based model, from its internal research. Both are vendor measurements on vendor data, and both measure the prediction half only.
How to test an AI read
Test any AI judgment before you trust it with the forecast. We learned this in our call-scoring research: one leading model changed 15% to 17% of its scores between identical runs, and a scorer that never read the transcript landed a better average error than the best real model. So a low average error says little; judge on ranking. Two checks carry over to forecasting:
- Re-run it. Score the same pipeline twice. If deals change category with no new activity, the read is noise.
- Judge it on ranking. Ask whether it puts deals in the same order as the managers who know them best, and whether that order matched what closed.
Two questions a forecast model answers
A forecast model can answer two different questions about a deal: will it close at all, and will it close in this period. They rest on different evidence. A deal with a strong champion and a stated cost may well close, but if nobody has walked the rep through the approval steps it can still land next quarter. When a tool gives a single probability, ask which question it answers. For this quarter's number, the second question matters more, and it depends on the decision and paper process more than on how keen the buyer sounds.
Take one illustrative deal. The champion, a head of finance operations, has said on a call what the manual process costs her team, and she wants it fixed. On "will it close", that deal looks strong. Then the manager asks the rep one question: "Who signs, and what does legal need before they do?" The rep does not know.
The champion has never mentioned procurement, and half the quarter has gone. The honest forecast is "likely to close, probably not this quarter" until the rep asks the champion to walk through the last purchase her team made of this size. A model that returns one probability blends those two answers, and the manager reads it as the first when the number depends on the second.
Where AI forecasts go wrong
Most errors come from the inputs and from how the team uses the output.
- Loose stage definitions. If commit means "the rep feels good" to one manager and "the signer is engaged" to another, the model learns both habits. Define each stage by the buyer evidence it needs before you tune anything.
- Close dates that drift. A close date pushed every month teaches the model little. Record why each date moved; slip reasons are some of the most useful data a team has.
- Thin or messy history. A model trained on a few dozen closed deals, or on a CRM where half the activity was never logged, will find patterns that are not there.
- A few large deals. When two or three deals make up most of the quarter, one slip moves the number more than any model can predict. Inspect those deals by hand every week.
- Treating the score as the decision. A risk score is a reason to look. The manager and rep still decide what the deal needs, and the rep still has to go and get it.
How to start
- Tighten the stage definitions. Write down the buyer evidence each stage needs, in your method's terms.
- Take a snapshot. Record the committed and best-case deals as they stand today, so there is something fixed to compare against.
- Run the tool alongside the current forecast. Keep the manager's call as the forecast for now, and compare both against what closed at the end of the period.
- Act on the flags (below).
- Review after two quarters. Keep the tool if its ranking matched the outcome and its flags led to work on the deals.
What to do with a flagged deal
A flag is only useful if someone acts on it before the next buyer conversation. For each committed deal the tool marks as high risk, the manager and rep work through three questions, in about ten minutes:
- What evidence is missing? Name it in buyer terms: no signer on a call, no cost stated by the buyer, no dated next step, no known approval path.
- Who can supply it, and when do we next speak to them? Usually the champion, on a call already booked. If nothing is booked, booking it is the first job.
- What will the rep say? Write the opening line and say it out loud, along with the likely pushback and the reply.
Then decide where the deal sits until that conversation happens. A deal with a plan and a date can stay in commit for a week. A deal with neither moves to best case now, which is the change that makes the number more accurate.
An illustrative quarter
This example is illustrative. In week 3 an AI forecast tool marks 6 of a team's 20 committed deals as high risk, and four of them have never had the signer on a call. Each of those reps gets a 20-minute coaching call that week, names the missing evidence, and says the opening line for the champion out loud before the call is booked.
- Committed deals20
- Marked high risk in week 36
- Signer never on a call4
- Signer engaged by week 62
By week 6, two of the four have the signer engaged and stay in commit. Two move to best case. The forecast is smaller, and the two deals that stayed in commit now have a signer engaged, because the flags turned into work on the deals. On our platform that rehearsal is a call with Caty, or with Keenan or Val for Gap Selling and ValueSelling Framework® teams. Bryce, the manager's coach, gives a Monday team read: which deals need attention this week and which deals carry buyer evidence. To see it on one of your own flagged deals, talk to us.
Next reads on AI and the forecast
- How to improve sales forecast accuracy: the full procedure, from snapshot to slip reasons.
- Deal risk guide: where deal risk comes from and what to do with a flag.
- How to run a forecast call: the weekly evidence questions.
FAQ
What data does AI sales forecasting use?
Most tools use CRM fields such as stage, amount, close date and time in stage, plus activity from email, calendar and calls. Conversation-based tools add what was said on calls. Outreach lists engagement across email, calls and meetings, CRM activity and prospect information as its deal health inputs.
Will AI forecasting replace the forecast call?
AI forecasting will not replace the forecast call. The AI narrows where the manager looks, while the call is where a rep is asked for buyer evidence and leaves with a plan to get what is missing. Use the AI read to choose which deals to discuss, then spend the call on those deals.
How long does AI forecasting take to get accurate?
How fast AI forecasting gets accurate depends on how much clean history the model has and how consistent your stage definitions are. Models trained on history learn a team's habits, including loose definitions of commit, so tighten the definitions first and measure the AI's read against a fixed snapshot for at least two quarters.
Does AI forecasting work for a small team?
AI forecasting can help a small team, though there is less history for a model to learn from. With a few dozen deals a quarter, the AI read is most useful as a check on each committed deal: is the signer on a call, is the next step dated, is the approval path known? That check needs no long history, and a manager can run it with each rep in the forecast call.