Sales

AI forecasting (win probability)

A model trained on your own history predicts the chance of winning each open opportunity, at no AI cost.

The "AI Forecast" page (in the side menu) uses YOUR history of won and lost opportunities to learn which characteristics lead to a win and, from that, estimate the probability of closing each open opportunity.

How it works

The model is trained automatically on the opportunities already closed (Won/Lost).
It analyzes characteristics such as stage, source, sale type, priority and value range.
Each open opportunity receives a win probability and a band (High, Medium, Low).
The "weighted pipeline" adds up the value of the open deals multiplied by the predicted probability, giving you a realistic forecast.

The screen also shows the factors that weigh most for and against a win, learned from your history.

💡 A minimum of history is needed (about 10 closed deals) for the model to train. The more opportunities you close as Won/Lost, the better the forecast.

Churn risk

On the same page, the "Churn risk" section learns from your history of lost accounts (lifecycle "Lost" or type "Churn") versus retained ones and estimates the risk of each ACTIVE account canceling. It uses leading signals such as health and renewal status, and lists accounts from highest to lowest risk.

Expansion and downgrade propensity

The same engine predicts which accounts are most likely to EXPAND (upsell/cross-sell) and which are most likely to DOWNGRADE/contract. It learns from your REAL revenue movement history: the MRR expansions and contractions of subscriptions over time. Features are account attributes (tier, size, industry, health, renewal status, plan). Ask the Copilot "which customers have expansion potential?" or "who might downgrade?" and it lists accounts from highest to lowest score, with the signals driving the prediction.

💡 The model trains on the history of subscription MRR changes. If you do not use the Subscription object yet or have few recorded expansions/contractions, the sample is insufficient and the result is flagged as untrained, without making up numbers.

Renewal forecast and future revenue

Ask the Copilot "which contracts will renew?" or "how much MRR/ARR will we have?". It projects MRR/ARR for the coming months from the current active MRR and the average net growth (new + expansion − contraction − churn) and estimates the renewal probability of each active subscription, trained on your history (active versus canceled subscriptions), listing them from lowest to highest (most at risk first). Each line also carries a contract-risk sub-score derived from the account renewal status. Features consider the plan, tenure and account attributes (tier, size, industry, health, renewal status).

💡 Everything is computed locally, with no AI cost. Without MRR movement history the projection stays flat, and with few resolved subscriptions the model is flagged as untrained, without making up numbers.

Forecast reliability (historical error and variance)

Ask "how reliable is my forecast?" or "am I hitting commit?". The historical error (MAPE) backtests the predictive model, comparing per period the forecast (sum of value × probability) against what was realized (won deals). The variance compares the declared forecast (Commit and Best case categories) with what was realized, showing whether you beat or fell short of what was committed.

💡 Because the system does not yet store a saved snapshot of each day’s forecast, the MAPE is computed over the whole history (an optimistic approximation) and the variance uses the record’s current category. Each answer carries a basis field explaining the limitation, so you read the number with the right context.

Product recommendation (best offer and cross-sell)

The Copilot also recommends WHICH product to offer each account, in the style "accounts that bought X also bought Y". It looks at your REAL sales history: for each customer, the products they already won (won opportunities and opportunity line items). From the co-occurrence between products, it suggests to the account the products that similar customers bought and that it does NOT yet have, from highest to lowest score (confidence times frequency). Each suggestion is classified as cross-sell (a product from another category), upsell (a category already bought, higher price) or next product, with the reason ("60% of accounts that bought A also bought B").

Ask "what should I offer this account?" for a specific account or "what are the best cross-sell opportunities?" to get the team queue, with the best offer per account.

💡 The recommendation comes from your own history. With few sales or no products bought together, the list comes back empty: the system does not invent a pair your data does not show.

Anomalies

The "Detected anomalies" block tracks weekly indicators (opportunities and leads created, deals won) and flags weeks that broke the pattern, for example a sharp drop in leads or a spike in wins. Detection uses a robust z-score (median + median absolute deviation), resistant to extreme values.

Beyond the live view, the scheduler runs detection automatically and NOTIFIES admins (bell notification) when a new anomaly shows up. Each anomaly is alerted only once: the same week of the same indicator never triggers a repeat alert. The "Anomaly history" on the page lists everything already detected and alerted.

In Settings > Organization you turn the alerts on or off, tune the sensitivity (z-score threshold; lower = more sensitive) and pick which indicators to monitor. The default is already on, with sensitivity 2.5 and all three indicators.

Smart routing

The "Smart routing" section ranks reps by their historical win rate (smoothed, so a small sample does not distort it) and by their current load. Use the ranking to direct new leads to whoever is most likely to convert, balancing the distribution.

💡 Everything is computed locally from your data. It uses no external AI providers and generates no cost.

Open this article inside the system →