What it actually does

Sales prediction does not tell you what will happen. It tells you which deals look most like the ones you have won before.

That is a narrower claim than the marketing, and it is still useful. A salesperson with forty open opportunities and time for fifteen calls has a real prioritisation problem. Ranking them by resemblance to past wins is a better starting point than working down the list by date.

The output is an ordering, not a forecast. Treat it that way and it helps. Treat it as a prediction and it will disappoint you.

What it learns from

The score comes from your own history — the deals you closed and the deals you lost.

It looks at patterns across things like:

• Deal size, and how that compares to what you usually win
• How long it has been in its current stage
• How much activity there has been, and how recently
• The customer’s industry, size and location
• Whether they have bought before, and what happened
• Which salesperson is on it
• Where the lead came from

None of this is magic. It is noticing that deals over a certain size from a certain source with no contact for three weeks tend not to close.

The requirement nobody mentions

You need enough history.

This is the single thing that determines whether sales prediction is useful or decorative in your business.

A few hundred closed deals is thin. Several thousand, across a couple of years, gives something to learn from. If you close twenty deals a year, a score is a guess with a confident interface — and it will look exactly as authoritative as a good one.

Two other data requirements matter as much:

You must record losses properly. A model trained only on wins cannot tell you what a loss looks like. If your team marks deals lost without a reason, or leaves them open forever, half the training data is missing.

Your pipeline must reflect reality. If deals sit in “Negotiation” for eight months because nobody moved them, the model learns that stage means nothing.

Where it genuinely helps

Three uses that work.

Ordering the day. Which five opportunities deserve a call this morning. This is the main one and it is worth having.

Spotting stalled deals. Opportunities whose score has dropped — activity has fallen off, or they have sat too long. These are usually deals people have quietly given up on without saying so.

Sanity-checking the forecast. When a salesperson says a deal will close this month and the score disagrees, that is a conversation worth having. Not an argument — a question.

Where it does not

Deciding which deals to abandon. A low score means “less like your past wins”, not “will not close”. Some of your best deals will be the ones that look unusual.

Replacing the pipeline review. A score does not know that the customer’s budget was frozen last week. Your salesperson does.

Forecasting revenue precisely. Aggregate scores give a rough direction. They do not give you a number to put in front of a board.

The failure mode to watch

This one is worth understanding, because it is subtle and self-reinforcing.

Salespeople stop working low-scored deals. Those deals then lose, because nobody worked them. The model sees they lost and becomes more confident that deals like that lose.

The score becomes true because people acted on it.

Two safeguards:

Keep working a sample of low-scored deals. Not all of them — some proportion, deliberately. It keeps the data honest and occasionally finds business you would have missed.

Check calibration periodically. Of the deals scored highly, how many actually closed? Of the low-scored ones you did work, how many closed? If those numbers do not match the score’s implication, it is not learning your business properly.

Getting the data right

Three habits, and they matter more than any setting.

Record losses with a reason. A short list — price, timing, competitor, no budget, no response. This is training data and it is also, separately, the most useful report in your CRM.

Keep stages accurate. Deals should move when something happens, not in a monthly tidy-up. A pipeline updated once a month teaches the model nothing about timing.

Close dead deals. An opportunity from eight months ago sitting in Negotiation is noise. Either work it or lose it.

How to introduce it

Run it silently first. Score the pipeline for a quarter without showing the team. Then compare: did the high-scored deals actually close more often? You now know whether it works in your business rather than in general.

Show it as a sort order, not a number. A percentage next to a customer’s name invites arguments and false confidence. “Suggested priority” is a more honest presentation of what it is.

Keep the pipeline review. The score orders the conversation. It does not replace it.

Watch what it does to behaviour. If salespeople start ignoring whole categories of deal, that is the failure mode above starting.

Measuring it

Before switching it on, write down what should improve:

Conversion rate. Are more of the deals worked being won?

Time to close. Is attention going to deals that were ready?

Deals that go quiet. Does the stalled-deal alert actually catch things earlier?

Without a baseline you cannot answer any of these, and you will keep the feature on because it feels modern rather than because it works.

The short version

Sales prediction is a prioritisation tool. Used to order a salesperson’s day, it earns its place.

It needs real volume of closed deals, honestly recorded losses, and a pipeline that reflects what is actually happening. Without those three it produces a number that looks authoritative and means very little.

And keep working some of the low-scored deals — otherwise the score becomes right by making itself right.