A forecast built on salespeople's gut feeling, single pipeline stages and the optimism of the quarter rarely gets better by being called a forecast. Testing AI for sales forecasting can give a considerably more accurate picture of future revenue, but only if you see it as commercial change work, not as yet another feature to switch on in the CRM.
For B2B companies with longer deals, several decision-makers and international growth, the need is obvious. Leadership needs to be able to plan capacity, investments and growth without being surprised every quarter by deals that slip. The problem is that many start at the wrong end: they buy an AI solution before they have decided which forecasting decisions should actually get better.
Why test AI for sales forecasting?
A traditional sales forecast often answers the question: what do the salespeople think they will close? An AI-supported forecast can instead answer: what do the history, customer behavior and current deal activity say about the likelihood that the revenue will actually come in?
The difference is bigger than it sounds. AI can combine signals that normally stay invisible in a manual forecast: changes in meeting frequency, response times, the number of active contacts at the account, how long the deal has been in a stage, discount levels, engagement from decision-makers and similarity to earlier won or lost deals. It can also identify risk in a pipeline that looks healthy on the surface.
That does not mean the model knows more about the customer's internal politics than the responsible salesperson. Nor does it know that a new CFO has just frozen all purchasing, if that information is never documented. AI should therefore not replace sales leadership's judgment. It should reduce the dependence on it where data can actually give a better answer.
The biggest benefit rarely comes from a nicer percentage on a dashboard. It comes from better decisions: which deals need executive sponsorship, where salespeople should spend time, whether pipeline coverage is enough and when marketing needs to create demand to close an upcoming gap.
Start with the decision, not the model
A pilot should be limited to a concrete business problem. For one company it can be predicting the quarter's new sales with greater precision. For another it is about understanding which deals are likely to slip into the next period. For a third, the main question is how many qualified opportunities are needed to reach a certain revenue level in the US or a new European market.
Choose one primary outcome and a few metrics. Forecast accuracy is obvious, but it needs to be defined. Do you measure the deviation on total revenue, number of won deals, deal value or forecast per segment? A model can be reasonably accurate at the total level but miss badly in the enterprise segment, where a few large deals decide the result.
It is also worth deciding which comparison is fair. AI should not be set against a perfect, idealized manual process. Compare against your current forecast: how early do you detect risk, how much deviation do you have at period end and how much manager time goes into chasing status? If the pilot does not improve anything that affects decisions or ways of working, the technical quality of the model matters less.
Data quality decides whether the forecast becomes useful
CRM data is not objective just because it exists in a system. If salespeople routinely push out close dates without reason, leave the next step empty or use pipeline stages in different ways, the model learns your process flaws. The result can look advanced but give false security.
That does not mean you first have to carry out a year-long data cleanup project. That is often a convenient way to avoid starting. You do, however, need to identify which fields and behaviors are consistent enough for a first pilot. In many B2B organizations, history from won and lost deals, clear stage definitions, amounts, expected close dates, ownership and activity data go a long way.
Three questions usually reveal quickly whether the foundation holds:
- Do you have enough closed deals to find patterns within the segments you want to forecast?
- Are your sales processes comparable, or do you mix, for example, small transactions and strategic enterprise deals in the same model?
- Is it possible to tell the difference between actual customer activity and internal CRM administration?
Be careful with historical bias
AI is good at finding patterns in what has already happened. That is also its limitation. If you have previously prioritized inbound leads with high activity but missed slower strategic accounts, the model can reinforce that behavior. If salespeople have always put extra energy into deals already judged promising, activity can look like a cause of winning when it is in fact a result of the prioritization.
Someone with commercial experience therefore needs to review the model's conclusions. Not just a data analyst and not just a system administrator. The question is not only whether the model is statistically reasonable, but whether the recommendations work in your go-to-market model.
How to run a pilot you can trust
Let the existing forecast process keep running in parallel for a limited period. The AI forecast should not become the answer key from day one. Compare the predictions week by week, especially on deals the model flags as risky or more likely than the salesperson's own assessment.
Then review the deviations together with sales leadership and RevOps. When the model misses, what was missing in the data? When the salespeople miss, were there signals that were ignored? It is in that discussion that the pilot starts to create value. You refine both the model and the way of working around the pipeline.
A good pilot also has a clear owner. Not an IT-led innovation group reporting on technology adoption, but a commercial lead who can make decisions when the insight calls for action. If the model shows that late stages in the pipeline are systematically overestimated, that must have consequences for qualification, sales coaching and how you calculate capacity.
This is also why we do not treat AI as a standalone implementation. Purasu links the forecasting work to CRM structure, sales process, lead scoring and the actual revenue plan. A model that only reports risk but does not change priorities is an expensive report generator.
When AI is not the right first step
In some organizations the problem is not the forecasting method but the absence of a working sales process. If you cannot explain what it takes for a deal to move from discovery to evaluation, there is too little signal to model. If every team uses its own definitions of qualified pipeline, a shared forecast becomes misleading regardless of technology.
The same applies at very low transaction volume. A company that closes only a few, very large deals per year generally needs to combine simpler data models with structured account reviews and leadership's market assessment. AI can still help make risk and activity patterns visible, but it should not be allowed to give false precision at the decimal level.
It is not a defeat to start with process and data. It is often the fastest route to a forecast you can use.
Scale only once the way of working has changed
A successful pilot does not just prove that AI can calculate better. It proves that the organization can act on the insights. Before you scale, you should therefore see concrete signs: earlier risk identification, fewer late surprises, better pipeline hygiene and clearer prioritization from sales leadership and marketing.
After that you can broaden to more segments, include renewal revenue or use the same signals to steer lead scoring and account prioritization. But do not scale because you have a model. Scale when the model helps you allocate time, money and leadership focus better than before.
The right first question is therefore not which AI platform to choose. Ask which uncertainty in your revenue engine you want to reduce, and which decisions get better when the answer no longer rests on guesses.