Guide
This guide is for CEOs, sales directors and CFOs who build their budget on figures from the sales team and are let down by them too often. We show three ways of forecasting sales, a simple way of measuring forecast accuracy, and the conditions a company must meet before AI can genuinely help.
In most mid-sized companies, the sales forecast is put together like this: the sales director asks the salespeople how much they will close in the quarter, adds up the answers and applies their own adjustment. It is quick, but it depends on mood and the pressure to hit targets.
The second method is based on the sales pipeline in the CRM. The value of each opportunity is multiplied by the probability of winning at its stage, and the results are added up. The third method is an AI model that estimates the likelihood of winning each deal by itself, based on the company’s history.
The table below compares these methods. They are not rungs on a ladder that you have to climb one by one within a year. Rather, they are tools for different states of data and different scales of sales.
| Method | How it works | When it is enough | Main weakness |
|---|---|---|---|
| Salespeople’s commitments | Each salesperson gives an amount; the manager adds them up and adjusts | A small company with a few large deals per quarter | People’s optimism or caution, no way of checking |
| Probability-weighted pipeline | Opportunity value × win rate at that stage | A few dozen customers in progress, a well-organised CRM | The same percentage for every opportunity at a given stage |
| AI model in the CRM | An algorithm assesses each opportunity based on the company’s history | Many opportunities, complete data from several quarters | Without good data, it performs worse than an experienced manager |
Forecast accuracy is a figure that few companies track, although it can be calculated in a simple spreadsheet. The formula: forecast error = (forecast – result) ÷ result × 100%. A positive figure means the forecast was too high, and a negative one that it was too low.
It matters which point in time you take the forecast from. Near the end of the quarter, it is easy to be accurate, because most deals are already closed. So record the forecast on a fixed date, for example in the first week of the quarter, and compare it with the result.
Look at two figures. The average absolute error tells you how far off the forecast is. The average signed error tells you whether it is always wrong in the same direction.
In the hypothetical example in the table below, the average absolute error is (20 + 9.1 + 13 + 0) ÷ 4 = 10.5%. The average signed error is (20 – 9.1 + 13 + 0) ÷ 4 = 6%. So the forecast is consistently too high, and it is worth looking for the cause in how salespeople assess their chances.
| Quarter | Forecast | Result | Forecast error |
|---|---|---|---|
| Q1 | 2.4 | 2.0 | +20% |
| Q2 | 2.0 | 2.2 | –9.1% |
| Q3 | 2.6 | 2.3 | +13.0% |
| Q4 | 2.5 | 2.5 | 0% |
A weighted forecast is the simplest step towards figures that can be checked. Each pipeline stage is given a probability of winning, and the forecast is the sum of opportunity values multiplied by that probability.
The most common mistake is entering percentages from memory or from a textbook. Calculate them from your own history: of the opportunities that reached the “quotation sent” stage in the last four quarters, how many did the company actually win? The table below shows the calculation on hypothetical data.
The weighted total in this example is PLN 1,020,000, although the total value of opportunities in the pipeline is PLN 3 million. The difference shows how misleading it is to report the raw pipeline value to the management board without weights.
| Stage | Value of opportunities in the pipeline | Historical win rate | Weighted value |
|---|---|---|---|
| Needs analysis | PLN 1,200,000 | 10% | PLN 120,000 |
| Quotation sent | PLN 900,000 | 30% | PLN 270,000 |
| Negotiation | PLN 600,000 | 60% | PLN 360,000 |
| Verbal confirmation | PLN 300,000 | 90% | PLN 270,000 |
| Total | PLN 3,000,000 | – | PLN 1,020,000 |
An AI model learns from closed opportunities: won and lost. If salespeople update the CRM once a month or enter random dates on opportunities, the model will learn random patterns.
For AI forecasting in Dynamics 365 Sales, Microsoft requires more than 10 closed opportunities with all the fields completed. But that is only the technical minimum. For a model to capture seasonality, it needs data from at least two full years, because with one year it cannot tell a season from a coincidence.
The list below sets out the conditions worth meeting before you start talking about AI licences. Each of them also improves an ordinary weighted forecast.
An AI model will learn from your company’s history only what your salespeople have recorded properly in the CRM.
Most popular CRM systems now have AI-based forecasting features, usually in their higher licence tiers. They differ in the details, but do similar things: they assess the likelihood of winning each deal and predict the total at the end of the period.
In Dynamics 365 Sales, the AI forecast appears as a separate column next to the salespeople’s forecast. The system also shows the factors that most influenced the result. This lets the manager see why the model rates a particular deal lower than the salesperson does.
Salesforce and HubSpot offer similar features, and tools such as Gong analyse recordings of sales calls. Prices and plan contents change often, so before deciding, check the current price list and ask which plan includes the feature you need.
For the management board, Copilot in Power BI can also be useful: it answers questions about the data in a report, such as which large opportunities have not changed stage for two weeks. However, it requires paid Fabric or Power BI Premium capacity. Microsoft also notes that for some features, questions in languages other than English are not officially supported.
A common misconception is: “we sell to businesses, so we don’t process personal data”. That is not true. The name, email address and phone number of a customer’s contact person are personal data, and meeting notes often contain more.
Before switching on AI, check three things. First, the legal basis on which you process this data. Second, whether you have a data processing agreement (DPA) with your CRM provider, and in which region the data is held. Third, whether the provider uses your data to train models.
If the tool records and analyses sales calls, participants must be told before the call. This applies to customers too, not just your own employees.
Improving your forecast is more a matter of several quarters of consistency than of a single project. The steps below can be started in any company that has a CRM, even a simple one.
The order is what matters most. AI switched on straight away will only replicate the mess in your data. Only at step 4 does it have something to learn from.
If a company closes a few large contracts a year, the model has nothing to learn from. A single contract can then shift the quarter’s result more than all the others put together. It is better to forecast such deals separately, assessing each one individually.
AI will not help just after a change in the sales model either, for example after entering a new market or changing the price list. History then stops describing reality, and the model will keep repeating old patterns for a long time.
Finally, the forecast will not improve if salespeople are afraid to show their real chances. When a low forecast brings a reprimand, people inflate the figures or hide deals until the last moment. That is a problem of workplace culture, and no tool will solve it.
Related topics in the knowledge base
FAQ
The questions sales and finance directors ask most often when they start measuring forecast accuracy.
About this page
About the author
Kacper Włodarczyk
CEO of ALGORCOMP
Kacper specializes in implementing Microsoft 365 Copilot, Copilot Studio, Power Platform (Power Automate, Power Apps, SharePoint) and AI agents for mid-sized B2B companies in Poland. He leads dozens of projects covering AI strategy, Power Platform governance, document workflow automation and sales processes. In his writing he focuses on the practical side of AI adoption in organizations — from the first proof of concept to company-wide rollout — with particular attention to data security, compliance (GDPR, NIS2, the AI Act) and return on investment.
Meet the teamIn a free consultation, we will review your sales pipeline and CRM data and calculate the error of your recent forecasts. We will then tell you whether the next step should be tidying up your data or an AI model.
Featured
How AI reads requests for quotation and tender specifications, builds a list of requirements and a draft quote – and which decisions must stay with the salesperson.
Which B2B marketing tasks to hand over to AI first, what each one needs, how to measure the effect, and which laws to keep in mind when handling content and customer data.
How to organise the handling of sales enquiries in a mid-sized B2B company: stages, response times, owners, metrics and automations that stop enquiries getting lost along the way.
Author: Kacper Włodarczyk, CEO of ALGORCOMP