• Written by Azhar Aliyar, Director

Most sales forecasts fail in the review, not in the model. The number is presented, somebody asks why it dropped for one region, nobody can answer in the meeting, and the question is taken away. By the time the answer comes back the decision has been made.

That failure mode is not about forecasting technique. It is about the distance between a forecast and the data underneath it.

The problem with an unquestionable number

A forecast is a claim about the future built from assumptions about the past. When those assumptions are buried in a spreadsheet that one person maintains, the forecast becomes something to accept or reject rather than something to improve.

Experienced sales leaders develop a rational scepticism about numbers they cannot interrogate. They discount the forecast and run on instinct, which is sometimes better and is never auditable. The organisation then has two forecasts, the official one and the real one, and plans against the wrong one.

Making the forecast open to questions is a structural fix. If the regional drop can be traced to two accounts and a delayed order in the space of a minute, the conversation moves from whether to trust the number to what to do about it.

The questions that actually get asked

In practice, forecast review questions cluster into a few shapes, and they are all comparative.

• What were the top selling products last month, and how did they compare with the month before.
• Which categories are trending against the same period last year, and by how much.
• Which accounts moved most between this forecast and the last one.
• Where is inventory out of step with the demand pattern we are now predicting.
• What did we assume about the promotional calendar, and has it changed.

None of these is analytically hard. All of them are tedious to assemble, which is why they get asked once a month instead of whenever they would be useful. This is the case generative AI for sales teams addresses directly.

Forecasting and inventory are one conversation

A demand forecast that does not reach the stock decision is an academic exercise. The sales team's view of what will sell and the operations team's view of what to hold are frequently maintained separately, reconciled in a monthly meeting, and out of step for the intervening four weeks.

When both sit behind the same question interface, the reconciliation becomes continuous. Asking where inventory is misaligned with the current demand signal is one question, not a meeting. The answer names products, not principles.

This is also where forecast error becomes visible in a way that motivates correction. A product consistently over forecast shows up as stock that will not move, and the pattern is easier to accept when it arrives as an observation rather than an accusation.

What to watch for

Grain mismatch

Sales plans at product family level, operations holds stock at SKU level, and finance reports by category. A system that silently converts between these will produce answers that are individually plausible and collectively inconsistent. The mapping has to be explicit, and where it is genuinely ambiguous the answer should say so.

Survivorship in the history

Historical sales record what was sold, not what was wanted. Periods when stock ran out look like periods of lower demand, and a forecast trained on them will under predict the next one. Anyone working with this data needs the stockout history alongside it, or the model faithfully learns the wrong lesson.

The confident wrong answer

A generative system will answer a question about a product with three months of history in the same tone as one with five years. The number of observations behind an answer matters and should be visible. A forecast for a product launched in June is a guess, and it should be presented as one.

Where to begin

Start with the review meeting you already hold. Write down the five questions that come up most often and currently get taken away as actions. Connect the sales history and inventory data those questions need, and answer them live in the next review.

If the answers match what the team believed, trust is established cheaply. If they do not, the disagreement is the most valuable thing the exercise will produce, because it has been sitting unexamined in the forecast for some time.

Who should be asking

There is a tempting mistake here, which is to give the capability to the analyst who already builds the forecast and consider the job done. That does make the analyst faster, and it leaves the structural problem exactly where it was.

The gain comes when the people who hold the commercial judgement can ask directly. A regional manager who can check their own numbers before a review arrives with a position rather than a question, and the review becomes a discussion between people who have both looked at the same data. That is a different meeting from one where a single analyst is the only person who has seen the detail.

It also distributes the error checking. An analyst reviewing forty products will not notice that one of them has an implausible history. The person who sells that product will notice immediately, because they know it launched in March and the chart shows two years.

Handling the exceptions honestly

Every sales history contains events that should not inform a forecast: the one-off bulk order, the recall, the quarter distorted by a system migration. If they are left in, the model reproduces them as if they were seasonality.

The practical approach is to keep a short, maintained list of these events and their dates, and to make the exclusion visible in the answer rather than silent. A forecast that says it has excluded a period is auditable. One that quietly smooths it is not, and the first person to notice the discrepancy will distrust everything else the system says.

To try this against your own sales data, schedule a demo.