AI sales forecast: 9 traps so far

Yesterday I could not tear myself away from the computer. When I lifted my head, it was half past eight in the evening. I had been sitting alone upstairs for about three hours.

I was teaching an AI agent to forecast orders.

The plan was simple.

Give the agent the data, the campaigns, the context, and let it forecast the number of orders for the next thirty days, every morning. And teach it to understand why the number lands where it lands on a given day.

I decided to build this more robustly than this particular output strictly requires. The reason is this. Once the agent understands what revenue is made of (what was caused by seasonality, by a short-term campaign, by an unexpected outage, and what effect overlapping campaigns had), a whole field of possibilities opens up for what else I can put it to work on.

One thing was clear from the start. Giving the agent a pile of numbers is not enough. For the result to be usable, it has to understand the connections between them. It has to be able to answer "if that seasonal campaign were not running, what would the chart look like?". It has to say "we expect this mid-month peak because of a retention campaign that ends in two weeks". It has to answer what-if questions and be able to attribute orders to specific circumstances or marketing activities.

The goal is clear.

Get the agent to a level where someone who makes order forecasts at the company says "fine, the machine takes this over, I will do something else". It is not easy.

THE NAIVE FIRST VERSION

If you tell an AI that you want to forecast orders and give it the sources of information for that, it will probably not go well. The AI will program something, something that looks expert, but when you look closely, you find that it does not work like this.

I tried it. I gave the AI the basic data source: how many orders came in the past. Access to the project tool that holds the marketing activities and campaigns. And from the marketing plan, it had the year-on-year revenue growth assumptions.

I gave it to the AI agent and it produced a formula:

baseline(2026-D) = actual(2025-D, weekday-aligned)
forecast(2026-D) = baseline × growth_target × campaign_multiplier

The multiplier (the number you multiply the baseline by to reflect the impact of a campaign) it pulled from history. A day at the peak of a particular campaign historically had some multiple of revenue compared to the state when no campaign was running. A different multiple for seasonal holidays, and so on.

At first sight the output looked quite good. Good enough to fine-tune. I started building a dashboard so I could visualise the result while tuning it.

When I asked it to explain the logic and visualised the data, a several-hour battle began.

ROUND 1: WHY WERE MY MULTIPLIERS LYING?

The model flagged one of the campaigns as the strongest. That was wrong.

I wrote to it: "That is completely off. This thematic week is one of the weaker ones. The other campaign running in parallel has a much bigger impact."

The problem was in the baseline (the reference state against which campaign impact is measured). The multiplier was computed as the ratio of (median of days when the campaign ran) to (median of other days). But "other days" included other parallel campaigns. The baseline was artificially inflated. The attribution of the revenue increase to each campaign was distorted in both directions: some campaigns overstated, others understated, depending on which other campaigns were running, either at the same time or in the days before or after them. In periods of overlapping campaigns, which is most of the year, the attribution made no sense at all.

After I objected, the agent rewrote the baseline definition to "median of days when no push campaign was running". But the result was not suitable as a starting point for analysis. There were few clean days. For some periods, campaigns and weekdays there were not even five examples. Campaigns overlap almost continuously.

ROUND 2: WHY IS AD ATTRIBUTION ONLY THE TIP OF THE ICEBERG?

Then came the attempt to add more context. The measurable impact of a campaign through ad attribution (assigning orders to a specific ad), conversions from the ad platforms.

Once again the agent could not interpret it correctly.

I wrote to it: "But you did not account for the consent rate. How many people refuse cookies."

Only a portion of orders gets matched through the ad platforms. The rest comes from non-consent customers who refused cookies: they do not show up in the ad platforms, but they do in the total number of orders. The agent knew about this gap, but did not include it in the prediction method.

After recalculation, the campaign numbers rose to more realistic values.

And straight away we ran into something else.

ROUND 3: HOW DOES ON-SITE COMMUNICATION CHANGE CONVERSION RATES?

I wrote to it: "A campaign does not have impact only through ads. When it is communicated on the website, the conversion rate rises for everyone who arrives, not just for those who clicked through from ads. Also for those who came organically, from Google, direct (visitors who typed the address themselves), and so on."