Every forecasting vendor is an AI company this year. The category language has converged so completely that the marketing no longer distinguishes between products.
Two questions still do, and I would ask them of us as readily as of anyone else.
Question one: is my data secure, and are you using it to train your models?
Ask both halves, because they are separate questions and vendors sometimes answer only the first.
Security is table stakes and the answers are verifiable. Training use is the part that gets glossed. There is a meaningful difference between a model trained on your historical data to serve you, and your data contributing to a model that serves everyone including your competitors.
Neither arrangement is automatically wrong. What matters is that you know which one you are agreeing to, since your closed-won history, pricing behavior and win rates by competitor are among the more commercially sensitive datasets you own.
Get the answer in the contract rather than in a sales conversation.
Question two: do the results pull from our internal data systems, and is there traceability back to the raw data?
This is the sharper question, and it is where products separate.
The reason is practical rather than philosophical. If you build a board deck from a tool's output and cannot trace how each number was produced, validating those numbers takes as long as building the deck yourself. The tool has moved the work rather than removed it.
Traceability means that for any number in the output you can follow the path back to the records that produced it. Not a description of the methodology. The actual rows.
| What good looks like | What to watch for |
|---|---|
| Any figure drills through to source records | A methodology page instead of a drill-through |
| Metric definitions are explicit and fixed | The same question phrased twice returns two answers |
| Results pull from your systems | Results computed from an upload you cannot re-run |
| Predictions come from a trained model | Predictions are a stage weight with new packaging |
Three follow-ups worth asking
How long until the model is trained on our data? A fully trained model built on your own historical sales performance is a 4 to 6 week exercise. An answer of days suggests you are getting a generic model with your logo on it. An answer of many months suggests the implementation is a consulting project. How does the model handle our bad data? The right answer is not that they will clean it first. Everyone has bad data, and as long as it is consistent, accurate predictions are possible. A vendor who requires clean data before starting is describing a project you will never finish. See your data is not uniquely bad. When in the quarter does the forecast become accurate? Accuracy in the final week is not useful, because by then the quarter has happened. Ask for accuracy measured at day 1, day 30, day 60 and day 90 rather than a single headline figure. See why a last-week forecast is worthless.Why these questions work
They are hard to answer well without having built the thing. Security and training terms require a real legal position. Traceability requires a semantic layer between raw data and the output. Training time reveals whether the model is yours. Data tolerance reveals whether they have implemented for real companies. Timing reveals whether the product forecasts or reports.
A vendor who answers all five clearly has done the work. One who redirects to accuracy percentages is selling a claim rather than a capability, and accuracy percentages are the easiest number in this category to quote without context.
Frequently Asked Questions
What should I ask an AI forecasting vendor first?
Two questions. Is my data secure and are you using it to train your models. And are the results pulling from our internal data systems with traceability back to the raw data. Both have answers that are easy to verify and hard to fake.
Why does traceability matter more than accuracy claims?
Because an untraceable number cannot be defended, and forecasting numbers get challenged in exactly the settings where defending them matters. If validating the output takes as long as producing it manually, the tool has not saved anything.
Is a chat interface evidence of AI forecasting capability?
No. A conversational layer added to existing reports is a different product from one whose predictions are model-driven. Machine learning and optimization do the forecasting work, and those are not the same technology as a language model.
Why ask about model training time?
Because it reveals whether the model is actually yours. A fully trained model on your own historical sales performance is a 4 to 6 week exercise. An answer of days suggests a generic model with your logo on it, and many months suggests a consulting project.
What is the wrong answer about data quality?
That they will clean your data first. Everyone has bad data, and as long as it is consistent accurate predictions are possible. A vendor requiring clean data before starting is describing a project you will never finish.
See how ORM turns these insights into action
ORM builds custom revenue forecast models for B2B SaaS companies. Not dashboards. Prescriptive analytics that tell you what to do next.
Schedule a Demo