A model can reproduce the measurements you already have and still leave an important question open: what explains them? For a business considering AI-assisted analysis, a persuasive answer should make room for other explanations that remain plausible.
In a November 6 guest research post on Anthropic’s website, AI for Science grantees Jakob Macke, Daniel Gedon and Stefan Wahl discuss building, fitting and interpreting scientific models. They explain that different models can fit the same observations, and that considering several plausible models can reveal uncertainty a single answer conceals. They also caution that these AI systems have mostly been tested on rediscovering known models; evidence on problems without known answers remains limited. This is a research discussion, not an announcement of a business forecasting feature.
Ask which question the model answers
The following is our business analysis, using an analogy rather than claiming that the research establishes results for business forecasting. Before accepting an AI-assisted explanation, separate the question you want answered from the material you gave the system.
Imagine a hypothetical retailer whose weekly orders rise after it changes its delivery information. An analyst asks an AI assistant to explain the increase. The assistant produces a neat account: clearer delivery information made shoppers more confident. That could be a useful idea to investigate. The order history alone, however, might also be compatible with a promotion, a returning customer group or an unusual week.
Those are invented possibilities for this example. The purpose of naming them is to make the proposed explanation more testable. A coherent story should not quietly become proof that the business has found the cause.
Keep the alternatives visible
A practical review can ask the analyst to record the leading explanation alongside one or two serious alternatives. Each should say what evidence it uses, what it leaves unexplained and what would make the team reconsider it. The goal is a useful comparison, not a long list of remote possibilities.
In the retailer example, the team might discover that the delivery change and a promotion happened together. It can then acknowledge that the available history does not separate their effects. That is a more informative result than assigning all the improvement to the preferred explanation.
Ask the assistant to help organize the comparison, but have a person check whether the alternatives actually match the records. Adding the phrase “consider other explanations” to a prompt does not establish that the resulting analysis is complete or correct.
Choose a next observation that matters
Once competing explanations are visible, the next question is what information would distinguish them. The team could examine an appropriate comparison period, investigate which customer groups changed, or seek specialist help designing a suitable test. Which approach is appropriate depends on the question and the business conditions.
Decide in advance how the additional information could change the decision. If every possible result would lead the team to the same conclusion, the proposed investigation may not address the real uncertainty. It may simply add another attractive chart to an already settled opinion.
Keep a record of the decision that was made with the evidence available at the time. If later information changes the explanation, reviewers should be able to understand why the earlier interpretation seemed reasonable and what prompted the revision.
Make uncertainty useful to the decision
A team does not always need one settled explanation before taking a limited next step. Several plausible explanations may support the same cautious action. In other cases they imply very different commitments, making further investigation more valuable.
The useful output is therefore more than a fitted line or a confident paragraph. It is an account of the question, the plausible alternatives and the decision the evidence can support. AI can help prepare that account; the people making the business decision still need to understand where its certainty ends.
AI-generated conceptual editorial image; fictional analyst and abstract sketches, not the researchers or a demonstration of models fitted to the same data.