AI Can Read Your Amazon Data. That Doesn't Mean It Understands Your Business.

One ASIN accounted for nearly 85% of an account's month-over-month sales decline. The numbers made it look like a performance problem. It wasn't. A promotion had ended, and without that context, AI could describe the pattern but not explain the business.

At the item level, the numbers looked concerning. Sales were down roughly -$1,700. Sessions had fallen -31%, conversion was down -32%, and average selling price was up +14%.

Several explanations seemed plausible. Perhaps the listing had changed. Maybe reviews had deteriorated, inventory availability had become inconsistent, or the quality of traffic had shifted.

None of those explanations was correct.

The item had been promoted around Easter during the previous month. That promotion increased traffic and conversion while temporarily lowering the average selling price. When the promotion ended, performance returned to its normal range.

What looked like deterioration was largely the result of comparing a normal month with an unusually strong promotional period.

The data accurately described what changed. It could not, by itself, explain why.

The most important input may be missing

This is one of the central challenges of applying AI to business data.

A language model is very good at recognizing patterns and generating plausible explanations for them. If traffic and conversion both decline while price rises, it can identify several scenarios that commonly produce that combination.

What it cannot know is whether any of those scenarios actually occurred.

The promotion was not visible in the performance data. That information may have lived in a marketing calendar, a planning document, an email thread, or simply in the memory of the person running the account. Without that context, neither an AI system nor a human analyst could confidently explain the result.

AI makes this problem harder to recognize because it can express an unsupported interpretation with remarkable fluency. A tentative hypothesis can arrive sounding like a finding, and the distinction is easy to miss.

Fluency is not evidence.

Establish the baseline, then follow the evidence

I've developed a simple mental model for performance analysis that separates three questions often treated as one:

What should have happened?
What changed in the observed data?
Why did those drivers move?

Each question requires a different kind of evidence.

The forecast answers the first question. It creates a counterfactual, or an estimate of what the business should have delivered based on seasonality, underlying trends, calendar effects, and other known conditions.

Without that baseline, a change can appear meaningful when it is actually expected. Sales can decline month over month and still be on plan. They can also grow compared with the prior month while falling short of what the business should have delivered.

The forecast tells us whether the result was above or below expectations. It does not, on its own, identify the operational drivers behind that result.

That is where MX Intelligence comes in.

At its quantitative core, MX Intelligence uses a deterministic bridge to compare two actual periods. The bridge decomposes the observed change into contributions from traffic, conversion, price, and product mix. It then traces those contributions through the account hierarchy to identify the products, campaigns, or categories where the movements occurred.

This boundary is important. The bridge cannot mathematically decompose the difference between actual performance and a forecast. A forecast is a modeled expectation, not a second period of observed operating data.

Instead, the forecast establishes whether there is a meaningful gap to investigate. The bridge identifies the largest movements in the observed data. MX Intelligence then adds the context of "what we know," including promotions, inventory constraints, pricing decisions, calendar shifts, and other business events that do not fit neatly into a mathematical decomposition.

The bridge establishes contribution, not causation. MX Intelligence moves the analysis toward causation by combining those measured contributions with the forecast signal and documented business context. This allows us to connect a result against expectations with the most likely drivers in the underlying data while remaining clear about what was calculated and what was inferred.

In this example, the bridge surfaced the single ASIN responsible for most of the month-over-month decline. It then isolated the combination of lower traffic, lower conversion, and a higher average selling price. The promotional history explained why those metrics moved together, while the forecast provided a check on whether current performance was actually outside the expected range.

Each layer plays a distinct role. The forecast establishes what should have happened. The bridge shows what changed between two actual periods and where those movements occurred. Business context helps explain why.

A reliable AI analysis needs all three.

Give AI facts it can trace

The best use of AI in performance analysis is not to let it search a raw dataset for a story. It is to give it a structured body of evidence and ask it to interpret and communicate what that evidence supports.

That foundation should include:

Consistent metric definitions
A relevant forecast or baseline
A reconciled period-over-period decomposition
Known business events such as promotions, inventory constraints, pricing changes, and retail calendar shifts
A clear distinction between measured contributions and inferred explanations

With that structure in place, AI becomes genuinely useful. It can summarize a complex analysis, identify the movements that matter most, connect quantitative findings with documented business context, and present the result in language an operator or executive can quickly understand.

It can also help expose uncertainty.

If the bridge shows that conversion declined but no relevant operational or promotional context is available, the correct output is not a definitive diagnosis. It is a set of hypotheses, clearly labeled as such, along with the additional evidence needed to evaluate them.

That is a better analytical standard for humans as well as machines.

The real test of an AI analysis

When evaluating an AI system for commerce data, I would not begin by asking whether its answers sound intelligent. Most current models can produce a convincing narrative.

I would ask more fundamental questions:

What information was the model given?
Which statements are calculations, and which are interpretations?
Can each important claim be traced to a metric or documented business event?
Does the system acknowledge when the available evidence is insufficient?
Can an operator reproduce the analysis without relying on the prose?

These questions become even more important as AI systems move from reporting into action.

A weak explanation in a monthly review is inconvenient. The same explanation connected to automated budget, inventory, or pricing decisions can become expensive.

The ability to act does not make an agent well informed.

Make every conclusion traceable

AI adds the most value when the analytical foundation is already in place. It can accelerate investigation, translate complex findings into clear language, and help operators focus their attention. But speed is only useful when the conclusion can be verified.

A trustworthy analysis should allow us to work backward from a recommendation to the interpretation, the supporting business context, the measured contribution, and ultimately the source data. If that chain breaks, the conclusion should be treated as a hypothesis rather than a fact.

The goal is not to make AI sound like an experienced operator. It is to give experienced operators a faster way to reach conclusions they can inspect, challenge, and defend.

AI should not be the source of truth. It should be the interface to it.