How Our Approach Is Improving Automated CPG Workflows

Consider a finance team reviewing a distributor deduction. The distributor has withheld money from a payment and provided a reason code. An AI system can read the remittance, identify the amount, categorize the charge, and produce a summary almost immediately.

So far, everything looks correct.

But determining whether the deduction is legitimate requires more than reading the charge. Was there an approved promotion? Did the agreement cover those specific products and dates? Does the amount reflect the agreed terms, or was an additional fee included?

The answer might be spread across a trade calendar, an email, an ERP record, and a distributor report. None of those sources necessarily uses the same identifiers or describes the transaction in the same way.

The same problem appears in demand planning. Imagine a brand sees declining sales at one distributor. At first glance, the data suggests weakening demand. But the distributor might have reduced inventory, the brand might have experienced a stockout, or the previous period might have included a promotion.

All three scenarios can produce similar numbers and require very different responses.

An experienced operator knows to consider these possibilities. They bring an understanding of the customer's history, the products involved, and how the business normally operates.

The difficulty isn't necessarily the complexity of the task. It's the amount of context required to complete it correctly.

And this isn't limited to large companies. At a small brand, much of that context might live in the founder's head. At an eight-figure business, it may be distributed across an ERP, retailer portals, spreadsheets, and several departments.

The information exists, but the relationships between it aren't always explicit.

A language model can recognize information and infer relationships between records. What it doesn't automatically have is an authoritative understanding of how those records relate within a particular business.

Take product identification. A product might have an internal SKU, a UPC, a retailer-assigned item number, and a distributor-specific code. Depending on the system, the record could describe an individual unit, a multipack, or a case. Those identifiers aren't interchangeable simply because the descriptions look similar.

The problem extends beyond products. A report labeled "sales" might represent shipments to a distributor, purchases by consumers, or invoices recorded in the ERP. Each is a valid measure of business activity, but they answer different questions.

Before an AI system can compare these records, it needs to establish what each one represents, how the records connect, and whether the comparison is appropriate.

This is where an otherwise capable model can produce an incorrect result.

Language models are good at identifying plausible relationships. But a plausible match isn't necessarily a verified one. Once an incorrect relationship is established, the system can build an entire analysis around it. The calculations may be correct, and the explanation may sound reasonable, yet the conclusion rests on the wrong premise.

This is also why asking the model to try again doesn't always solve the problem.

If an agreement is missing, additional reasoning cannot establish what that agreement authorized. If two product identifiers haven't been reliably mapped, another attempt may simply produce a different guess.

A model can reconsider its answer, but it cannot reliably recover business evidence that was never available to it.

Even after someone corrects the mistake, that correction needs to become part of the system's verified operating context. Otherwise, the next transaction can encounter the same ambiguity.

The issue is less about making a model think harder and more about giving it the information and verification process needed to reach the right conclusion.

At Natora, we're approaching this problem by working backward from the decisions operators make every day.

Rather than treating every document or data source independently, the goal is to establish a consistent understanding of how information relates across the business. Retailer and distributor records describe one part of an operation, while internal accounting systems, ERP data, spreadsheets, agreements, and email often describe the rest.

Connecting those sources is only the beginning. The more important work involves establishing relationships between products, customers, transactions, dates, units, and commercial terms so the system can interpret information within the correct business context.

That requires treating interpretation and verification as separate but complementary parts of the process. Language models are particularly useful for understanding unstructured information, recognizing relevant evidence, and explaining discrepancies. However, calculations, record matching, and other business rules often require explicit validation rather than relying on the model's interpretation alone.

The distinction matters because a system shouldn't treat every conclusion it generates as equally reliable. An answer supported by consistent records is fundamentally different from one built on incomplete information or assumptions.

Our approach is to make those differences visible by keeping the underlying evidence connected to the result. When information is missing, contradictory, or insufficient to support a decision, the system should recognize that limitation and surface it for review rather than produce an unsupported conclusion.

Human review is therefore part of the design, but it shouldn't be required for every task. Routine work with sufficient evidence can be handled with greater automation, while exceptions and consequential decisions receive more attention. Over time, verified corrections can help improve the mappings, rules, and test cases used across similar workflows.

The objective is to reduce the amount of work an operator needs to reconstruct, not simply generate an answer faster.

For operators, implementing AI starts with understanding the reasoning already taking place inside the business.

Much of that reasoning is rarely documented because experienced employees have developed it through repetition. They know which sources to trust, what inconsistencies to investigate, and which details can change the meaning of a transaction.

When businesses attempt to automate these workflows, they often discover that the challenge isn't a lack of data. It's that the knowledge connecting that data has never been formally established.

This is a useful way to evaluate where AI belongs in an organization. Rather than starting with the technology and looking for tasks it can perform, start with the decisions your team makes repeatedly and identify what allows someone to reach a reliable conclusion.

The exercise reveals which information needs to be connected, which relationships need to be verified, and where judgment is still necessary.

It also changes how success should be measured. Reading a report in five seconds instead of five minutes is valuable, but it says little about whether the system can finish the job. If someone still needs to spend twenty minutes reconstructing evidence and correcting the result, much of that benefit disappears.

A better measure is how much of the complete workflow can be handled reliably, how often exceptions require intervention, and whether the team reaches the correct outcome with less effort.

For a smaller brand, this might begin with a few recurring financial or operational questions and the files already used to answer them. For a larger business, the challenge may be connecting established systems and creating consistency across departments.

The scale changes, but the underlying principle remains the same: start with the business decision and work backward to the context needed to support it.

AI gives CPG businesses an opportunity to automate work that has historically depended on people moving between systems, reconstructing information, and applying knowledge that rarely exists in one place.

But the value doesn't come from placing a language model on top of those systems and expecting it to understand how the business works.

It comes from connecting the information, defining the relationships, verifying the results, and knowing when human judgment is necessary.

The model is an important part of that process. It just isn't the entire process.

The most useful AI system isn't one that always produces an answer. It's one that can establish when the evidence is sufficient to make a decision and when it isn't.