What Is an AI Agent in CPG Analytics?
An AI agent in CPG analytics is an AI system that plans and executes multi-step analysis across syndicated data, retailer portals, and internal files, choosing measures, running queries, and assembling answers, rather than responding to a single prompt. That single sentence separates a real analytics agent from most of what gets sold under the label.
An agent is not a chatbot
The two companies behind the most widely used AI models publish compatible definitions. Anthropic separates workflows, where models and tools follow steps scripted in advance, from agents, which direct their own process and tool use; its practical summary is that agents are just models using tools in a loop, checking progress against real feedback from the environment at each step [1]. OpenAI defines agents as systems that independently accomplish tasks on your behalf, and is explicit that bolting a language model onto a fixed flow (simple chatbots, single-turn assistants) does not qualify [2]. Both labs also advise using the simplest solution that works: fixed workflows for well-defined repeating tasks, agents only where the path cannot be scripted in advance [1].
Put that in CPG terms. A chatbot bolted onto your SPINS extract maps one prompt to one query. Ask it a question that requires judgment (which measure, which time period, which product level, which market) and that single translation step fails quietly. An agent decomposes the question into steps, makes each choice explicitly, executes against the data, checks the result, and recovers when something looks wrong.
CPG analytics sits squarely in agent territory. The question stream is open-ended (every buyer email is different), the right analysis path cannot be hardcoded in advance, and the judgment calls change with every question. A category review is not a form to fill in. It is a small investigation, and investigations need an investigator.
The three parts of an agent
OpenAI describes an agent as three core components: a model, tools, and instructions [2]. Each maps onto CPG analytics work, and each has a distinct failure mode when it is missing.
| Component | In CPG analytics | Without it |
|---|---|---|
| Model (the brain) | Interprets the question, plans the analysis, decides what "velocity dipped" should mean for your category | A rigid workflow that breaks on the first unanticipated question |
| Tools (the hands) | Queries against SPINS, Circana, or Nielsen extracts, retailer portal exports, and internal files | A system that can only talk about your data |
| Instructions (the playbook) | Provider rulebooks for how each measure may be calculated and rolled up, plus brand context learned over time | Confident wrong numbers |
Anthropic stresses that agents must gain ground truth from the environment at each step, such as the actual result of a tool call [1]. In analytics, that means the agent looks at what the query returned before deciding what to do next. In Sous, the playbook's brand context half is what the system learns about your products, categories, retailers, and competitive set over time.
A worked example
A founder asks: "Why did my velocity dip in the Southeast last period?"
A chatbot translates that into one query, guesses at a velocity formula, and returns a number. Maybe the right one. You cannot tell.
An agent working inside a Sous workbook runs a plan:
- Resolve terms. "Velocity" is a blanket term covering several calculations, and this brand reads it as dollars per point of % ACV distribution; "the Southeast" is a specific market geography in this dataset; and "last period" is the most recent completed 4-week period, not the partial one in progress and not the same period a year ago.
- Decompose the question. Did velocity actually dip, or did distribution grow faster than sales?
- Query sales and TDP separately, checking each against the provider's aggregation rules.
- Compare against the prior period and the category.
- Write the answer: velocity fell 6% because 2 new items entered 40 more stores and have not built rate of sale yet, while base items held steady.
The queries stay attached to the workbook, so anyone can audit the math. Same question, but one system gives you a number and the other gives you an answer with its work shown.
The workbook also changes the economics. Because the analysis lives as a structured document (queries, charts, and narrative together) rather than a chat transcript, it can be saved as a template and re-run when the next period's data lands. The velocity investigation you commissioned in March quietly becomes the velocity monitor you own in April, and that difference compounds every period.
What an agent is not
The label is getting stretched, so it helps to name the impostors.
- A dashboard with a search box. Search over prebuilt charts still only answers the questions someone predicted at build time.
- A raw question-to-query box. A system that turns your sentence straight into a single database query skips the planning and verification loop that makes an answer trustworthy, and syndicated data punishes that harder than most (see "how accurate are AI agents on syndicated data?").
- A general-purpose AI model with a spreadsheet pasted in. No tools, no rulebook, no memory of your brand, and no way to check its own arithmetic against the source data. Paste 10,000 rows of Circana extract into a chat window and you get a system that will happily discuss your business while being unable to verify a single claim it makes about it.
- A scripted workflow wearing an agent costume. A process that always runs the same 12 steps against the same extract is what Anthropic calls a workflow: useful, predictable, and incapable of handling the question that was not anticipated when the steps were written [1]. Ask the unanticipated question and it either fails or silently answers a different question it does know how to answer.
The through-line is the loop: plan, act, check, adjust. A system that cannot run that loop on your syndicated data is a chat interface, whatever the marketing says. The loop creates a category of its own, agentic analytics, and for how it survives contact with SPINS, Circana, and Nielsen data specifically, see how an agent works with syndicated data.