Retail and brand teams are under unprecedented pressure today and the world is shifting faster than systems can adjust. Customer expectations reset in real time, tariffs and input costs are repricing entire categories overnight, and planning assumptions that held last quarter no longer apply.
In this environment, many executives are looking for one AI agent to read the market, interpret requests, pull the right data, apply the right business logic, forecast demand and generate decisions across the entire operation.
It sounds like the perfect answer to every challenge hitting retail right now, but it’s not. And the retailers deploying AI this way are quietly building systems designed to fail.
From Prompts to Agentic Workflows
AI, as most people understand it, is a single exchange—prompt in, answer out. Retail decisions, however, are never single exchanges, but chains of interdependent steps and moving parts. Companies make multiple seasonal buys each year, across every category they sell, and placing each one involves reading prior sell-through, checking open-to-buy budgets, applying margin targets, and committing to quantities across sizes and colorways.
A multi-agent approach to AI keeps those steps intact rather than collapsing them into a single prompt-and-response. One agent interprets the request. A second retrieves the relevant data. The next applies the policy or business logic. And another produces the output. Each agent passes a defined output to the next, making the process explicit, auditable and controllable.
The workflow is the same, but the structure supporting it finally matches the complexity of operations.
The One Agent Problem
Think of the steps involved in a product return. They include interpreting the request, matching it to an order, applying the correct policy and then generating a response. When a single AI agent handles all of it, those steps collapse into a single output.
But what happens if the request is misread at the start? Then the entire system is anchored to that mistake. A return is treated as a billing issue, the wrong policy is pulled, and the customer receives a response that sounds correct but definitely is not.
With a single agent, workflows tend to degrade in three predictable ways: errors compound because there’s no checkpoint between steps to catch them, transparency disappears because there’s no record of how the output was produced and flexibility suffers because every new task layers on to the same process. With a single agent, it’s much harder to identify exactly where the agent went wrong, and one mistake can easily cascade through the entire workflow.
The Fashion Forecasting Challenge
Fashion is an excellent use case for a multi-agent approach because it’s an industry built on future bets. Teams commit to sizes, colors, fabrics and quantities months in advance. Yet in 2023, the industry produced an estimated 2.5–5 billion items of excess stock equating to roughly $70–$140 billion in losses, indicating just how difficult it is to accurately gauge demand.
Improving these decisions requires multiple analyses, including reviewing past collections, identifying which attributes mattered, mapping those attributes to sell-through and comparing them to current demand signals.
One AI agent asked to “forecast demand” has to do all of that in one pass. But just as no retailer or brand executive would ask a planner to do trend analysis, historical reporting, demand planning and competitive research simultaneously, no retailer should expect a single agent to, either—at least not at the level of precision, craftsmanship and detail today’s consumers demand.
A multi-agent approach distributes the work, with the first agent scanning product images from prior seasons and labeling size, color, fabrication and print. The next takes those tags and translates them into structured data buyers can actually use. A third maps that data against sell-through rates, markdown cadence and regional performance. A fourth cross-references those performance patterns with current search trends, social signals and competitor assortments.
Each agent is responsible for a narrow task and each generates an output the next step can use. The result is not a single answer, but a structured view of the decision, enabling human teams to work through a level of complexity that would otherwise be unmanageable.
Start with the Workflow, Not the Agent
Most failures in AI deployments are not failures of the model itself, but failures at the boundaries between steps, so teams should build one agent for each task. Retail teams looking to build effective agentic systems should first analyze each component of the workflow, asking “Where does the work break into steps?” “Where do errors enter?” and “Where does a human need visibility or control?”
Those are the points where retailers should introduce agents. They should also ensure a clear output and handoff at each step, and build in explicit points where a human can review, override or redirect before the workflow continues.
Don’t Forget to Build in a Data Strategy
Retail and brand leaders should also have a clear data strategy for their agentic workflows, as siloed data is one of the main challenges to deploying AI effectively in retail. Companies need to ensure each individual AI agent generates data that the other agents in the system can use. Planning feeds buying, buying feeds merchandising, and so on through inventory, logistics, sales and service. Especially in retail, each agent needs to function as a link in a strong data chain.
To build strong agentic workflows, retailers and brands should start with a specific challenge, break it into its component tasks, create a focused agent for each task, and build in specific points where human teams can review, validate and overrule if needed. This approach reduces the chance of a single point of failure bringing down an entire system and ensures appropriate AI boundaries that keep human intelligence at the center of business decisions that matter.
The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.
