Agentic AI for retail planning

What agentic AI means in retail

An AI agent is software that can decide and act, not only predict. It works in a loop: it senses what is happening by reading data from databases, APIs and live feeds; it thinks by running that data through models and rules; and it acts by triggering a business process, such as raising an alert, drafting an order or calling another system’s API. Agentic AI is the practice of building systems out of such agents, each with a goal, the data and tools it needs, and the constraints it must respect.

What makes a system agentic is not one clever model. It is the orchestration of several steps and tools toward a goal, with little manual handling in between. A demand forecast on its own is analytics. A system that forecasts, compares the forecast with the plan, works out the order that closes the gap, checks the order against budget and supplier rules, and sends it to the ERP is agentic.

That distinction matters for retail because most retailers already have plenty of prediction. They have forecasts, dashboards and weekly reports. What they lack is the step from a number on a screen to a decision made in time, written into the system that executes it, and recorded so someone can check it later. Agentic AI is aimed at that step.

Retail uses agents in many places. Customer-facing agents answer questions, recommend products and help with fit. Operational agents forecast demand, manage stock across stores and warehouses, and set prices. This guide covers the second group as it applies to merchandise planning, sometimes called retail planning AI: the decisions that set what a retailer buys, where it goes and what it sells for.

Why merchandise planning suits agentic AI

Merchandise planning has the shape of a problem agents handle well. The decisions recur every week. They draw on a lot of data: sales, stock, receipts, orders, prices, calendars and plans, by product, store, channel and week. They have hard constraints: budgets, margin floors, pack sizes, vendor minimums and space. And they depend on each other, so a change in one decision should flow into the next.

Today, most of that flow runs through people and spreadsheets. A planner re-forecasts the plan, a buyer reads the new open-to-buy, an allocator sees a broken size run in the weekly report, a pricing analyst picks the markdown. Each handoff takes days, and each team works from a different extract of the data. The result is decisions made late, on numbers that have already moved, with the reasons scattered across email and workbooks.

Agents change the timing. They can check every product, store and week overnight, find what moved, explain why, and have a proposed fix ready when the planner starts the day. They can pass a change from one decision to the next without waiting for a meeting. And because every step runs in software, every step can be recorded.

The planning decisions agents can run

Merchandise planning is a chain of decisions, from the season’s money down to the price of a single style in a single store cluster. Each has its own data, cadence and owner, and each can have its own specialist agent.

• Merchandise financial planning (MFP): the season’s targets for sales, margin, receipts and inventory, re-projected as actuals land. An agent watches actuals against plan, explains each variance and proposes a re-projected plan.
• Open to buy (OTB): the budget left to buy by department, channel and month. An agent rolls OTB forward from the latest demand and proposes moving budget when demand shifts or a cost changes.
• Assortment planning: which options each store cluster carries, and how deep. An agent ranks options by local demand, space, margin and exit risk, and proposes adds, keeps and drops.
• Allocation: packs, size curves and transfers, store by store. An agent watches stock and weeks of cover by size and proposes transfers or pack changes before a size run breaks.
• Replenishment: reordering continuing products as they sell. An agent drafts the daily order within pack, minimum-order and vendor rules.
• Pricing and markdown: price ladders inside price and margin limits, and how deep and how soon to mark down. An agent models elasticity by cluster and proposes the ladder with its expected margin.

The value comes from connecting them. An approved plan feeds open to buy. Open to buy limits what the assortment can buy. The approved range is what allocation plans from. Allocation and replenishment move the stock, and pricing clears what is left. When each decision has an agent and the agents share one data model, a change at the top of the chain carries through to the bottom instead of waiting for each team’s next review. The retail planning glossary defines the terms used across these decisions.

How an agentic planning system is built

Under the agents, an agentic planning system has the same few layers, whoever builds it.

1. A data foundation. Sales, inventory, orders, plans, product attributes and calendars from every source, cleaned and mapped to one model. Retail data sits in POS, e-commerce, ERP and planning systems, and joining it is the first and often the hardest job. Agents are only as good as the data they read.
2. Prediction. Time-series forecasting for demand, and classification, regression and clustering models for questions such as which stores behave alike or which products are at risk.
3. Prescription. Optimization that turns predictions into decisions under constraints: how many units to send to each store given capacity, which markdown clears the stock at the least margin cost, how to split a budget. Prediction says what will happen; optimization says what to do about it.
4. Business rules and guardrails. The limits a decision must respect: never price below cost, stay inside the margin floor, respect pack sizes and minimums. Domain experts set them, and the agents check against them before proposing anything.
5. Agents and orchestration. Specialist agents, each responsible for one decision, and a coordinating agent that routes work between them, resolves conflicts and keeps a record. Each agent has the context it needs, such as product data, store data and past cases.
6. Integration. Agents read from and write back to the systems that run the business through their APIs and event streams, so a decision becomes an order, a transfer or a price change in the system of record.

Building it as separate modules with clear interfaces keeps it maintainable. A forecasting module can be retrained or replaced without touching the replenishment logic, as long as the interface between them stays the same.

What changes for planning teams

Agentic AI does not remove the planner, the buyer or the allocator. It changes where their week goes. Today much of it goes to gathering numbers: exporting reports, rebuilding the same forecast in several workbooks, and reconciling versions before a meeting can start. With agents doing the checking overnight, the variance is already found and explained when the planner starts the day, and the time goes to deciding.

The skills that matter shift with it. Planners spend less time on spreadsheet mechanics and more on judgment: whether a proposal makes commercial sense, which constraint should bend, when to override and why. Their knowledge also becomes part of the system, as the rules, limits and approval paths the agents follow, so it is written down instead of living in one person’s head.

That only works if the teams who own the decisions help shape the system. Projects that bring merchandise planners, buyers, allocators and finance in from the start, alongside the data and engineering people, get rules that reflect how the business really runs, and users who trust the proposals because they helped define them. A feedback channel after launch, where users flag proposals that look wrong, keeps that trust as the system changes.

Keeping people in control

The question every planning leader asks first is who decides. Merchandise decisions carry real money, and a wrong markdown or an over-buy is expensive. A well-built agentic system answers that question with five rules.

• Explain every proposal. A recommendation should arrive with its evidence: what moved, why, the data it used and how confident the system is. Planners trust what they can check.
• Keep a person in the loop, especially at first. The owner of each decision reviews and approves before anything changes, and every override is recorded with its reason.
• Set limits and escalate exceptions. When a proposal would break a limit, or the agent meets a situation it was not built for, it should flag it to a person instead of acting.
• Record everything and make it reversible. Every proposal, approval and writeback should be traceable to its source data, and every change should have a way back.
• Widen autonomy gradually. As an agent proves reliable on a decision, the owner can let it act on its own inside limits they set, and review what it did afterwards.

Autonomy, in other words, is a setting that changes over time, not a property of the software. Teams usually start with agents that propose and people who approve, and move routine, low-risk decisions to autonomous handling once the track record is there.

How OODARIS applies agentic AI to retail planning

OODARIS is the Agentic Operating System for Retail Decisions, for fashion and footwear retailers. It watches the business overnight, explains what moved, proposes the fix, and writes the approved action back into your systems. Every night it checks plan, stock, sell-through and price across ten capabilities in plan, buy, flow and sell, from demand forecasting to order fulfillment. By morning the person who owns each decision has the cause and a proposed fix with its expected margin.

Twelve specialist agents watch the plan, buys, ranges, stock and prices, and each proposes a change with its evidence. The AI Orchestrator connects them: it routes each task to the right specialist, validates incoming data and records where every number came from, reviews each proposal before it reaches a person, and replans when a tool or an agent fails. OODIE is the agent your team talks to: ask it anything about OODARIS, run scenarios and simulations, or change a value, always with approval. The agents page lists every agent with what it watches and proposes.

Every decision moves through the same five stages: assemble context, scope and route work, generate a recommendation, review and approve, and write back and learn. Confidence thresholds, human overrides and approvals, and decision records travel with the decision. One trace ID links the source inputs, the agent and model runs, the approval, the writeback and the outcome, and each writeback has a rollback path.

How much OODARIS does is your call, in four levels:

• OODIE answers questions, runs scenarios and simulations, and changes a value when you ask. Every change to a value needs approval.
• Advisory shows what moved and why. Nothing is proposed or written.
• Prescriptive, the default, proposes the next best action with its reason and expected impact, and the decision’s owner approves before anything is written back.
• Autonomous acts on its own inside guardrails the decision’s owner sets. Every action is logged with its reason and can be reversed.

OODARIS reads from the systems a retailer already runs, including SAP ECC or S/4HANA, Microsoft D365, Oracle Retail, NetSuite, APTOS, POS, Shopify and data warehouses, and maps products, locations, calendars and measures from every source to one retail data model with lineage back to the source record. Each writeback uses the target system’s own APIs or event paths and is logged with the decision that caused it. How it works and the architecture page have the detail.

Merchandise financial planning and open-to-buy run in production today, and five capabilities are available to new clients. In production, OODARIS replaced a legacy planning system and more than 20 spreadsheets for MFP, and open-to-buy feeds the customer’s own reports and dashboards. The customers page walks through each decision.

How to start: one decision, one pilot

Agentic AI projects fail for familiar reasons: no clear objective, data that is not ready, a big-bang rollout that never reaches production, and users who were not involved and do not trust the output. The way around them is to start narrow and measure.

1. Pick one decision where margin leaks first, and say what success means in business terms, such as fewer broken size runs or shallower markdowns for the same sell-through.
2. Check the data. A pilot usually needs sales transaction history, access to POS or ERP data through an API or a file export, and inventory data, ideally at SKU level.
3. Agree the control. Decide in advance what the result will be compared against, such as a matched group of stores or categories, so the outcome can be read honestly.
4. Involve the people who own the decision. Planners, buyers and allocators know the constraints the agents must respect, and they are the ones who will approve the proposals.
5. Run a real cycle with named approvers, read the result against the control, and only then decide whether to scale or add the next decision.
6. Keep it current. Retrain models as seasons and behavior change, and review the guardrails as the agents prove themselves.

An OODARIS engagement follows the same path: foundations (pick the decision, connect the source data and agree the KPIs), configure (map your data, set the approval rules and tune OODARIS to your price ladders), a live cycle (planners run a real decision in OODARIS, with named approvers and controlled writeback), and measure (read the result against the control, then decide whether to scale). You can start with one decision; each capability you add after it reuses the same data and integrations.

Questions to ask before you buy

When you compare agentic AI for retail planning, these questions separate a working system from a demo:

• Which planning decisions does it run end to end, and which only produce a report?
• What does it write back, to which systems, and through which interfaces?
• Who approves each change, and can the approval rules follow our own approval model?
• Can every recommendation be traced to its source data, and every change rolled back?
• What does it need from us to start, and how soon is the first live decision?
• How will results be measured, and against what control?
• Does it replace the tools we no longer need, or sit beside them as one more system?

The research behind it

OODARIS founder Fatih Nayebi wrote Foundations of Agentic AI for Retail, a practical guide to implementing AI agents in retail operations. The book introduces the BDI (Belief-Desire-Intention) framework and the OODARIS loop for retail AI: Observe, Orient, Decide, Act, Reflect, Iterate, Serve. OODARIS turns that research into software.