Agentic merchandising

Agentic merchandising uses goal-directed AI agents to continuously monitor retail signals, evaluate commercial trade-offs, and execute approved merchandising decisions across digital and physical storefronts. It covers assortment, pricing, promotions, inventory, search, browse, recommendations, supplier workflows, and in-store execution as one connected system rather than separate workstreams.

What separates it from earlier AI merchandising is the move from analysis to action. Rather than operating with unchecked autonomy, agents work within merchant-defined goals, margin thresholds, approval gates, and audit controls.

The scope is the retailer’s side of the counter. Agentic commerce handles the customer-facing side, from shopping agents through checkout and post-purchase. Agentic merchandising governs the products, offers, inventory, and visibility that customers encounter.

How agentic merchandising works

Agentic merchandising replaces periodic batch reviews with a continuous decision-and-action loop. Rather than waiting for weekly sales reports or manual spreadsheet audits, AI agents observe retail signals, evaluate trade-offs against business objectives, call underlying optimization tools, and either execute approved changes or request merchant sign-off.

1. Perceive signals. The system ingests live inputs from across the business: search queries and zero-result rates, point-of-sale transactions, inventory levels, competitor prices, promotion schedules, supplier lead times, weather forecasts, and shelf images.
2. Interpret context. Agents read raw data in context. A drop in regional conversion is cross-referenced against local warehouse inventory, competitor discounting, and delivery delays to identify the cause rather than the symptom.
3. Plan against objectives. Instead of optimizing one variable in isolation, agents weigh competing targets, such as protecting gross margin against accelerating sell-through to clear seasonal stock.
4. Call optimization tools. Deterministic services still do the math. Agents query demand forecasting models, price elasticity engines, and catalog enrichment pipelines, deciding which to invoke and what to do with the result. Agents coordinate analytical models; they do not replace them.
5. Act or escalate. When the decision falls within its permitted range, the agent executes through API connections to commerce platforms, pricing engines, and ERP systems. Anything above the threshold or outside policy routes to a category manager for explicit approval.
6. Measure and learn. Every adjustment is tracked against its pre-decision forecast, so the next decision is informed by how the last one performed.

Orchestrating cross-functional trade-offs

Retail decisions are interconnected. A pricing agent that matches a competitor without consulting supply data can trigger an immediate stockout.

An orchestration layer acts as the referee among specialized agents. Price agents monitor elasticity, margins, and competitor moves, adjusting pricing within defined floors. Inventory agents track lead times, capacity, and allocation across locations. Discovery agents adjust ranking, sorting, and recommendations against inventory depth and commercial priority. Promotion agents measure incrementality and guard against margin dilution. Supplier and store execution agents handle vendor terms and shelf-level tasks.

When goals conflict, the orchestration layer resolves the trade-off against category strategy. Constrained stock means suppressed promotional placement and protected price floors. Excess inventory nearing storage penalties means markdowns coordinated with raised search visibility, so the two levers pull together instead of against each other.

Agentic vs. traditional and other AI merchandising

Retailers have automated catalog and pricing decisions for decades. Agentic systems differ from earlier approaches in how they structure decision rights, system access, and feedback loops.

Comparing retail merchandising approaches

Manual and rule-based merchandising executes merchant heuristics and explicit if-then logic. It looks backward, working from batch sales data, and has no autonomy of its own. Access is limited to read-only reporting and manual updates through a user interface, and teams learn from results only in a manual post-campaign review. Merchants keep full control, since they execute every rule, sort order, and price change directly.

Predictive optimization forecasts demand and calculates price or stock recommendations mathematically. It looks forward across weeks or seasons and plays an assistive role, generating suggested values rather than acting on them. The models read from transactional data feeds and improve through periodic retraining against historical sales. Merchants control the outcome by approving or rejecting recommendations in batches.

Generative copilots draft copy, summarize supplier performance, and answer ad hoc questions. They work interactively, in the moment, and their autonomy is task-specific: they suggest text or produce drafts. They run read-only queries against connected documents and learn from user feedback within a session. The merchant reviews and edits whatever the copilot generates.

Agentic merchandising monitors operational signals continuously, models trade-offs across multiple variables, and executes approved adjustments. It operates proactively and around the clock, responding to intraday shifts. Its autonomy is bounded: it acts only within guardrails and margin floors. Unlike earlier models, it holds read and write access to pricing, catalog, and ERP systems, and it runs a closed feedback loop that compares actual impact against its pre-decision forecasts. Merchants move from executing decisions to setting strategy, defining boundaries, and approving exceptions.

Agentic commerce orchestrates end-to-end shopping journeys across consumer agents, checkout, and settlement. It runs in real time, driven by each shopping session, with bounded autonomy on the consumer side of the transaction. Access comes through protocols that connect to consumer wallets and checkout APIs, and the system measures success by tracking conversion, payment, and settlement. Control sits with consumer consent and the limits set by merchant policy.

Understanding the operational shift

Each of the first three approaches leaves the merchandiser doing the heavy lifting.

With traditional e-commerce merchandising tools, merchants configure ranking boosts, pin products to category pages, and schedule promotional markdowns by hand. That manual approach suits planned seasonal campaigns, but it breaks down when local stock depletes unexpectedly, or a competitor adjusts prices intraday. 

Predictive models provided sharper analytical inputs without changing who executed the work: an algorithm recommended a markdown, and a human still had to key that change into an ERP screen. Generative copilots made information retrieval conversational, but drafting text is not the same as operating a category. Copilots hold no commercial objectives, track no business outcomes, and take no actions.

Agentic merchandising shifts software from an advisor to an operator. If informed that inventory is aging in a regional distribution center, an agentic merchandising assistant drafts a markdown schedule, models the margin impact against contracted vendor funding, checks stock across nearby locations, and presents a coordinated execution plan for approval.

The fifth row represents a different operational scope rather than a subsequent maturity stage. Agentic commerce governs the buyer side of the transaction, coordinating autonomous shopping agents and agentic payments. The two systems meet where the shopper’s agent queries the retailer’s catalog, pricing, and availability. Agentic merchandising determines what that customer agent discovers.

Core use cases across the merchandising lifecycle

Merchandising decisions have historically run on separate operational cadences: category plans are set months ahead, prices change weekly or daily, search rules respond to query logs, and physical shelves catch up through periodic audits. Agentic merchandising creates a unified operational fabric where analytical and execution models share live telemetry, so changes in demand, inventory, or margins inform downstream adjustments rather than waiting for manual intervention.

1. Assortment and category planning

Category planning balances local demand against inventory holding costs and operational complexity. Planning agents evaluate competitor product assortments, regional point-of-sale patterns, category margin profiles, and unmet search demand. Instead of relying on manual spreadsheet reconciliation, agents model rationalization scenarios, forecast demand transfer across substitutes, and draft localized line recommendations for merchant review.

Merchandisers retain authority over brand strategy and commercial relationships:

• Selecting strategic vendor brands that anchor the category
• Negotiating exclusive vendor contracts and private-label introductions
• Approving seasonal line architecture and brand-halo investments

While planning agents model commercial scenarios, operationalizing them requires enterprise-grade orchestration. Architectural patterns demonstrated in multi-agent enterprise workflows show how specialized agents can coordinate analysis across corporate functions while maintaining full governance and auditability for human decision-makers.

Primary KPI: Gross margin dollars.

2. Pricing, promotions, and markdowns

In fast-moving retail categories, competitor price moves and localized demand shifts outpace manual review cycles. Pricing agents analyze price elasticity, inventory age, competitor price points, and contracted vendor funding to model margin-balanced price adjustments and draft stepped markdown schedules.

Merchandisers define the commercial guardrails and approve major exceptions:

• Setting non-negotiable gross margin floors by department
• Authorizing headline promotional campaigns and seasonal events
• Approving out-of-cadence markdown proposals on aging inventory

Connecting these models to an omnichannel pricing engine allows agents to push governed price updates simultaneously across online storefronts and electronic shelf labels while strictly enforcing merchant-defined boundary rules.

Primary KPI: Gross margin percentage.

3. Inventory, allocation, and replenishment

Stock allocation determines whether commercial demand converts into revenue or sits as idle working capital. Replenishment agents track point-of-sale velocity, supplier lead-time variance, open-to-buy budgets, and distribution center capacity. By anticipating stock depletion from forward demand signals rather than waiting for shelves to empty, agents propose stock rebalances and draft purchase orders.

Merchandisers govern the financial and operational parameters:

• Approving capital expenditure outlays for major inventory reorders
• Reallocating regional distribution center buffer capacities
• Authorizing revised supplier lead-time buffers and penalty terms

Deploying pre-built algorithms for inventory allocation optimization ensures inventory shifts toward projected regional demand before stockouts occur.

Primary KPI: Sell-through rate.

4. Search, browse, and catalog discovery

Discovery is where category strategies become visible to customers. Search and ranking agents continuously tune category grids and facet prioritization based on shopper queries, click behavior, margins, and available stock depth, preventing search engines from elevating depleted items simply because of past click volume.

Catalog completeness directly determines the ceiling of discovery performance:

• Missing attributes, inaccurate dimensions, or inconsistent brand spellings break search facets and derail algorithmic recommendations.
• Attribute enrichment within the merchandising experience platform fills catalog gaps using plain-language merchant instructions. One AI model extracts missing specifications from product descriptions and imagery, while an independent validation model checks the values, automatically publishing high-confidence updates and routing low-confidence exceptions to merchant review queues.
• Pairing automated merchandising tools with catalog optimization accelerates product onboarding, as demonstrated when apparel brand G-Star used AI catalog enrichment to reduce onboarding cycles from weeks to days.
• Integrating catalog intelligence with a dedicated Vertex AI search engine allows retailers to ground discovery models in structured product metadata while retaining merchant ranking controls.

Merchandisers control commercial presentation:

• Defining sponsored brand placement rules
• Establishing manual product boosts for marketing campaigns
• Overriding automated search facets for seasonal events

In a live implementation, European luxury department store Galeries Lafayette deployed AI search and merchandising across a catalog of more than 600,000 SKUs. The platform gave merchandising teams granular control over discovery rules and testing, driving revenue lift and improved catalog discovery during peak trading periods.

Primary KPI: Revenue per visitor.

5. Conversational commerce and sales enablement

When customers research high-consideration purchases or store associates guide clients on the retail floor, keyword search bars fail to capture nuanced requirements. In conversational merchandising, agents act as consultative product advisors, translating open-ended queries into precise product recommendations backed by real-time inventory and merchandising rules.

Conversational agents streamline discovery across client-facing touchpoints:

• Interpreting natural-language customer intent and matching requests to product specifications
• Guiding shoppers through complex trade-offs across style, compatibility, and price
• Retrieving warranty, financing, and policy terms to support associate-led selling

In physical showrooms, a leading specialty sleep retailer empowered over 6,000 retail sales associates with an AI sales assistant that consolidates fragmented product manuals, financing tables, and warranty documentation into an instant conversational copilot. For digital touchpoints, integrating a custom shopping agent with the Google customer engagement suite enables proactive customer interactions across messaging channels and digital storefronts.

Primary KPI: Conversion rate.

6. Supplier and procurement workflows

Supplier terms, vendor funding agreements, and fulfillment track records are frequently split across contracts, enterprise resource planning systems, and emails. Procurement agents aggregate supplier histories, monitor compliance against contractual service-level agreements, identify lead-time drift, and draft evidence-backed negotiation summaries.

Because supplier relationships require relationship-driven judgment, agents operate under strict supervisory boundaries:

• Auditing supplier fill rates against contractual delivery terms
• Flagging invoice line-item discrepancies and calculating chargebacks
• Preparing quarterly vendor scorecards with fulfillment variance metrics

Similar automated reconciliation powers an expense management agent, which reconciles transaction ledgers, flags contract billing variances, and prepares dispute documentation.

Merchandisers conduct strategic negotiations, approve vendor allowances, and sign commercial contracts.

Primary KPI: On-time in-full (OTIF) rate.

7. In-store and physical shelf execution

Merchandising plans created at headquarters fall flat if store shelves are disorganized, tagged incorrectly, or missing stock. Computer-vision agents audit physical store conditions by analyzing photos from store associates, mobile cameras, or automated shelf sensors, identifying execution gaps between the planned assortment and the store shelf.

Store execution agents flag physical merchandising discrepancies:

• Detecting out-of-stock items and misplaced facings from image feeds
• Identifying pricing mismatches between shelf tags and central databases
• Dispatching targeted restocking tasks to store associates’ handheld devices

Merchandisers resolve recurring space allocation constraints and supplier packaging defects.

In a large-scale consumer goods deployment, PepsiCo automated shelf intelligence using computer vision and optical character recognition. The system verifies planogram compliance, detects shelf availability gaps, and validates price-tag accuracy from submitted store photography, alerting field teams to fix shelf issues before peak shopping hours.

Primary KPI: On-shelf availability.

Benefits and how to measure them

Evaluating agentic merchandising means looking past model accuracy to how autonomous actions affect revenue, operating speed, and commercial safety.

Merchandising performance scorecard

Tier	Core metric	Measurement method	Target outcome
Commercial	Revenue per visitor	A/B testing agentic ranking against baseline grids	Higher session value without margin erosion
Commercial	GMROI	Inventory turns against realized gross margin.	More margin dollars per inventory dollar
Commercial	Full-price sell-through	Units sold before the first markdown event	Less reliance on reactive clearance
Operational	Merchant touch time	Hours spent on onboarding and facet curation	Time redirected from data entry to supplier strategy
Operational	Time-to-shelf	Delta from catalog ingestion to live indexing	New SKU discoverability cut from weeks to days
Governance	Merchant override rate	Share of recommendations edited or rejected	Early warning of model drift or bad inputs
Governance	Guardrail trips	Agent actions blocked by hard margin rules	No out-of-policy prices reaching production

Each tier answers a different question. Commercial metrics ask whether the business is better off, operational metrics ask whether the way work happens has changed, and governance metrics ask whether the agents can be trusted to keep operating.

Commercial and operational returns

• Revenue capture from dynamic discovery. Traditional search ranks on historical clicks, often promoting items with broken size runs or thin local availability. Discovery agents weigh live stock depth, conversion velocity, and margin to surface available alternatives instead. In a phased deployment across 600,000 SKUs, Galeries Lafayette recorded a 7% revenue increase and 8% higher average basket value, achieved by pairing AI ranking with merchant oversight and continuous experimentation.
• Working capital efficiency. Replenishment agents anticipate localized stockouts from forward demand signals rather than waiting for shelves to empty, while pricing agents model elasticity to recommend gradual markdowns early. Both protect GMROI and reduce end-of-season write-downs.
• Merchant operational capacity. Automating attribute extraction, brand normalization, and discrepancy detection frees category managers from repetitive catalog work, so their hours go to supplier negotiation and line architecture instead.

Governance and system safety

Agents acting on margin and brand visibility need their own reliability measured, along with their commercial output.

• Tracking decision drift. AI agent observability lets teams inspect execution paths, track cost, and catch declining confidence before flawed decisions reach shoppers. Golden datasets and manual log review stop keeping pace once agents run at volume.
• Monitoring merchant override rates. A sudden rise in rejections points to bad catalog inputs, uncalibrated elasticity thresholds, or supply conditions the agent has not encountered. The trend matters more than the absolute number.
• Enforcing deterministic guardrails. Minimum margin floors and capital outlay limits are programmatic boundaries rather than model behavior, which is what stops agent speed from outrunning commercial policy.

The merchandiser’s role in an agentic operating model

The most common failure in agentic merchandising is organizational: a team that never adjusts to working alongside agents, so they run in advisory mode indefinitely while merchants keep executing by hand. What changes in the nature of the judgment merchandisers apply, and the organizational structure must change with it.

Decisions merchants retain

• Setting category strategy, brand identity, and line architecture
• Managing supplier partnerships and negotiating commercial terms
• Defining the trade-off policy between sales volume and margin preservation
• Exercising creative judgment on seasonal storytelling and premium placements

Commercial accountability does not transfer with the work. Whoever owns the category still answers for its performance, regardless of how many of its decisions an agent proposed.

How daily responsibilities change

Day-to-day work moves toward supervisory orchestration. Merchandisers interrogate agent rationales, evaluate edge-case exceptions, and validate commercial outcomes.

Previously	In an agentic model
Keying markdowns into ERP screens	Choosing between modelled markdown scenarios
Rebuilding facet rules after a range change	Setting the commercial logic those rules follow
Chasing suppliers for missing attributes	Negotiating terms and category investment
Reconciling reports that disagree	Deciding escalated exceptions with the reasoning attached

Organizational requirements

Operating this way asks for structural clarity as much as new tooling:

• Establishing end-to-end decision owners for each category
• Instituting dedicated AI product ownership to bridge merchandising and engineering teams
• Establishing policy stewardship to maintain compliance and margin floors
• Updating review cadences to focus on exception queues rather than blanket reporting

The skills shift accordingly. The strongest category managers in this model are the ones who can explain why a recommendation is wrong, not the ones fastest at executing it. Building that capability is an operating model question rather than a tooling one, and it usually takes longer than the technical deployment.

Implementation, governance, and a practical starting path

Agentic merchandising programs typically stumble when early project scope outpaces organizational capacity. A pilot spanning every category and decision type produces enough exceptions to exhaust the team reviewing them, and the agents get switched off before they prove anything.

Defining a decision-rights framework

Governance starts by setting explicit autonomy boundaries before any agent touches a production system.

Autonomy level	Permitted agent action	Operational scope	Human oversight
Recommend	Generates proposals with reasoning attached	Complex markdown curves, seasonal line rationalization	Merchant reviews and executes manually.
Act with approval	Prepares complete transaction payloads	High-value purchase orders, out-of-band promotional shifts	The merchant signs off through an approval queue.
Act within bounds	Executes automatically within hard constraints	Routine facet re-ranking, high-confidence attribute tags	Programmatic margin floors; merchant audit logs
Human-only	Restricted entirely from algorithmic action	Vendor negotiations, brand partnership terms	Excluded from automated decision pipelines

Decision types move up this scale as evidence accumulates. Something starts as recommend-only and earns autonomy by sustaining a low override rate across significant transaction volumes over time.

A four-step adoption path

1. Select a focused, high-impact workflow. Choose one measurable process with clear P&L accountability: category-page re-ranking, markdown timing, or catalog attribute enrichment. Avoid assortment and supplier decisions at this stage, since their feedback loops run in months rather than days.
2. Build the governed data foundation. Connect real-time inventory, unified product specifications, margin thresholds, and point-of-sale telemetry through two-way APIs.
3. Validate in shadow mode. Run agents in parallel with existing teams, comparing what the agent would have done against what merchants actually did. This surfaces calibration problems while they are still cheap: thresholds set too tight, elasticity assumptions that do not hold locally, and catalog gaps distorting inputs. Log the reasoning paths behind each proposal, and test rollback procedures and circuit breakers here before anything depends on them.
4. Scale through staged autonomy. Move from recommendation to bounded execution only after decision quality, override rates, and safety thresholds hold across repeated business cycles. Expand one dimension at a time: a category, a decision type, or an authority level, but not several at once, because a problem emerging after three simultaneous changes takes longer to diagnose than the expansion saved.

Platform foundations

Merchandising agents can only act on what they can see and reach. Three foundations decide whether a pilot scales or stalls:

• Catalog completeness, since ranking and enrichment agents inherit the gaps
• Live inventory visibility, since pricing and allocation decisions made on stale stock positions are wrong before they execute
• Write access to execution systems, since an agent that can only recommend leaves merchants doing the work

Retailers standardizing on Google Cloud can shorten that groundwork with pre-composed AI search and merchandising. Where merchandising agents also need to coordinate with customer-facing systems, the GAIN Platform provides the shared catalog, orchestration, and guardrail layer both sides operate on.

Assessing readiness usually starts with one high-priority merchandising workflow, its data dependencies, and the commercial guardrails around it.

Share

Follow

Subscribe

Follow