How To Give Your Merch Team Superpowers With Agentic AI
Agentic AI is moving from recommendation to execution, and the operators who get the transition right will pull ahead.
Every retailer I’ve worked with has at least one person who is brilliant at pricing. They know the product, they read the sales data, they have an instinct for when to mark down and by how much. On a single line, they’re better than any model. But across hundreds or thousands of lines, with data changing daily, no individual keeps up.
This is where agentic AI enters the conversation, and for me it’s the most significant shift in how retail operations will work over the next few years. We’ve spent the past decade getting comfortable with AI as a recommendation engine: here’s what the model thinks, now you decide. Agentic AI goes further, analysing, deciding, and executing within boundaries you set, while the human moves from approving each action to defining the rules the system operates within.
I think most boards are right to be cautious about this. Handing execution to an algorithm requires trust, and trust has to be earned through a deliberate progression. But the commercial prize for getting it right is substantial, and the cost of being slow is growing.
The pricing example
Pricing is a good place to see how this works in practice because the feedback loop is tight and the outcomes are measurable. An agent takes in sales velocity, stock position, margin data, and competitor pricing, runs analysis against defined objectives, and makes a recommendation. In the early stages a human reviews and approves each recommendation, and once confidence is established the agent executes autonomously within guardrails.
I’ve implemented a margin optimisation trial where we improved overall margin by 1%. The process was heavily human-in-the-loop, with extensive manual oversight at every stage. Now consider what happens when you automate and scale the same logic across a business turning over a billion pounds. A 1% margin improvement is £10 million to the bottom line, with no new stores, no restructuring, and no headcount changes. A number any board recognises.
The sequence matters. Start with full human oversight, measure the model’s recommendations against a control, and build confidence before handing over execution within defined boundaries.
I’ve used pricing as the example here because results show up quickly and are hard to argue with, but the same logic applies across a full retail operating model. Replenishment, store allocation, demand forecasting, promotional planning: these are all areas where the work involves absorbing large volumes of data, running analysis against defined objectives, and making frequent decisions at a granularity no human team sustains. The shift from recommendation to autonomous execution within guardrails works wherever the decision is repeatable, data-rich, and time-sensitive.
Getting the guardrails right
The guardrails matter more than the model. For a pricing agent, this means setting minimum and maximum thresholds on the things the business cares about: minimum margin levels, maximum markdown depth, markdown increments, and stock level boundaries for promotional periods.
What makes this interesting is the guardrails aren’t fixed, they change depending on the objective. If the goal is stock clearance ahead of a new season, the margin floor drops and the volume target goes up. If the goal is margin maximisation on a core range, the markdown rules tighten and availability becomes the priority. You need a hierarchy of importance across those levers, and the hierarchy shifts with the commercial context.
This is where a well-constructed agent earns its keep: adjusting across hundreds of lines simultaneously, responding to daily sales data, and optimising against the right objective with the right constraints. No merchandiser, regardless of how good they are, does this at scale.
The flip side is worth stating plainly. An agent operating at speed and scale with the wrong guardrails, or against a data error, creates damage at speed and scale too. An automated 40% markdown across a core range because of a stock file glitch is a material event, not an irritation. The guardrails need kill switches: hard limits the agent won’t breach regardless of what the data tells it, exception alerts for moves outside normal parameters, and a clear escalation path when the model encounters something it hasn’t seen before. The same discipline you’d apply to giving a new buyer delegated authority applies here, except the agent works faster and won’t flag its own uncertainty unless you build in the mechanism to do so.
The resistance is predictable (and mostly wrong)
When you talk to merchandising teams about this, the pushback is immediate. Handing over execution feels like an admission they aren’t capable, and for people who’ve built careers on their judgement and product knowledge, an agent doing the job feels like a threat to their professional identity.
I understand the reaction, but the argument falls apart when you consider volume. Optimising pricing on one product is fine. Doing it across 500 lines simultaneously, reacting to daily sell-through data on each, and rebalancing across competing objectives: no human team does this well at scale. BCG’s research supports the point: agentic systems accounted for 17% of total AI value in 2025 and are projected to reach 29% by 2028, with back-office operations and supply chain among the fastest adoption areas.[1]
The framing I’ve found works with teams is capacity. The agent handles the volume, the frequency, and the computational complexity, while the merchandiser defines the strategy, sets the guardrails, interprets the exceptions, and focuses on the work where human judgement genuinely adds value: range architecture, supplier negotiation, brand positioning. These are tasks requiring commercial instinct, relationship management, and contextual awareness no model replicates, and they’re the parts of the role most merchandisers built their careers to do.
The prerequisite nobody wants to talk about
If you’ve been following this series, you’ll recognise what comes next. An agent making pricing decisions across hundreds of lines needs accurate, timely data on stock positions, sales velocity, cost prices, and margin targets. Subtle optimisations on poor data compound errors rather than improve outcomes. As I put it to a board recently: putting lipstick on a pig.
I wrote about this earlier in the series. When I was leading a supply chain transformation, we took OTIF from 13% to 70% before any AI was involved, because the data had to be right first.[2] Gartner’s 2026 research echoes the point: 48% of retailers plan to deploy AI agents in the next twelve months, but many remain stuck in pilot mode because they haven’t aligned the data, processes, and teams needed for full deployment.[3]
Agentic AI amplifies whatever it finds, good data or bad. If your stock file is wrong, your pricing agent will optimise confidently against the wrong numbers.
What this means for your business
If you’re a CEO or COO looking at agentic AI, the progression I’d recommend is straightforward. Start with a contained use case where the data is clean and the feedback loop is short. Pricing, markdown optimisation, and replenishment are good candidates. Run it with full human oversight, measure outcomes against a control, and build confidence in the model’s recommendations before you hand over execution.
Define your guardrails in terms the finance team recognises: margin floors, stock turn targets, working capital limits. Make the hierarchy of objectives explicit so the agent knows whether it’s optimising for clearance or margin on any given range at any given time.
And be direct with the team about what changes. The merchandiser’s role shifts from executing hundreds of individual pricing decisions to setting strategy, defining constraints, and managing by exception. For most good merchandisers this is a better job, because the repetitive, high-volume analytical work is the part they don’t enjoy anyway.
The retailers who get this right won’t announce it with a press release. They’ll show it in their margin, their stock turn, and their markdown rate. The ones who wait will wonder why the gap is widening.
This is part of the series “Practical AI for Operators.” Previous articles covered where AI adds value, when not to use it, experimentation and scaling, why pilots fail at rollout, data quality, accountability in algorithmic decision-making, the build vs buy decision, and the economics of AI investment. Subscribe for the rest of the series.
References
[1]: Dataiku, “Supply Chain AI Trends 2026”: BCG reports agentic systems accounted for 17% of total AI value in 2025, projected to reach 29% by 2028. https://www.dataiku.com/stories/blog/supply-chain-ai-trends-2026
[2]: See earlier in this series: “Why Most Fashion Retailers Still Can’t Get AI to Work” https://open.substack.com/pub/scottmrobertson/p/the-supply-chain-data-problem-ai
[3]: Kore.ai, citing Gartner 2026 CIO Agenda for Retail: 48% of retailers plan to deploy AI agents in the next 12 months. https://www.kore.ai/blog/agentic-ai-in-retail