Sizing and replenishment decisions sit at the heart of retail profitability. They are also among the most underinvested, under-governed and misunderstood areas in the business.

That gap, between what the data makes possible and what most organisations are actually doing, is wider than most planning teams would like to admit.

Recently, we brought together three practitioners — Frances Fountain (Fabletics), Matthew Pawson (Victoria Beckham Ltd) & Valentina Labate (PittaRosso) — for a direct conversation about the decisions that quietly move the P&L, and what it actually takes to get them right.

🎥 Watch the full discussion below & read on for the full write-up.

Note: We're aware that we're talking about the importance of technology…but we also know that sometimes it fails. Unfortunately, Matthew drops out for a few minutes due to connectivity, but he comes back quickly. Rest assured, this is still a conversation worth watching!

Replenishment Cannot Fix a Broken Buy

The first and perhaps most important truth of the conversation was that replenishment is a downstream mechanism. It can optimise. It can adjust. But it cannot fully recover from a size curve that was set wrong at the buy stage.

By the time you get to the demand signals and you see your sell through in your sales, you've already missed some of your full price selling opportunity — Matthew Pawson, Victoria Beckham Ltd

Lead times compound the problem; by the time poor size performance is visible in the data and a corrective order can be placed, the selling window has often passed. In fast fashion and e-commerce, the customer who couldn't find their size has already moved on to a competitor offering something similar.

For seasonal products, the window to react is even narrower. If the size curve is wrong and the product is already in stores, what follows is often just the redistribution of an already imperfect setup.

The implication is structural; planning and buying teams must work together earlier and more closely than most organisations currently manage. Merchandising teams own the size-level performance data and should be building the metrics and tools that inform buying decisions. Buyers bring product knowledge — silhouette nuance, fit behaviour by category, which styles run small.

When those two sources of intelligence are combined upstream, replenishment becomes a genuine optimisation tool rather than a damage-limitation exercise.

The practical reality is that orders are often placed before fits are even finalised, and guesswork fills the gap. That guesswork is expensive.

The Article-to-Size-Level Shift Changes Everything

One of the most significant operational shifts a planning team can make is moving from article-level performance tracking to size-level replenishment logic. It sounds incremental. The impact is not.

At article level, a product can appear healthy — strong sell-through, low markdowns, everything looking fine. Move to size level, and the picture changes entirely.

The product-level performance could look very strong. But underneath that, there can be a lot of inefficiency that you simply don't see. When you move to size-level replenishment, that becomes visible.
— Valentina Labate, PittaRosso

PittaRosso's experience after implementing AI-driven replenishment was illustrative. Their best-selling products had significant stock sitting in the central warehouse, yet store-level replenishment was still running too low. At first it felt like a model issue, but in reality, it exposed a structural misalignment between base-level demand and stock availability. Something invisible at article level.

Matthew framed the same dynamic from a merchandising perspective: a 70% sell-through on a SKU looks efficient until you realise that your core sizes have sold out and you are only left with fringe sizes. The product looks like a reasonable performer. In reality, it was a far better performer than the data suggests, constrained by availability rather than demand.

Sizing and replenishment efficiencies are all about being able to see where your ceiling is in terms of your sales levels. You can always see where your floor is. But what you want to be able to see is where you can go further.
— Matthew Pawson, Victoria Beckham Ltd

The ceiling is what most teams cannot currently see. Size-level logic is how you start to see it.

Which Signals Are Worth Trusting

More data does not automatically produce better decisions. A recurring theme across the conversation was the risk of over-complicating signal inputs, and the importance of knowing which signals to trust and which to discount.

The most reliable signals remain the most direct: stock-outs, sell-through, and weeks of coverage, because these are closest to real customer behaviour. They are not perfect — sell-through without accounting for size availability can mislead significantly — but they are stable and actionable.

There is often a tendency to over-complicate things. Adding more inputs like climate data, promotions or qualitative feedback can be helpful, but they are also much noisier than the direct signals. If your core signal isn't solid, adding more data doesn't really improve the decisions.
— Valentina Labate, PittaRosso

E-commerce data opens additional signal layers that remain underused in many organisations. Viewed availability at size level in platforms like Google Analytics shows how much customer interaction exists at a granular level. Out-of-stock sign-up volumes demonstrate latent demand that never converted. Back-in-stock email performance reveals how many customers were waiting, not absent. These signals capture missed demand, something that traditional size curve methodology, built on what sold rather than what could have sold, structurally cannot.

If you're already selling out on day one or day two of that product, your size curve could be completely different. We can only look at the facts we have rather than what the potential opportunity was.
– Frances Fountain, Fabletics

AI can begin to address this. By extrapolating from demand patterns across comparable products and factoring in periods of zero availability, it can approximate what a size curve would have looked like with full availability throughout. That estimated demand, not just observed sales, is the more accurate input for future buying decisions.

The counterbalancing caution is against over-granularisation. Drilling into one specific short in one specific colour can generate a size analysis that is technically rigorous and practically useless. A shorts customer is broadly the same size regardless of colour. The category-level size curve is usually more informative than the style-level one.

Knowing when to roll back up is as important as knowing when to go granular.

One Size Fits All Actively Costs You

Uniform replenishment logic applied across a diverse store estate, product mix and channel split is not a neutral choice. It is a choice with measurable costs.

Differentiation is not just a nice to have. It's actually where a lot of value is either captured or lost — Valentina Labate, PittaRosso

The logic differs materially across at least three dimensions:

  • Continuity versus seasonal are fundamentally different planning problems. Continuity products give teams time to react, which means bolder initial allocations carry less risk — stock can be redistributed, replenished and managed over a longer window. Seasonal products do not allow for that. The margin for error at the initial buy is narrow, so if the size curve is wrong and the product is already in stores, extracting and redistributing it in time to recover full-price sell-through is rarely realistic.
  • Store tier demands differentiated logic because demand is not uniform across a network. Flat allocations — sending the same size curve to every store for operational simplicity — consistently create shortages in high-demand locations and excess in low-demand ones. A store that outperforms in the first two weeks on a seasonal line is not necessarily the store that warrants additional replenishment depth if it is not a top-tier performer over the full season.
  • Channel introduces further complexity. Online availability should be maintained as close to 100% as possible, even when physical constraints mean some stores cannot receive a full size run. When a customer cannot find their size in store, there is a recovery path through digital. But when online availability fails too, that path closes.

Valentina shared a practical benchmark from PittaRosso's experience: keep the initial allocation to no more than 50% of total stock. Allocating 80% or more upfront creates a rigid distribution that requires inter-store transfers when demand diverges from the plan. Those transfers cost money, delay availability and often arrive after the peak selling window has passed.

Holding back meaningful replenishment stock and deploying it where demand actually materialises is more efficient, but only if the replenishment infrastructure — speed, frequency, algorithmic support — can support it.

The markdown dimension compounds the stakes further.

Getting your size curve right is the main reason to manage markdowns. It's so difficult when you've got the wrong sizes to sell them, even at a high discount. Your markdown sales tend to follow the same size curve as your full price sales — Frances Fountain, Fabletics

A wrong size curve does not just cost full-price revenue on the way in. It compounds into markdown inefficiency on the way out.

When AI Surfaces Inefficiencies, Who Owns the Problem?

AI-driven replenishment does not create problems. It makes previously invisible problems visible. That distinction matters, because the natural reaction to an AI surfacing an uncomfortable inefficiency is to treat it as a system failure rather than an organisational one.

AI didn't create a problem; it made it visible. And once you see it, you can start acting on it — Valentina Labate, PittaRosso

The harder question is who is responsible for acting on what becomes visible. In most organisations, the inefficiencies surfaced by size-level AI, for example misalignment between demand and stock, buying decisions that created structural imbalance, parameter settings that produced the wrong outcomes, do not belong to any single function. They sit across buying, planning, supply chain and sometimes finance.

Assigning ownership to one function does not work, because the problems are cross-functional, and so must be the governance.

What you need is shared ownership with planning, buying, supply chain all aligned on the same outcome. That usually requires stronger governance, clearer responsibilities, and shared KPIs — Valentina Labate, PittaRosso

Without shared KPIs and explicit accountability, AI generates insight that organisations acknowledge without acting on. That insight becomes a report. That report becomes a known issue. That known issue persists.

At some point, someone needs to take ownership for actually actioning on the data. Making sure your orders increase for the future, correcting that demand curve, working with finance to make sure you have the budget to make those sales — Frances Fountain, Fabletics

The Trust Problem Is Organisational, Not Technical

Getting teams to trust and adopt new replenishment logic is consistently cited as one of the hardest parts of any implementation.

I think the reluctance to use AI is really starting to shift. People are now on board, they just want to know that what they're doing has some credibility to it — Frances Fountain, Fabletics

Building that credibility requires more than a good model. It requires proof that the model is making decisions that a skilled planner would recognise as sound. And it requires showing, with enough time and enough controlled testing, that outcomes improve.

The governance around AI adoption is its own project. Super users need to be developed. Training needs to be structured and accessible enough to survive team turnover because new team members will inherit the tool without the institutional knowledge that shaped its initial configuration. Onboarding, that is not roadmapped properly, tends to produce inconsistent usage, which produces inconsistent outcomes, which gives sceptics the evidence they need to resist.

Leadership buy-in is equally non-negotiable. These tools require capital or operational expenditure, and the value they deliver is cumulative rather than immediate. Making the case to senior stakeholders and keeping that case credible as the tool evolves, requires a narrative that goes beyond month-one metrics.

It's really hard for team members, especially if you're allocators or administrative level, from breaking out of the habit of doing the day to day routine of looking at stock and moving it to store.
— Matthew Pawson, Victoria Beckham Ltd

That difficulty is not resistance to change, but a rational response of experienced professionals being asked to trust a system that is new and unfamiliar. Building that trust takes structured exposure, not just top-down instruction.

Translating Lost Sales Into Curve Adjustments (Without Overcorrecting)

Lost sales analysis is among the most seductive and most dangerous inputs to size curve adjustment. The danger is not in using it, but in over-indexing on it.

You can't just say we lost ten units, so we should buy ten more. You need to go deeper: how does adding those ten reflect on the other sizes, how does demand change, how does it play out across a chain of 300 stores?
— Valentina Labate, PittaRosso

The adjustment problem is not linear; adding units to an undersold size changes the expected sell-through of the whole curve, alters the markdown risk profile and shifts the stock position across a network. Getting those relationships right — particularly at scale — is not something that can be done reliably by eyeballing the numbers.

The approach that works is structured test-and-learn: make a small, controlled adjustment in a defined cluster or product group, track a small set of KPIs — sell-through, lost sales and weeks of coverage — and wait long enough to see a genuine signal before drawing conclusions.

The timing constraint is real. Too short an observation window and the data is noise. Too long and decisions slow to the point where the learning is no longer commercially relevant.

Matthew offered a practical calibration: even buying to 80% efficiency through known customer preferences is achievable without AI. The incremental gain from automation — potentially ten percentage points of improved efficiency — goes directly to full-price sell-through. That gain compounds across a season and across categories.

The point is not that human judgment is insufficient, but that automation makes the improvement faster, more consistent and scalable across a store estate too large for manual management.

The Execution Loop: Rhythm, Thresholds & Knowing When to Override

Governance at a strategic level — who owns the problem, which teams are aligned — is one thing. Governance at an operational level is another, and arguably harder. It is the day-to-day discipline of running the system: what the feedback rhythm looks like, how quickly teams act on what they see, and, most critically, when to trust the model and when the planner must step in.

These are not abstract questions. Without clear answers, even well-implemented AI replenishment tools drift. Parameters get adjusted reactively rather than deliberately. Interventions happen inconsistently. And because no one has defined what a legitimate override looks like versus an emotional one, the system gradually loses the confidence of the team running it.

The practical execution loop that emerged from the conversation has three components working in sequence: field feedback, parameter adjustment, and KPI readout, all with enough time between each step to distinguish signal from noise.

You need to give enough time, but not so much that you slow down decisions. Disciplined iteration, not trying to optimise everything at once, is the only way to validate whether your change has actually worked.
— Valentina Labate, PittaRosso

The KPI set that governs that loop should be tight — the same core three that should anchor any signal analysis: stock-outs, sell-through and weeks of coverage. Expanding that set before those signals are solid adds noise rather than clarity. The temptation to monitor everything is understandable, but it is also a reliable way to ensure nothing gets acted on with confidence.

On the question of when the planner should override the model, the conversation pointed toward a principle rather than a formula: the model earns trust through consistency, and the planner earns the right to intervene through judgment that the model cannot replicate. Qualitative context — an unexpected local event, a marketing push that wasn't in the data, a product quality issue that returns are only starting to reflect — is legitimately outside the model's reach. Those are valid override triggers.

What is not a valid trigger is discomfort. Planners who override because the model's recommendation feels wrong, without being able to articulate specifically why, are the single largest source of degraded model performance over time. That is not a criticism of planners, rather a description of what happens when override thresholds are not defined in advance.

The best output you get from these tools is where they're plugged into your data. But you also need rigidity — trust in both your sales and your stock data — before you can trust what the tool is telling you to do.
— Matthew Pawson, Victoria Beckham Ltd

The data integrity point is foundational. Replenishment logic built on unreliable stock positions — discrepancies between the system and physical reality — will produce recommendations that experienced planners will correctly distrust. In those cases, the problem is not the model itself, but the inputs. Fixing those inputs, and building the discipline to keep them clean, is a prerequisite to any meaningful improvement in replenishment accuracy.

PittaRosso's experience after moving to AI-driven replenishment illustrated what a working operational rhythm can unlock. Before the shift, manual allocation meant that more than 15–20% of total stock volume could not be allocated at the right time to the right locations. It was too granular, too slow, and always arriving after the demand peak had passed. With an algorithm running the allocation logic, the volume of pieces moving from warehouse to stores each week increased substantially, and the timing improved to match demand rather than chase it.

That improvement did not come from the technology alone, but from a team that had defined what the model was responsible for, what the planner was responsible for, and what a good intervention looked like.

The rhythm was deliberate. The thresholds were explicit. And the iteration was genuinely disciplined: one change, tracked against a small set of KPIs, with enough time to know if it worked before the next change was made.

Size Availability Is Still Being Overlooked as a Leading KPI

One consistent thread across the discussion was the gap between what most planning teams measure and what they should be measuring. Size availability — the proportion of time that core sizes are actually in stock when a customer is looking for them — remains a leading indicator that most organisations track inconsistently, if at all.

It's not maybe a traditional merchandising KPI, which is why it doesn't get as much attention or support as it should — Frances Fountain, Fabletics

The gap between sell-through and size availability can be enormous. A product with strong sell-through may have achieved those sales with consistent stock-outs in key sizes throughout the selling period. The sell-through figure suggests success. The size availability data suggests significant missed demand.

Tracking size availability properly, weighted by the sizes that matter most rather than as a simple average, exposes the ceiling that most product-level and even article-level data obscures. It also creates accountability for maintaining availability in real time rather than diagnosing absence after the season ends.

The Bottleneck Is Governance, Not Technology

The tools exist. The data exists. In most retail organisations of meaningful scale, the technology is no longer the limiting factor.

What limits progress is the organisational discipline to use the tools consistently, the governance structures to act on what they surface, and the cross-functional alignment to stop buying, planning and supply chain from optimising independently for their own metrics at the expense of shared outcomes.

Ultimately, replenishment that actually works is not a planning team problem. It is not a technology problem. It is an organisational design problem, one that most businesses are still tackling with the structures they built before the data existed to demand something better.

The opportunity is significant and the path to it is clear. The constraint is not knowledge, but the willingness to act on it.

We will be continuing this conversation, and many others like it, at Re:Plan USA (27-28 May) and Re:Plan Europe (13-14 October). Find out more and register your place.