Executive Master in AI for Fashion · Module 3, Day 2 · 6 November 2026
An online order creates revenue potential. What happens next—fulfilment, delivery, customer care and a possible return—helps determine how much value the business retains.
During Day 2 of Module 3 of Digital Fashion Academy’s Executive Master in AI for Fashion, Marcello Valerio, Pini Usha and Michela Petrinca explored how teams can use data and AI to improve those decisions. The sessions connected returns management, demand forecasting and ecommerce operations, with a shared emphasis on turning analysis into practical action.
Marcello Valerio: predicting a return is only the beginning
“The important thing is what we do with them, right?”
— Marcello Valerio, discussing predicted returns
Marcello challenged participants to look beyond the prediction itself. Knowing that returns are coming can help a company prepare its warehouse and financial forecast. The larger opportunity lies in deciding how to support the customer, recover product value and improve the next purchase.
He structured the discussion around three stages: before, during and after the return.
Before a return, brands can analyse order patterns, product history and customer feedback to identify recurring problems. Written comments add detail that a standard reason such as “too small” cannot capture. Teams can use that feedback to improve product descriptions, fit guidance and the information available before purchase.
During the return, the customer’s context matters. Someone who ordered several sizes to try at home may need a different response from someone who bought a single garment in the wrong size. Marcello explored how relevant exchanges and product recommendations can help a customer find a better option when they want one.
After the return, the task becomes operational: assess the item’s condition, decide where it should go and choose an appropriate route back to sale. He discussed how AI-assisted inspection and classification can support this work, while recognising the difficulty of automating every aspect of checking a used garment.
Practical takeaway: close the feedback loop
Returns data should inform product, merchandising, customer service and logistics decisions. Marcello also stressed the need for clear responsibility across the full journey: when each team sees only one part of the cost, opportunities to improve the overall result can be missed.
A participant’s question about fit changes brought this into focus. Return reasons become more useful when teams compare them with subsequent behaviour: did the customer exchange for a larger size, and did that replacement solve the problem?
Pini Usha: choose the forecasting method that fits the data
“AI does not solve problems that do not have enough data.”
— Pini Usha, Buffers AI
Pini separated two decisions that are often discussed together: estimating demand and deciding how much stock to buy or move. The forecasting approach depends on the product, the available history and the lead time.
His session distinguished three situations:
- New products with no sales history: use category-level history, budgets and planning assumptions to build a buying plan.
- Products with sparse sales: aggregate information at a higher level, such as style or colour, then use historical size distributions to inform the detail.
- Products with sufficient history: compare suitable forecasting methods against observed results.
In the Buffers AI demonstration, Pini showed how competing models can be evaluated through backtesting: predicting a historical period and comparing those predictions with the sales that actually occurred. The process allows the system to select a method based on its performance for the series being assessed.
When participants asked whether machine learning necessarily improves on established statistical methods, Pini explained that it does not always win in the examples he presented. Data availability and the characteristics of the series influence which approach performs best.
Practical takeaway: test assumptions about allocation
Pini also challenged the assumption that every size curve follows a neat bell shape. Different product families can have different distributions. Those differences matter when allocating a new product to stores and adjusting replenishment as sales arrive.
For planners and merchandisers, the lesson was to connect the forecast with actual operational constraints—lead times, delivery schedules and stock policies—and test methods against a relevant baseline.
Michela Petrinca: connect AI to the operation behind the checkout
“It’s about connecting systems, connecting tools, connecting infrastructure”
— Michela Petrinca, Calicantus
Michela, Account & Digital Strategist at Calicantus, examined the work that takes place after a customer places an order. Her central concern was how ecommerce teams protect profitability while delivering the service customers expect.
A customer-facing assistant needs access to the operational information behind its answers. Michela discussed the connections between warehouse management, enterprise resource planning, order management, payments, customer service and the product catalogue.
She explored four stages of the post-checkout journey:
- Order validation: assess transactions and operational requirements, with human review where needed.
- Fulfilment: choose how to serve an order using the available inventory and delivery options.
- Delivery and tracking: detect problems and communicate proactively.
- Customer care and returns: capture what customers report and use it to improve the experience.
Her delivery example made the benefit tangible: identifying an anomaly and informing the customer can help teams address uncertainty before a “Where is my order?” enquiry arrives.
Practical takeaway: use customer feedback to improve product information
Emails, chats and return notes contain observations about fit, colour and expectations. Michela described a workflow that groups recurring observations, connects them to products and helps teams update catalogue information.
She also highlighted the economics of running AI services. Teams need to consider the volume of information supplied to a model, the responses it generates and the time customers wait. These costs belong in the business case alongside implementation and supervision.
Her review criteria connected service with commercial outcomes: net sales, return costs, acquisition cost and customer lifetime value, fulfilment speed and delivery reliability, and first-contact resolution alongside cost per ticket.
What connects the three lessons?
The sessions showed how planning, returns and customer care can inform one another.
A forecast guides stock decisions. Fulfilment determines how an order reaches the customer. Returns and support conversations reveal where product expectations or processes need attention. When teams connect those signals, they can make more informed decisions across the customer journey.
DFA’s synthesis: start with a specific operational decision, identify the evidence needed to improve it, and measure the result. AI can support that work through prediction, classification and automation; teams still need to define the objective, review exceptions and act on what they learn.
Thank you to our instructors and participants
Thank you to Marcello Valerio, Pini Usha of Buffers AI, and Michela Petrinca of Calicantus for sharing their methods, demonstrations and operational experience.
Thank you also to our participants for the questions on fit consistency, forecasting accuracy and the relationship between established expertise and new tools. Those discussions helped connect the sessions with the decisions fashion professionals make in their daily work.
Explore Module 3: AI for Sourcing, Production and Supply Chain.