
During my years working across corporate functions and operations at Versace and Giorgio Armani, I became familiar with the friction that arises when technology is organised around individual departments without a systemic view of how they need to work together. An application can perform its specific function well while the process connecting product development, purchasing, production and sales still depends on people manually joining the pieces.
The fragmented set-up I am describing has a familiar division of responsibilities: commercial teams rely on the ERP, the product office on the PLM, purchasing on the MRP, e-commerce on the PIM and marketing on the DAM. MRP also supports part of the production-planning process, while much of the scheduling can still happen in Excel. These systems have different purposes, but their information needs to be connected. Too often, the connection is a makeshift spreadsheet assembled outside them: data is exported, matched and circulated rather than exchanged through maintained interfaces such as APIs, so the next team may be working from a snapshot that is already out of date.
For me, the underlying problem is organisational as well as architectural. If nobody designs and owns the cross-departmental process, buying a better application for each team leaves the gaps between them intact. My work on virtual showrooms, digital prototypes and traceability has reinforced the need to follow those dependencies beyond the boundaries of a single project.
That is the opportunity I see for an AI layer above enterprise systems in luxury fashion, provided it is built on reliable integration and shared responsibilities. The existing applications retain their authority; AI helps people interpret exceptions, understand their consequences and prepare a response across departments. Putting a conversational interface over disconnected Excel extracts would reproduce the same fragmentation, with more convincing answers.
The following hypothetical scenarios illustrate how this could work. Their quantities are illustrative, rather than measured results from implementations at either company.
Keeping the systems of record in charge
An enterprise resource planning system, or ERP, records orders, purchasing, inventory and financial transactions. Product lifecycle management, or PLM, manages product development and technical specifications. Product information management, or PIM, structures information for commercial channels, while digital asset management, or DAM, manages the images and other media associated with a product. Material requirements planning, or MRP, calculates material needs against demand and supply; capacity checks may require additional planning tools.
These functions deserve to retain their official records. Calling a system “legacy” does not mean that its business rules have become disposable.
Connecting them requires explicit relationships between product identifiers, material lots, orders and assets, together with agreed rules for who owns and updates each record. Where the installed systems support them, APIs and event-driven connections can carry changes without another manual export. An API alone does not guarantee real-time coordination: the interfaces, update timing and failure handling must be designed and monitored. Any remaining spreadsheet dependency needs to be visible, with a known owner and refresh time, rather than silently treated as current information.
The proposed AI layer uses those authorised connections to read records and documents, link them through verified identifiers and prepare explanations or actions. Calculations of availability, cost and feasibility remain with the relevant applications and deterministic planning engines. The integration layer translates between their data models; the language model does not get to improvise what an order status or reservation means. Microsoft documents this separation of meanings through its anti-corruption layer pattern, an integration principle rather than evidence that every legacy system supports the same workflow. [1]

A document index or cache can make retrieval easier, but it must remain a derived view with a known source and refresh policy. Otherwise, the company has quietly acquired another unofficial master database, this time with a conversational interface.
PLM: a supplier reply should change the plan, not just the inbox
Consider an overshirt whose fabric development is six days late. The approved technical pack, which specifies how the garment should be made, is at version seven. A supplier replies that the colour-test sample will be shipped on Friday. That is useful information, but it does not establish when the sample will arrive, whether its colour will be approved or when the production sample can proceed.
An AI assistant could connect the response to the correct style and approved document, translate it for the team and prepare a follow-up in the supplier’s language. It would show the remaining dependencies, so that “shipped” does not become “on schedule” merely because an email sounds reassuring. Sourcing approves the communication; the responsible planner approves changes to milestone dates.


The same approach could help quality teams investigate recurring sample defects by retrieving comparable reports and their evidence. It should present a pattern to examine, not diagnose a fabric problem from a persuasive summary.
The expected benefit is less copying and chasing, with more time to address the delay while alternatives remain available. The approved specification and the quality decision stay with the people and systems responsible for them.
PIM: product content should inherit evidence, not imagination
A product page has its own version of the coordination problem. Composition may sit in a technical pack, care instructions in another approved document and a certification in a supplier attachment. Someone has to assemble those facts, translate them and make them fit each channel’s requirements.
In the illustrative T-shirt case, the AI extracts a composition of 95% cotton and 5% elastane from an approved source and prepares English, German and French content using the same facts. Each field remains linked to its document and version. A person reviews the translations, sizing information and channel mapping before publication.

The more revealing decision concerns the claim “organic cotton”. If the certificate cannot be connected to the relevant material and product, the claim remains excluded. A model’s familiarity with a supplier cannot repair that missing relationship.

This would reduce repeated data entry and contradictory descriptions, while making unsupported claims easier to spot. Channel readiness still depends on that channel’s actual schema and validation. An internally complete listing is not necessarily accepted by a retailer, and sending the record is not the same as confirming its publication.
DAM: an approved image needs a reliable connection to the garment
I would extend the same logic to the DAM. In my earlier analysis of digital asset production in luxury fashion, the useful starting point was the whole process: sample availability affects the shooting schedule, studio work produces the image, approval determines whether it can be used, and product information determines where it belongs. [2]
An AI layer could flag a missing approved image before a launch, propose a link between an asset and the correct product variant, or retrieve an existing wholesale image for possible consumer-facing reuse. Those suggestions must be checked against product identifiers, the approval history, usage rights and channel guidelines. Two garments that look similar are not necessarily the same SKU, the stock-keeping unit used to identify the sellable item.

The opportunity is to remove avoidable handovers and duplicate production where reuse is genuinely appropriate. It should also expose discrepancies: if a product specification changes after photography, the asset may need review. Generating an attractive replacement image would not establish that it accurately represents the garment being sold.
DPP: a missing origin remains missing
The Digital Product Passport, or DPP, makes the connection between product information and supporting evidence especially important. It is not simply another legacy application: it is a product-information framework whose preparation depends on evidence held across the supply chain.
Consider a pair of jeans for which the garment maker, fabric lot and yarn supplier are documented, but the origin of the cotton is not. An AI assistant could show that gap, locate the relevant supplier contact and draft an evidence request. It could also flag an expiring certificate and identify which products require review. It should never infer the cotton’s origin from the supplier’s address.

The expected benefit is a dossier that can be reconstructed and challenged, rather than a folder that becomes more impressive as PDFs accumulate. Under the EU’s Ecodesign for Sustainable Products Regulation, exact textile requirements and their application depend on the relevant delegated rules. A completed internal checklist is neither a published passport nor proof of legal compliance. [3] [4]


That distinction matters when material allocations change. Linking a new lot to a garment should prompt a review of affected evidence and claims, not silently transfer the previous lot’s sustainability story.
ERP: a delayed purchase order is also a customer decision
The ERP example begins with a supplier confirmation that delays 180 jackets by seven days. Two customer orders depend on them, but only one delivery promise is at immediate risk. The difficult work is connecting the purchasing exception to the service commitment, then comparing responses that someone can actually authorise.
The AI could extract the proposed dates, identify the affected orders and prepare alternatives: wait for standard delivery, expedite 120 units, or use 60 uncommitted units for a partial delivery. Stock, transport feasibility and the financial implications must come from the relevant operational systems, not estimates invented by the model.

In the simulation, targeted express transport adds €480 and reduces contribution after direct service costs from €6,300 to €5,820. Those illustrative amounts make the trade-off visible; they do not demonstrate a return on investment. The buyer and spending approver decide whether protecting the delivery promise justifies that cost. Splitting a purchase-order line also requires a supported operation in the actual ERP.

Invoice, purchase-order and goods-receipt discrepancies fit the same approach. AI can explain a possible mismatch and retrieve its evidence, while deterministic matching and payment authorisation retain their existing controls. In either case, the gain is a better-prepared decision, not permission for the assistant to spend money.
MRP: fabric in the warehouse may already belong to another promise
This is the example that best captures the value of a systemic view. A warehouse holds 600 metres of approved fabric, all reserved for 300 jackets. The jackets cannot enter production because their zips are late. Meanwhile, a confirmed order for 200 overshirts uses the same fabric in its approved bill of materials, the BOM, and its other components are available.
The AI could identify the second order as a candidate for investigation. It cannot declare the fabric free. Physical stock and uncommitted availability are different quantities, and the original reservation represents a commitment.

Using the example’s assumption of two metres per garment, producing the overshirts would consume 400 metres. That leaves 200 metres for the jackets and creates a replenishment requirement of 400 metres before their complete production run can start. The shortage does not disappear because the first order is temporarily blocked.
MRP and, where necessary, a capacity-planning or optimisation engine would test the proposal against reservations, replenishment dates and production constraints. Quality must accept the replacement lot, procurement must validate its availability, and contractual requirements must permit the combination. A requirement for a single dye lot could rule the proposal out entirely. Promised replenishment becomes usable stock only after receipt and the necessary quality release.


This could reduce the time fabric spends immobilised without hiding the risk being transferred to the jacket order. The same logic applies to deadstock: material left after a season is a candidate for reuse only if available quantities, approved specifications and lot quality support it. Demand forecasts can inform the assessment, but they do not override reservations or make a capacity constraint disappear.
Systems thinking is what connects those decisions
Systems thinking means examining relationships, feedback and constraints rather than treating each activity as an isolated optimisation. Donella Meadows identified information flows, rules and system goals as places where an intervention can change how a system behaves. [5] That is a useful lens for this architecture, because the quality of the outcome depends on what happens between departments.
The shared process therefore needs an accountable owner and decisions agreed across business units. IT can implement the connections, but product, purchasing, production, commercial, e-commerce and marketing teams have to define which changes matter, who can approve them and what information the next team needs. Those agreements determine what the AI should connect and what it must leave to a specialist.
Reallocating fabric changes a production reservation. It can also change purchasing needs, customer delivery risk and the lot genealogy behind a product claim. The AI layer should help make those consequences visible. Each authoritative system must then record the changes relevant to its responsibility, with reconciliation confirming what actually happened.

People belong in that design as well. A planner can approve a production priority without having authority to approve a sustainability claim. A buyer’s approval does not substitute for a quality release. If faster responses in one department create more work or risk in another, the company has moved the bottleneck rather than improved the process.
That is why I would judge an AI initiative against complete order fulfilment and the reliability of the resulting product information, not the volume of emails it drafts. Local speed is useful only when it supports the shared outcome.
Making it work today without putting operations at risk
The tools for document interpretation and controlled access to enterprise services make this architecture worth testing today. I would start by mapping one process across departments, including the Excel files currently connecting its steps, then agree its ownership and establish reliable access to the relevant records. The difficult work remains specific to the company: accessible interfaces, shared identifiers, refresh timing and operations that the installed systems actually support. Preserving the legacy core does not mean that integration requires no engineering or change management.
I would begin with one product family and a read-only comparison of AI recommendations against actual planner decisions. The pilot would track the time between an exception and an accountable decision, fabric immobilisation in metre-days, avoidable rework and orders delivered completely against their original promise. A partial delivery should remain visible as partial. Benefits have to be observed against a baseline before anyone assigns a savings target.
Before enabling actions, approvals must bind a named person to the exact records, changes and costs reviewed. If a source version changes, the proposal needs another check. Permissions belong outside the model, enforced by narrowly scoped services. Supplier emails and attachments are data to inspect, including possible malicious instructions, not a route to bypass those controls. OWASP identifies both indirect prompt injection and excessive agent privileges as risks that require controls beyond a carefully worded prompt. [6] [7]
Execution needs the same discipline. Retrying a request must not duplicate its effect, and an update that succeeds in one system but fails in another must remain visibly incomplete. Approval, execution and reconciliation are separate states; a conversational confirmation cannot collapse them into one.
For me, the attraction of this approach is that it preserves the specialised functions a luxury business relies on while making their relationships usable across departments. Its systems retain the rules, its specialists retain judgement, and AI helps people act on a shared process. The test is whether a change can reach the teams it affects, with its implications and approvals intact, without someone rebuilding the connection in another spreadsheet.
References
- Microsoft Azure Architecture Center, Anti-Corruption Layer pattern
- Fabrizio Rinaldi, Digital asset production in luxury fashion
- European Commission, Textile Apparel and the Digital Product Passport
- Regulation EU 2024/1781, Ecodesign for Sustainable Products
- Donella Meadows, Leverage points: places to intervene in a system
- OWASP GenAI Security Project, Prompt Injection
- OWASP GenAI Security Project, Excessive Agency