PRINTABLE VERSION HERE  

AI capability is becoming more accessible, while the cost of production systems depends on the models, context, tools, loops and human review required by each workflow. For department stores, advantage will come less from access to a model than from control of the data, definitions, customer relationship and judgment that make its output useful.

Boston Consulting Group’s June 2026 analysis describes AI intelligence as increasingly scalable and accessible. Its July 2026 cost framework exposes the other side of that shift. Agentic systems may retrieve documents, make several model calls, use tools, carry context across steps and repeat actions before an output is accepted. Human review and correction add another layer of cost. Access to intelligence is getting cheaper while the cost of an accepted outcome can remain variable and difficult to predict.1

That tension matters because the cheapest model call does not necessarily produce the cheapest useful result. The relevant cost is what the store spends to reach an answer a buyer or a merchandiser can act on without sending it back. It also shifts the strategic decision from model access to organisational control. The store must decide which data enters the workflow, which commercial definitions it uses, how its output is evaluated, who can override it and whether the customer relationship remains with the store.

The argument for “homegrown” AI begins there. Homegrown AI can combine external models, private cloud and selected local infrastructure. What the department store retains is the business logic beneath those systems and the option to run workloads locally when economics, sensitivity or control justify it.

What homegrown AI means 

From token cost to business value 

For a department store, the bill that matters is the cost of an accepted outcome. BCG proposes measuring the economic return from an AI outcome against both the tokens consumed and the human effort required to initiate, review, correct, approve and operate the workflow.

Model price alone is therefore an incomplete guide. A more capable model that answers a buyer’s stock query correctly the first time can cost less than a cheaper one that takes three attempts and a merchandiser’s correction. For high-volume routine work, such as tagging product copy, the cheaper model usually wins. The comparison has to be made on the whole task, at the required quality, speed and level of risk.2

The three kinds of AI spend also behave differently on the profit-and-loss account. Building a reusable merchandising assistant is closer to an investment, running internal reports is an operating expense, and customer-facing inference behaves more like a cost of goods sold, rising with every conversation. Treating all three as one technology line hides where AI protects margin and where it quietly consumes it.

Reducing the time required to prepare a trading report does not automatically create value. The return appears only when the saved time changes a commercial decision: earlier action on inventory, more time with suppliers or better service. If the buyer still places the same order on the same day, the only thing that improved is the report.

Owning the intelligence layer 

BCG’s token-competition argument weakens the case for treating access to a foundation model, a general-purpose model supplied by an external provider, as a lasting source of advantage.

For a department store, the part a supplier cannot bring is the store’s own meaning. Available inventory has to account for merchandise moving between locations. Concession sales have to be attributed consistently across the store, department and brand. Returns have to be read differently when they signal poor fit, product quality or a commercial problem. Those definitions shape what buyers, merchandisers and store teams can do with an AI answer. A general model arrives without them, and no supplier can supply them. Paolo Pedersoli of Jakala made the same argument at the NRF Big Show 2026, in a session titled “AI without semantic capital is of no use”: as powerful models and generic data become widely accessible, the scarce asset is an organisation’s own codified definitions and decision logic, made machine-readable rather than left as tribal knowledge.3

BCG’s 10–20–70 framework, which Forbes reported the consultancy was urging in a January 2026 report on agentic AI, assigns around 10% of AI-transformation effort to algorithms, 20% to technology and data and 70% to people and processes4.

Model location and organisational control are separate decisions, but location still carries consequences. For stable, high-volume workloads on controlled data, local infrastructure can offer tighter custody of sensitive information and greater cost predictability. It also brings hardware utilisation, energy, engineering, security and replacement costs. External models avoid much of that infrastructure burden but introduce dependence on provider pricing, interfaces and governance.

The trade-off is workload-specific. High-volume, stable work on sensitive data, such as enriching product records or answering associate policy questions across the estate, is the kind that can justify local infrastructure. Occasional, bursty or experimental work rarely does. In both cases, the commercial definitions and evaluation standards should remain portable enough for the store to change the technology underneath them.

The operating model beneath the technology

When AI scales the wrong process

An IADS member provides a department-store example of AI built around explicit ownership of the underlying information. At a recent E-commerce Directors Meeting, the department store described moving ownership of data quality from IT to its e-commerce specialists, pairing AI enrichment tools with human validation: the enrichment tools produced usable attributes only once the people who trade the products owned the definitions behind each attribute. The reported improvement in search and conversion followed that change in ownership and human validation alongside AI enrichment.

BCG found that the organisations producing stronger returns were redesigning end-to-end processes, concentrating on a limited number of high-value applications and tracking whether the gains reached the profit-and-loss account. Adding AI to an unchanged process may increase speed while leaving conflicting definitions, unclear ownership or badly designed incentives untouched. The NRF Big Show 2026 department-store panel drew the same line between AI applied to legacy systems and AI-native design, citing a 90% reduction in the cost of product data management as the kind of gain that follows only when a process is rebuilt rather than accelerated.5

A report assembled from incompatible systems can still produce conflicting answers. Product information can be returned quickly while remaining incomplete or inconsistently governed. A workforce system can reduce scheduled labour against a target that damages service. In each case, execution improves before the organisation has agreed what a good result means.

Starbucks provides a separate warning from the wider retail sector. Reuters reported that the company rolled out automated inventory counting across more than 11,000 company-operated North American stores in September 2025 and discontinued it in May 2026 after persistent misidentification and missing items forced employees to check or repeat its work.6

Reuters’ January 2026 reporting places those failures inside a supply chain already affected by fragmented suppliers, outdated systems, forecasting problems and limited store storage. The counting errors created new verification work for employees, while the fragmented suppliers, forecasting problems and storage constraints remained untouched. Nine months of automation had changed the counting process without resolving the supply chain beneath it.

Where build-or-buy moves the work

In Korea, Lotte shows why build-or-buy is the wrong binary. The department store used Strategy’s commercial technology but deployed eight analytics agents on a governed semantic layer applying common definitions across departments. Strategy reports more than 9,500 AI-supported analyses and answers delivered up to ten times faster in six months.7

Lotte answered the technology question by buying and was left with the harder half. The definitions its eight agents run on had to be written, agreed and policed inside the business, and they now live on a supplier's platform, which is where portability has to be defended rather than assumed. Buying a model or platform does not remove the work of deciding what inventory, sales, customers and performance mean.

Department stores should vary a system’s authority according to the consequences of an error and the amount of human interpretation the task requires. Product and policy retrieval can tolerate more automation because the answer can be checked against an approved source. A buying recommendation that changes an order still needs the buyer who understands the category and trading context. A staffing recommendation needs a store manager able to judge the service consequence.8

Recruitment, promotion, discipline and worker monitoring carry still higher stakes. The EU AI Act, which IADS identified as a governance framework for retail in its 2024 analysis of AI accountability, reflects that higher threshold by classifying specified employment uses (recruitment, promotion or termination, certain task-allocation systems and worker monitoring) as high-risk, subject to the regulation’s conditions and exceptions. AI may organise evidence and identify inconsistencies, but accountability for the final decision must remain with qualified people.

At the NRF Big Show 2026, the IADS panel on what AI can and cannot do for department stores placed governance at leadership level rather than inside IT alone. Data quality and taxonomy were identified as foundational constraints, while the complexity of product assortments and databases was singled out as an opportunity for automation. The warning was equally clear: efficiency gains should not be converted into cuts to sales staff or store investment but redirected towards customer experience and differentiation. The 2024 IADS Exclusive Navigating the AI maze in retail beyond the black box argued for proportionate oversight concentrated on high-risk areas, and for AI that supports rather than replaces human judgment.9 The authority given to the system should fall as the commercial, human or legal consequence of being wrong rises.

The customer relationship is part of the infrastructure

Distribution without disintermediation

The internal intelligence layer is only one part of AI sovereignty. The other sits between the department store and the customer. At NRF 2026, futurist and author Jason del Rey described AI platforms moving beyond product research towards transactions, with conversational interfaces experimenting with completing purchases inside the chat rather than sending customers to the retailer’s own website. He treated the shift as under way rather than settled. His practical conclusion was narrower than the threat: experimenting on external platforms is worth doing so products remain discoverable, but the priority is making the store’s own digital experience at least as good as the assistant’s. He described abandoning a retailer’s own site and getting a faster, more accurate answer from ChatGPT.

IADS members had already identified the underlying issue at the 2025 E-commerce Directors Meeting. Across their different app and omnichannel strategies, participants agreed that owning the primary digital doorway mattered to protecting brand equity and margins against marketplace giants. NRF added another layer. IADS summarised REI chief executive Mary Beth Laughton’s approach as “selective openness”: participating in external AI platforms while deciding which content, data and expertise remain exclusive to the retailer’s own channels.

The completed order tells a department store what the customer bought. The conversation that preceded it can contain the occasion, budget, preferences, rejected alternatives and future intent. If an external agent keeps that context and sends the store only an order, the richer part of the customer relationship sits outside the business.

Walmart offers one response. The March 2026 Forbes analysis describes the company placing its own assistant inside external interfaces while retaining more of the customer journey and checkout. The relevant lesson for department stores is control over continuity: an outside platform can introduce the customer, while identity, consent, loyalty history and the ability to continue serving that customer remain with the store.10

Selective openness buys reach by surrendering some control over discovery. The strategic choice is what the department store is willing to distribute and what it still needs customers to return to its own channels to receive. Based on IADS’s e-commerce and NRF report, IADS’s position is that department stores should use external AI agents as distribution channels without allowing them to become the permanent owner of the customer relationship.

What a general platform cannot reproduce

Physical and digital relationships remain intertwined for younger consumers. Edelman Gen Z Lab’s 2024 Rarely Heard Voices research found the cohort divided between enthusiasm and resistance towards AI. Seven in ten Gen Zers reported fact-checking online information even while shopping, while 49% of older Gen Zers preferred making purchases in stores and 40% of 14- to 17-year-olds shopped equally online and in store.11

The research also describes a “trust loop” in which brand action, advocacy and continued purchasing reinforce one another. That relationship extends beyond a single conversion.

Product comparison is where a general AI platform is strong. City-level presence, an alterations counter and a relationship built over several buying seasons sit outside its reach, and as more product discovery moves into AI interfaces, those capabilities carry more of the reason a customer chooses the department store itself.

The workforce is part of the intelligence layer

AI does not create expertise

Much of the knowledge that makes a department store function is not captured in its databases. It sits in how buyers interpret the relationship between sales and returns, how store managers balance labour targets against service requirements, and how associates adjust their approach to different customers. That judgment accumulates across products, seasons, trading cycles and exceptions. Transactional data records what happened; the experience of the people around it often explains why it happened and whether the same response should be used again.

A controlled experiment at IG Group, a financial-services company, helps isolate the role of domain knowledge. The study involved 78 employees divided among experienced writers, marketing specialists with adjacent knowledge and technologists more distant from the task. AI helped the adjacent specialists close much of the gap with experienced writers, but it did not turn employees without the relevant domain knowledge into experts. The researchers described an “AI wall”: the further users were from the required knowledge, the less able they were to evaluate and improve the output.12

Although the setting was not retail, the mechanism is relevant to department stores: employees still need enough underlying knowledge to recognise missing context, weak reasoning and commercially inappropriate answers. AI can accelerate people who already possess enough domain knowledge to judge the result; without that foundation, it risks increasing dependence rather than competence. The IADS NRF 2026 report reached the same conclusion from a retail perspective: Luc Julia, former co-inventor of Siri and Chief Innovation Director at Renault, and Pascal Malfoy, CEO of ADEO (Leroy Merlin), argued that AI does not remove the need for professional expertise, and that the strongest results come from combining domain knowledge with intelligent tools.

When production becomes faster, the scarce capability shifts towards judgment: whether a markdown recommendation suits the season, what trading context the model was never given, and who answers for the decision if it is wrong. Experience cannot be produced at the speed of a model response.13

The IADS 2024 White Paper on middle management treats reverse mentoring as one tool among several, alongside traditional and peer mentoring. That combination is particularly relevant to AI. Younger employees can bring familiarity with emerging interfaces and experimentation; experienced colleagues bring organisational memory, customer understanding and the ability to recognise exceptions. The IG Group experiment suggests why both matter: fluency with the tool did not compensate for distance from the underlying domain.

When efficiency consumes the apprenticeship

The workforce risk extends beyond adoption. It concerns what employees no longer learn once AI begins performing the work. First drafts, routine comparisons, straightforward customer requests, and low-risk recommendations can look expendable. They are also how people develop judgment: by doing work repeatedly, having it corrected and seeing the consequences of low-risk mistakes. Automating those activities without redesigning learning improves output today while weakening the organisation’s future expertise. Keeping some formative work in human hands also has a cost: the department store pays people to perform work that a competitor may automate. That choice should therefore be deliberate. The organisation needs to know which tasks are still buying experience, not merely labour.

BCG’s June 2026 study When everyone uses AI, companies risk losing critical skills, a global survey of 70 C-suite leaders and senior executives, found that half were already observing de-skilling, while more than 60% expected it to become a material threat within three to five years. The capabilities considered most exposed included judgment and decision making, problem understanding and framing, creative thinking, analysis and causal reasoning, and solution generation and evaluation, the same capabilities employees need when evaluating AI-generated work.14

The risk is especially relevant to department stores because much of their management and service capability has traditionally been developed through progression. Employees acquire judgment through customer-facing work, operational responsibility, recurring trading cycles and exposure to exceptions before moving into broader roles. The IADS 2024 White Paper on middle management describes a path in which sales associates could move through store management and potentially into director roles, gaining broader customer, operational, and commercial responsibility along the way. Centralisation and repeated reductions in the in-store management layer have already damaged succession planning and narrowed that traditional route into management.

AI can narrow it further by removing the lower-risk tasks and recurring exceptions through which employees learn to interpret evidence, make decisions under uncertainty and understand the consequences of being wrong. The risk is less that AI erodes leadership skills directly than that future managers assume accountability for work they have had fewer opportunities to perform and judge.15 When formative tasks are automated without being replaced by structured learning, an already fragile route from execution to expertise becomes narrower still.

A department store cannot rely indefinitely on experienced buyers and store managers to check AI output if it has automated the work through which the next generation of them would have been trained.

The recent IADS Exclusive, The $10 trillion management problem, reached the same conclusion from a management perspective. Technology reaches the shop floor through managers, yet manager engagement has fallen sharply, even as organisations expect those managers to translate AI into new working practices.16

The IG Group experiment, BCG’s de-skilling findings and Gallup’s evidence on manager support point to the same requirement for AI training. Shop-floor employees need more than familiarity with an interface. They need enough product and operational knowledge to identify a plausible but wrong answer, escalate uncertainty and reject output when the commercial context does not fit. A human in the loop adds little protection when that person lacks the experience or standing to challenge the system.

What department stores should build first

Start with governed knowledge

At NRF 2026, IADS reported a consistent message across several AI sessions: begin with a concrete business problem, clear objectives and the data required to solve it. Luc Julia and Pascal Malfoy argued for specialised, domain-specific applications rather than broad AI deployments, while the department-store panel identified product assortments, databases, data quality and taxonomy as practical areas where the sector’s complexity can be reduced.

BCG’s workflow framework and NIST’s AI Risk Management Framework add the evaluation discipline. BCG begins with a defined outcome, while NIST requires organisations to establish intended purpose, responsibilities, risks, measurement and continuing oversight. Applied to a department store, these principles favour a bounded and verifiable first project rather than a broad autonomous system.17

An associate-facing product-and-policy assistant fits those conditions. Three things already on the record point to it: the NRF department-store panel named product assortments, databases, data quality and taxonomy as the sector's most automatable complexity, the same panel cited a 90% reduction in the cost of product data management, and the quoted IADS member has shown that a department store can take ownership of that data layer without outsourcing it. It can answer questions about products, services, returns, delivery, appointments and approved procedures. Its answers can be grounded in approved product, service and policy information, the source can be shown to the associate, and unresolved questions can be escalated.

Lowe’s provides an adjacent retail example. Its Mylow Companion was deployed across more than 1,700 stores to give associates access to product, inventory and project information. A department-store version would have to accommodate the additional complexity of concessions, services, appointments, returns and store-specific policies.18

Preparing that knowledge is where the pilot becomes difficult. The IADS member treated data quality as a question of commercial ownership rather than a purely technical clean-up exercise, moving responsibility from IT to its e-commerce specialists and retaining human validation. The same issue would arise in an associate assistant: POS, CRM, inventory and product systems may disagree over whether an item is available, which department or concession owns the sale, or how a return should be classified. Store policies can also differ by service, market and location. The department store therefore has to decide which definition takes precedence, who owns it and how quickly an outdated answer is removed. None of this is free. Reconciling definitions, keeping formative work in human hands and staffing real oversight are costs the store carries before any savings appears, and they compete for the same budget as stores, people and logistics.

NIST requires systems to be tested before and during operation and reassessed against their intended purpose. For a first department-store pilot, that means testing the same set of real associate questions across the viable local, cloud, hybrid and non-AI options. The comparison should measure answer quality, response time, cost per accepted answer, escalation and manual correction, while also recording employee rework and variation across stores and languages.

Measure rework, not activity

The pilot should also test whether experienced employees consider the answers commercially useful, whether newer employees can explain why an answer is correct and whether the assistant builds understanding rather than conceals gaps.

Oversight should follow the retail decision itself, applying the NIST requirements set out above. An associate needs to see the product or policy source behind an answer and know when to escalate it. A buyer reviewing an AI recommendation needs the sales, returns and stock evidence behind it. A store manager needs the authority to reject a staffing recommendation that meets a labour target but leaves the floor unable to serve customers. Decisions affecting recruitment, promotion or discipline require stronger human control again.

Scale only the question types the pilot answered reliably, and leave the rest with people. The test of the pilot is whether associates keep using it once the launch attention fades, whether the cost per accepted answer is one the business would pay at scale, and whether the people supervising it can explain why an answer was right.

Owning the intelligence layer buys one thing above all: the freedom to replace the model.

A department store that keeps its commercial definitions, evaluation standards, workflows and customer identity separate from a supplier can move as the technology moves. It can adopt a better model, bring a sensitive workload closer to home or renegotiate with a provider without reconstructing the business around that decision. Once those assets are embedded inside one platform, switching becomes slower, more expensive and more dependent on the supplier.

Homegrown AI should preserve that freedom. Where a workload runs can change with economics, sensitivity and control. The product history, customer logic, evaluation standards and operating definitions that give the business its meaning must remain portable and under the department store’s control. The model can change. The department store should not have to rebuild itself around every change.

Credits: IADS (Maya Sankoh)