Retail: the real ROI of AI isn’t decided by which model you choose, but by how you govern it

The question that has occupied retail in recent years was a simple one: what can AI actually do? Retailers have experimented, launched internal assistants, automated content creation, and tested new use cases in customer service, merchandising and product data management. These initiatives have proven that the technology works. What they have really done, though, is open up a far more demanding question: at what cost does it work once deployed at scale?

A recent exercise illustrates the problem well. Drawing up a snapshot comparison of pricing across the leading generative AI providers, in order to inform architecture decisions, seems like sound practice. It is also an exercise with a very short shelf life. Within the space of a few weeks, three of the four providers under review moved to a new model generation in the mid-tier segment — the segment that accounts for the bulk of production usage. Anthropic replaced its Sonnet model with a newer, cheaper version. OpenAI migrated its range to a new model family. Google has scheduled the end of life of its reference Flash model for mid-October. Only Mistral held steady in this segment.

This isn’t a technical footnote. It is real-time proof that the real question isn’t which AI model to choose today, but accepting that the choice will need revisiting in the months ahead. The real question then becomes: which solution allows you to switch models quickly, without rebuilding everything each time? An agile architecture lets you follow the market in both directions — moving to a more capable model as soon as one appears, or to a cheaper one as soon as a price drop justifies it. It’s this capacity for rapid adjustment, more than the initial model choice, that determines the real profitability of an AI architecture over time.

Moving to industrialisation changes the economics

During the experimentation phase, costs seem secondary: a handful of users, limited volumes, a narrow scope. At scale, cost centres multiply — model calls and token consumption, infrastructure, integration with existing tools, data preparation, oversight, security, compliance, workflow maintenance, and a growing number of vendors.

The true cost of AI, then, is not simply the model’s price tag. But the price gaps between models remain a governance issue in their own right, precisely because they shift so quickly. Pricing observed in August 2026, as an order of magnitude — likely to have moved on by the time this piece is read: Mistral charges $0.15 per million input tokens and $0.60 per million output tokens for its Small 4 model, against $1.50 and $7.50 for Medium 3.5. Anthropic prices Sonnet 5 at $2 and $10 — a promotional rate that has become permanent, after a planned September increase was cancelled. Google’s latest-generation Flash model sits at around $1.50 and $7.50, with even cheaper entry-level tiers expected to follow. These figures, accurate at the time of writing, will very likely already be out of date by the time this piece is read — which is precisely the point: choosing a model is an economic decision, but one that needs revisiting, not a decision made once and left alone.

Why integrate AI into MaPS System rather than using standalone tools?

Many artificial intelligence tools already allow you to generate text, translate content, or edit an image. The difference with MaPS AI lies in integrating these capabilities directly into the heart of the product repository and operational processes.

When a third-party tool is used separately from the PIM or DAM, users typically have to extract information, formulate their prompt, copy the result, verify it, and manually re-enter it into the correct record and field. At scale, this method becomes virtually impossible to industrialise.

With MaPS AI, operations can be integrated directly into MDM, PIM, and DAM workflows without switching interfaces. Product context, category, attributes, media, target markets, and brand rules can all be leveraged automatically.

This integration specifically enables you to:

• Process hundreds or thousands of SKUs in a single operation;
• Apply the same rules across an entire catalogue;
• Save results directly into the relevant fields;
• Trigger automatic checks after each operation;
• Submit results for human validation;
• Track usage and model consumption (FinOps approach);
• Choose the most suitable AI model for each specific need.

MaPS AI adopts a vendor-agnostic approach. The platform can orchestrate various models based on their capabilities, required quality levels, processing volumes, or the company’s technical and financial constraints.

1. Marketing and E-commerce: Generating Context-Aware Content

Creating product pages is often one of the main bottlenecks for marketing and e-commerce teams.

A single SKU may require multiple descriptions depending on the brand, market, channel, language, or customer segment. Content created for a marketplace must meet different requirements than a brand site page, a B2B specification sheet, or a mobile app.

MaPS AI can generate this content directly from the reliable data stored in the PIM.

Generate Contextualised Product Descriptions

Based on available attributes, MaPS AI can produce:

• A short description for search results lists;
• A long description for product detail pages;
• Selling points and consumer benefits;
• Optimised titles and SEO metadata;
• Content tailored to marketplace requirements;
• Tailored versions for B2B and B2C audiences.

Guidelines can incorporate brand tone of voice, required or excluded terms, target length, and channel-specific constraints. Rather than generating generic copy, the AI leverages the structured product context and brand editorial guidelines.

Harmonise an Existing Catalogue

Catalogues fed by multiple brands or suppliers often suffer from inconsistent content. Some entries are highly detailed, others incomplete. Tones vary, units are inconsistent, and product benefits are expressed in different ways.

MaPS AI can review this content, identify discrepancies, and propose a catalogue-wide harmonisation.

This makes it possible to standardise:

• Title structures and description lengths;
• Vocabulary and editorial tone of voice;
• Feature presentations and measurement units;
• Mandatory disclosures and legal notices.

Tailor Content to Different Channels

Every channel comes with its own formats, character limits, and publishing guidelines. MaPS AI can adapt source data into multiple tailored versions without requiring teams to rewrite text manually.

For example, a comprehensive description can be adapted into:

• A summary for mobile apps;
• A sales hook for marketing campaigns;
• Optimised copy for marketplaces;
• Technical spec sheet for B2B portals;
• Sales talk tracks for commercial teams;
• Structured content for search engines.

2. International Expansion: Translating Without Losing Brand Context

Translating thousands of product listings is costly and often slows down entry into new markets.

Literal translation isn’t enough: industry terminology, regional language variants, brand voice, and target market preferences must all be respected. MaPS AI can automate a context-aware initial translation directly from PIM content.

Translations can take into account:

• Brand glossaries, along with mandatory or prohibited terms;
• Differences between British and American English;
• Regional variations within the same language;
• Product context, category, and technical specifications;
• Character limits allowed by the channel.

Workflows can then incorporate human validation for sensitive markets, products, or high-stakes content. AI accelerates the initial groundwork while local teams ensure final accuracy and relevance.

3. Creative Studio & Media Teams: Industrialising Workflows in the DAM

Media asset production involves repetitive tasks: background removal, renaming, cropping, format conversion, quality control, and channel adaptation. Performing these manually across thousands of files ties up significant resources.

Integrated into the DAM, MaPS AI can support the entire media asset lifecycle.

Automatically Classify and Document Images

AI can analyse images to:

• Identify the product shown and recognise the shot type;
• Distinguish packshots from lifestyle imagery;
• Generate keywords and alt text (ALT) for accessibility;
• Detect missing media assets and verify visual consistency;
• Automatically map the file to the correct product.

Automate Background Removal and Format Adaptation

MaPS AI can orchestrate automated processing tasks such as:

• Background removal or replacement;
• Cropping, resizing, and format conversion;
• Aspect ratio adaptation required by marketplaces;
• Resolution enhancement and quality verification.

Results are stored in the DAM and automatically associated with the correct product version.

Create New Visual Variations

Depending on configured models and services, AI can also generate new lifestyle renderings from existing assets. For instance, a studio packshot can be placed into various environments or adapted for seasonal campaigns—always governed by brand rules and approval stages.

4. Category Management: Analysing Offerings and Prioritising Actions

Category managers must prioritise amongst thousands of SKUs. They need to identify products requiring enrichment, underrepresented segments, positioning gaps, and actionable insights from sales data or customer reviews.

MaPS AI converts data within MaPS System into directly actionable recommendations.

Prioritise Product Enrichment

Instead of uniformly enriching the entire catalogue, AI cross-references key indicators:

• Sell-through rate, inventory levels, and margin;
• Traffic, conversion rate, and product page completeness;
• Store distribution footprint and seasonality.

Across a catalogue of 5,000 SKUs, MaPS AI highlights products with high commercial potential but incomplete product pages. Teams immediately know where to focus their efforts.

Analyse Assortment Diversity

MaPS AI generates category breakdown summaries—analysing features such as opening mechanisms, noise levels, power output, or range-specific capabilities. Results can be saved in MaPS System or exported as decision-making reports.

Leverage Customer Reviews

Manually sifting through thousands of reviews is rarely feasible. MaPS AI aggregates, summarises, and extracts recurring feedback (e.g., hard-to-open packaging, unclear instructions, sizing confusion).

These insights directly inform product detail improvements, supplier feedback, and range planning decisions.

5. Data Management: Checking, Completing, and Standardising Data

Data teams spend a significant amount of time catching errors, harmonising information, and filling in missing attributes.

MaPS AI assists teams in these operations without replacing deterministic rules already embedded in the MDM or PIM.

Detect Anomalies

AI can flag statistically or semantically inconsistent values, such as:

• A book weighing 50 kilograms or a shoe measuring two metres;
• A likely incorrect unit or a description contradicting technical attributes;
• An image that does not match the stated color attribute;
• Pricing data significantly outlying the rest of the product range.

Convert and Standardise Units

Supplier data is often provided in varying units (kilograms, pounds, inches, centimetres). MaPS AI helps identify units, convert them, and map them to the proper attribute in your data model.

Classify Products

Based on product descriptions or supplier documentation, AI automatically suggests a category, product family, risk tier, or list of relevant attributes for review.

6. Logistics: Ensuring Data Accuracy to Protect Shipping Costs

Inaccurate weight or dimension metrics lead to freight miscalculation, improper order fulfillment, or higher return rates. MaPS AI helps logistics teams validate data upon ingestion.

• Verify weights and dimensions: Automatic benchmarking against product category averages, similar items, and visual assets.
• Recommend shipping modes: Suggest packaging types, vehicle requirements, or specialized delivery services (e.g., standard vs. freight, climate-controlled).
• Identify special handling requirements: Automatic detection of fragile goods, hazardous materials, oversized items, or food products.

7. Procurement & Suppliers: Accelerating Data Onboarding at the Source

Buyers receive data in various formats: spreadsheets, catalogues, technical sheets, and PDFs. These documents must then be parsed, interpreted, and ingested into the PIM.

MaPS AI can automatically extract relevant data from these sources to pre-fill product records (SKUs, dimensions, weight, materials, regulatory, and logistics data).

This automation significantly reduces supplier onboarding time while improving collected data quality.

8. Compliance & Regulation: Identifying Sensitive Information

Regulatory obligations vary by product, market, and sales channel. MaPS AI supports teams with:

• Identifying mandatory disclosures and safety warnings;
• Automated parsing of compliance documentation;
• Detecting anomalies in HS/customs codes;
• Mapping products against target market regulatory frameworks.

9. Digital & Channel Management: Auditing Data Post-Syndication

High-quality data in the PIM does not always guarantee accurate display across websites, marketplaces, or reseller portals. Legacy titles or outdated assets may persist on external channels.

MaPS AI monitors syndicated data and compares published content against the master repository, generating channel-by-channel anomaly reports for digital teams.

10. Competitive Intelligence: Turning External Data into Insights

Using ethically scraped/authorised external data, MaPS AI analyses competitor pricing, range composition, key selling propositions, and customer feedback.

Prior to launching a new product, teams can identify market expectations, pain points, and high-value features.

Thousands of Scenarios Built Around Your Business Rules and Data

The examples above represent only a fraction of potential use cases. Every business operates with its own unique data, workflows, and priorities.

A tailored use case can be built by combining:

• A trigger event;
• A set of source data;
• One or more specialized AI models;
• Business rules and human-in-the-loop approval steps;
• An action executed directly within MaPS System.

Shorten Time-to-Market Without Sacrificing Quality

Automation isn’t just about scaling output volume—it also dramatically accelerates the onboarding of new products and brands.

In the Easypara project, onboarding a brand and its entire SKU line previously took over a week; with MaPS System, this was reduced to just two days.

The goal is to eliminate low-value tasks, accelerate data preparation, and empower teams to focus on quality oversight and strategic decision-making.