What AI Inside a PIM Actually Does - and Where It Breaks

AI in a PIM does a handful of jobs well, such as mapping supplier files, checking data quality and drafting descriptions, and it fails in places you can predict once you have seen a few projects. You need both sides when you plan, because the jobs it does well only pay off if you have also dealt with the places where it breaks. This guide covers both, then sets out a staged rollout, the governance that has to run alongside it, and the EU Digital Product Passport, which forces the same data work on a legal timeline. It is part of a series on AI and product information management: part 1, The product data problem no one is fixing fast enough, covers why the product data status quo is eroding competitiveness.

What AI does inside a PIM

Three kinds of AI do most of the work on a product catalog. Natural language processing reads and writes text, such as supplier descriptions and product copy. Machine learning learns patterns from your existing records and from the corrections your team has made, which is what drives classification and anomaly detection. Computer vision reads images, for example to tag a color or to check whether a photo meets a channel's requirements. Generative AI, the large language models behind tools such as ChatGPT, adds the ability to draft text. Each of these jobs still needs a person who knows which rule the model applies, because a model does not know your attribute rules until you teach it, and someone has to notice when it applies them wrongly.

1. 1✓Put files in one shape
2. 2✓Pull attributes out of text
3. 3✓Fill a gap from context
4. 4✓Check values
5. 5✓Map to each channel
6. 6✓Suggest search tags

Supplier file

Item desc

chair vintage wood

→ Name

Wt

-10 kg

→ Weight

Col

–

→ Color

“A red, vintage, wooden chair with a curved back.”

Your PIM record Example

Name

Vintage wooden chair

Material

Wood

Color

RedColor maps toAmazon: Product ColorShopify: Variant Color

Weight

-10 kg⚠ Impossible value: sent to a person

Search tagsvintagewoodenredchair

✓ A person reviews before publication

Supplier onboarding: generative AI on supplier files

Supplier files rarely arrive in the shape your PIM expects. A wholesaler that buys from many suppliers receives spreadsheets with different column names, units written in different ways, empty fields and values that cannot be true, and much of the onboarding effort goes into making those files fit before anyone can enrich a product. Once a model has been taught your standards, it can take on a good part of that work.

The first job is to put every file into one shape. The model learns the formats and terms a particular supplier uses, then maps incoming rows onto the format your PIM already accepts, including product names, SKU patterns and the place each product takes in your category tree. The second is to pull attributes out of raw text. Descriptions, customer feedback and technical specifications often arrive as running text, and the model can extract the attributes you need and place them on your data model, so enrichment starts from structured fields instead of a block of prose.

The model can also fill a gap from context. If the color field is empty but the description names the color, it can write the attribute, and a fashion retailer receiving many supplier feeds can fill missing material composition or care instructions in the same way, from patterns already in the data. Finally, attributes can be compared with sources you trust, and the same pass flags missing values, impossible ones (a weight of -10 kg) and spelling mistakes. Each hit is either corrected or sent to a person, depending on how confident you are in the rule behind it. The work does not disappear: someone still decides which supplier terms map to which of your values, and reviews the first batches from a new supplier until the mapping holds.

Data quality checks

Machine learning can be trained on past records and on corrections people already made. From those it learns what a normal record looks like, so it can flag anomalies, inconsistencies and mistakes, and suggest a cleansing or standardization step, such as writing every voltage in the same unit. It can also check a record as it comes in, against rules you set: incomplete attributes, missing or wrong numbers, items that miss a regulation. Running that check before publication matters because a bad value does not stay in the PIM. Once it is published it appears on the webshop, in every marketplace feed and in the print catalog, and correcting it then means chasing it through every channel.

Classification and taxonomy

A model can classify products from their attributes and from earlier classifications. A technology retailer can separate "High-end Gaming Laptops" from "Budget Laptops" on processor, RAM and graphics card, and keep that split as the range changes, so a newly added laptop lands in the right category without someone reading every specification. On taxonomy, AI can propose a hierarchy and the links between categories from attributes and semantic similarity. That helps most in a catalog that has grown through acquisitions or new suppliers, where the same kind of product sits in several places under different names. A cleaner structure lets shoppers reach the product with fewer wrong turns, and the filters on a category page work better because the products in it share the same attributes. The proposal still needs someone who owns the category tree, because a model can group products that look alike in the data but are bought by different customers for different reasons.

Channel mapping

Every marketplace and channel has its own structure and its own required attributes, and each one changes them from time to time. A model can learn those requirements and propose the map from your attributes to them. It can learn, for example, that Color maps to Product Color on Amazon and to Variant Color on Shopify. You then maintain fewer hand-built maps per channel, which matters most when you add a new marketplace and would otherwise build its map field by field. Someone still approves the proposed map, because a wrong match, such as a net weight sent where the channel expects a shipping weight, passes every technical check and only shows up later as a rejected listing or a customer complaint.

Descriptions, enrichment and translation

Drafting copy from attributes is faster than writing each SKU by hand, and the tone stays closer to one standard than when several people write. That frees marketing time for the work the model should not own, such as campaign copy and the products that carry the brand. An appliance retailer with a large intake can generate descriptions that cover energy efficiency, how the product is used, and the features that set it apart. Given a list of laptop specifications, a model can write what they mean for a buyer who is not technical. For a new line of monitors, it can turn screen size, resolution, refresh rate and connectivity into a structured first draft that someone then reviews. AI translation for the language pairs with the most traffic follows the same pattern: the model writes the first version, and a person who knows the market checks product names, units and the terms your industry uses.

Tags and product questions in the shop

A model reads descriptions and attributes and suggests tags for on-site search. For a "red, vintage, wooden chair" it suggests vintage, wooden, red and chair. It can also answer common product questions from the attributes ("Does this phone have wireless charging?"), which leaves support capacity for the harder ones. Both uses depend entirely on the attributes underneath, so a tag or an answer generated from a wrong attribute gives the shopper wrong information in a confident tone.

Around the PIM

Some AI uses depend on product data without living in the PIM: demand estimates from past sales and market trends, recommendations from preferences and purchase history, image recognition that reads color, pattern, form and texture for visual search, and semantic search that works from context and intent. Each needs the enriched record to reach marketing automation, the CMS and the commerce platform. A generated recommendation is still generated text, and the product data underneath it still has to be right, which is why these uses tend to expose gaps in the PIM rather than fix them.

Where AI in a PIM breaks

The same failure points come back across all of these jobs, and most of them have more to do with the process around the model than with the model itself. The first is that the model does not know your rules. Without your attribute definitions and content standards you get generic output, and supplier cleansing, classification, descriptions and recommendations are separate switches that each need their own rules. A team that has configured description drafting carefully has not yet configured classification, even when both run in the same PIM.

The second is the material the model learns from and works with. Training sets carry their own mess: classification is only as consistent as the examples the model learned from, so a catalog where the same product type sits in two categories teaches the model to repeat that split. Outside material carries a similar risk. Photos, user-generated content and related products found outside the PIM have to be checked against the record you trust before they reach a product page, because nothing in the material itself tells you whether it belongs to the current version of the product.

Handoffs introduce errors of their own. When output is copied by hand between a chat tool and the PIM, copies drift apart: someone edits the text in one place and not the other, or pastes a description onto the wrong variant. Fluent text makes these errors harder to spot, because generated copy reads well whether or not the attribute underneath is right. A description that explains the benefits of a large screen in polished sentences is still wrong if the screen-size attribute it was built from is wrong.

Quality also drifts. Output that was good when the feature was configured does not stay good on its own, because new suppliers, new product types and changes to the model all shift the result, so it needs sampling and review long after go-live. Finally, compliance is a different job from enrichment. An AI-powered PIM is not automatically ready for a Digital Product Passport, which needs verified supplier and lifecycle data with audit trails, as the section on passports below explains.

That suggests an order. Start with extraction, enrichment and quality checks, where the rules are easiest to write down and the results easiest to check. Classification, taxonomy and rule-based validation build on those, because they need consistent attributes to work from. In each case the catalog improves only where someone still owns the rule and reviews what the model does with it. Why a model cannot clean data it has no rules for is covered in Cleaning data: why AI isn't a magic fix.

Relay, bridge, highway: where the model sits

ChatGPT and open-source models are easy to try, and many product teams have already used one to draft a description. Turning that experiment into a product data workflow is harder, because the model then has to fit into the path from ERP to PIM to the shop. Three positions describe where it can sit on that path.

Relay. Generative tools sit outside the systems that store and publish product data. Product data starts in the ERP: inventory, price, SKU. A person moves it into a generative tool to draft descriptions, then moves the text into the PIM, which publishes to the shop and other channels. You get the model without changing the process, and you also get the failure mode of handoffs: inconsistent copies, and time lost to switching systems even though the sentences themselves were generated. The gain is a faster first draft, with a person still responsible for what lands in the record.

Bridge. The generative features move into the systems you already run. The ERP can categorize products and clean inconsistencies in master data with a built-in model, and the PIM can draft descriptions and build sets. Enriched data goes to the shop and other channels without a paste step. That removes most hand-made mistakes, and because the process runs the same way every time, it becomes easier to review and to measure.

Highway. The systems are built with the model inside the flow. The ERP feeds the PIM, and the model in the PIM drafts descriptions, categorizes products, and predicts pricing and sales performance. It keeps learning from feedback, from how people use it, and from new data, so the quality of its output depends on the quality of that feedback as much as on the first configuration.

The practical test is whether enrichment, mapping and publishing still depend on a person moving text between screens. If they do, you are in relay. Starting there is a reasonable way to learn what a model can do with your data, as long as you count the time and the errors that the handoffs cost.

Relay

The model sits outside your systems. A person carries the text across.

Faster first draft, handoff errors

Bridge

The model sits inside the ERP and the PIM. No paste step.

Fewer hand-made mistakes

Highway

The systems are built with the model in the flow, and it keeps learning.

Drafts, categorizes, predicts

Does a person still move text between screens? Then you are in relay.

How to roll out AI in a PIM: four stages

The most common mistake is to start with technology selection instead of a data audit. A team that picks an AI feature first has no baseline to measure it against and no standards to configure it with, so the first months go into discovering problems that an audit would have shown up front. The four stages below put the work in an order where each stage produces the evidence for the next.

1. Months 1 to 3

   Foundation: audit the data, set the standards

   Before you evaluate any AI capability, establish the current state of the catalog: completeness by product category, average enrichment time per SKU, defects by attribute type, the main data sources and how reliable they are, and the number of records in each language. That baseline shows where AI will help most, gives you a way to measure return, and surfaces the quality problems that would undermine any AI you deploy.

   At the same time, write the standards down: attribute requirements per channel, category hierarchy rules, terminology and brand voice. They cannot stay vague when you configure AI, because their precision decides how consistent the output is. Time spent on them before any AI configuration saves remediation later.

2. Months 3 to 6

   Targeted activation: the cleanest slice first

   Apply AI to the cleanest, highest-priority part of the catalog first, typically the SKUs that carry most of the revenue. The quality work is justified there, and the results are easiest to read commercially. Three use cases, in this order: generated descriptions for high-volume standard products, where manual effort is highest and the output simplest; gap detection and completeness scoring across the whole catalog; and AI translation for the language pairs with the most traffic.

   Measure everything against the baseline from stage 1. The aim at this stage is a set of validated results that justify the next one, so resist the pressure to switch on every feature at once. A feature that cannot show an improvement on the cleanest slice will not show one on the rest of the catalog either.

3. Months 6 to 12

   Scaled deployment: the whole catalog, with governance

   Extend the AI workflows to the full catalog, with the standards and governance from the earlier stages. Workflow automation pays off most here: task routing, automatic approval triggers for records that meet quality thresholds, and AI-assisted import mapping for supplier feeds. Keeping AI-assisted content work inside the system that already manages and publishes product information means reviews and approvals happen before content goes out.

   Add AI on the asset side as well: image tagging, quality checks against channel requirements, and tracking of usage rights, so a supplier image needs less handling before it is ready to publish.

   If you have Digital Product Passport obligations, this is the stage to build the DPP attributes and processes into the PIM data model: lifecycle data fields, supplier data validation and audit trails. Build them in now rather than retrofit them later as a parallel compliance project.

4. Months 12 to 24

   Predictive intelligence: link quality to results

   Connect catalog quality to commercial outcomes: conversion, returns, search performance and channel rejections. Those feedback loops let content models improve, and once you can show that better product data lifts conversion, funding decisions get easier. At that point the PIM also becomes a source of commercial insight, beyond its role in managing data.

   Worth piloting at this stage: conversational interfaces, where people onboard products in natural language instead of filling in forms, which shortens training for new users; and autonomous enrichment, where the system gathers attribute data from outside sources such as manufacturer specifications, regulatory databases and standards bodies, and pre-fills records for human review.

Governance across every stage

Every stage of the rollout needs governance, and most organizations underinvest in it because it produces no visible output of its own. Three elements cannot be skipped.

The first is content governance, and it has to be in place before content generation runs at scale. That means explicit brand voice guidelines, training sets curated from your best-performing content, review thresholds per product category and commercial importance, and sampling that checks AI output continuously instead of assuming it stays good after configuration. Apply extra scrutiny to content tied to trusted brand names, where buyers expect accuracy. Keep people on evaluative and creative work, where the perception of automation affects credibility.

The second is data lineage and audit trails, built into the architecture. For every record you should know which model generated it, when, from which input, and who reviewed it before publication. DPP compliance makes this mandatory for regulated product categories. It is good practice everywhere, because it lets you trace and fix quality problems systematically: when a wrong value turns up, you can find every other record produced the same way instead of correcting one complaint at a time.

The third is supplier data coordination, started early. AI performance and DPP compliance both depend on clean, structured supplier data, so set clear requirements for material origins, component details and sustainability data, and agree on the format in which suppliers deliver them. The most common cause of delay is an AI system that cannot perform because the upstream data is inconsistent, incomplete or arrives in formats that need manual work. A manufacturer that sources components from many suppliers can find that agreeing on one data template takes longer than configuring the model.

The roles and rules behind this are covered in product information governance.

Digital Product Passports: the same data work

Most PIM conversations are about efficiency and conversion. The EU Digital Product Passport (DPP) deserves the same attention, because it forces investment in product data on a legal timeline, whether or not the commercial case has been made internally.

A DPP is a machine-readable record per product, reached through a data carrier such as a QR code, NFC or RFID. It covers material composition, origin, environmental impact, recycled content, repairability and end-of-life instructions. It is a data infrastructure requirement rather than a marketing document: regulation decides what goes in it, and every value has to be correct, complete and traceable. Batteries are the first product group with a mandatory passport, from February 2027 under the EU Battery Regulation. Other product groups, textiles and furniture among them, follow in phases through product rules under the Ecodesign for Sustainable Products Regulation (ESPR). The obligations follow the product, so they also apply to manufacturers outside the EU that sell into it.

Most manufacturers today rely on scattered supplier documents, ad-hoc lifecycle assessments, manual processes and static product reports. A DPP needs data flows most PIM implementations were not designed for: verified material origins, component-level sustainability data, lifecycle data from suppliers, and version-controlled audit trails. That is why an AI-powered PIM is not automatically DPP-ready. AI can help fill and check those fields once the flows exist, but it cannot verify where a material came from if no supplier has told you.

The two investments are largely the same investment. AI and DPP compliance both need clean, structured, centralized product data, governance and supplier coordination. One roadmap that serves both gets more from the same infrastructure spend than two separate programs, which would each build their own version of the same product data. The passport itself is covered in Digital Product Passport, and its overlap with packaging rules in PPWR and the Digital Product Passport.

What the organizations that get this right do

The companies that get commercial returns from AI in their PIM share a recognizable profile. They treat product data as an asset that earns revenue and budget for it accordingly, instead of treating it as a cost to keep down. They measure the catalog before they deploy AI, so they can show what changed, and they keep to the sequence: data foundation first, targeted activation second, scaled deployment third, predictive intelligence fourth.

They also put governance in place before content generation goes live at scale, and they coordinate supplier data requirements early and explicitly, instead of discovering halfway through the project that suppliers cannot deliver what the model needs. And they run DPP compliance and AI on one data foundation, with one budget, instead of as separate initiatives that each pay for the same groundwork.

None of this works on a catalog that nobody owns, because a model can only apply rules that someone has written down and keeps up to date. If you are still deciding whether a PIM is the right foundation, start with what a PIM is and how to select the right PIM.

Where to go from here

Both tools are free and independent of any vendor.

Not sure you need a PIM yet?

A short check that shows whether a dedicated PIM helps your catalog, or whether spreadsheets and ERP still do the job.

Check if a PIM fits

Ready to compare vendors?

Answer questions about your scale, sector and must-haves and get a ranked shortlist.

Build my vendor shortlist