Data quality software used to be something IT bought after a bad audit. In 2026, it decides whether AI projects reach production. This article explains what the tools do now and where they fall short for product data. The second half covers how PIM keeps product data accurate before it gets published.
Key Takeaways
- Poor data is the main reason AI projects stall. Gartner tied AI project abandonment directly to missing AI-ready data.
- Modern data quality software pairs rules for known problems with observability for unexpected changes. Both create alerts, so someone has to own them.
- AI agents fix data and break it at the same speed. Treat them as data producers with approval steps and audit logs.
- A completeness score shows that fields are filled. It doesn't show that the values are correct.
- For product data, PIM is the place to block errors before they reach a channel. Tools further downstream can only report them.
- EU manufacturers face fixed dates. Battery passports become mandatory in February 2027, and more product groups follow.
Why Data Quality Became An AI Problem
A Gartner survey of 1,203 data management leaders in July 2024 found that 63% of organizations either lacked the right data management practices for AI or weren't sure they had them. From that, Gartner predicted that through 2026, organizations would abandon 60% of AI projects unsupported by AI-ready data.
The logic is simple. A dashboard shows a strange number, and someone frowns. An AI agent takes the wrong weight or the outdated price and acts on it. Nobody frowns until a customer does.
Gartner's analyst summed it up in the same release:
"If the data has issues, then the data is not ready for AI."
The release also treats AI-ready data as an ongoing practice. It recommends data observability and better metadata management, and it expects data requirements to shift with each new AI use case. So a one-time cleanup project won't hold for long.
What Data Quality Software Does In 2026
Rules And Observability Catch Different Problems
Rule-based checks cover what you already know. A GTIN has 14 digits. A price can't be negative. A voltage value needs a unit. You write the rule once and the tool flags every record that breaks it.
Observability covers what nobody wrote a rule for. The tool learns normal patterns for freshness, volume, schema, and value distribution, then alerts when something shifts. A supplier feed that usually delivers 4,000 rows arrives with 400. A column changes type after an ERP update.
Most mature teams run both. The usual failure is alert overload. Switch on every detector and the team soon stops reading alerts. Decide who receives which alert and tune thresholds in the first weeks, or the tool turns into expensive background noise.
AI Moved Inside The Tools
Plain-language rule authoring has become a common feature. A data steward writes "every cable needs a cross-section in mm²" and the tool generates the check. Many tools also suggest rules from profiling results, propose fixes, and answer questions about quality results in chat. Some vendors now ship agent-style "data stewards" that work through issue backlogs.
This lowers the skill barrier for writing checks. Someone still has to know what the rule should say, because the tool will happily write a perfect check for a wrong requirement.
Quality Gets Defined By Who Uses The Data
Good data now means data that fits a specific consumer. A product record can work fine for the ERP and fail on a marketplace that requires twelve attributes the ERP never stored. Data contracts formalize this. The team producing the data and the team consuming it agree on fields, formats, owners and service levels, and the tool checks against that agreement.
Unstructured data is in scope now too. Datasheets, PDFs, and images feed AI systems, so the tools are expected to extract attributes from them and check the results.
Checks Move Closer To The Source
Cleaning data in the warehouse every night repeats the same fixes forever. So organizations push checks into the systems where data is created. MDM handles master records. PIM handles product content.
Risks To Plan For
The biggest risk is AI plans running ahead of data readiness. The Gartner numbers above describe it well enough. The fix is boring: assess data readiness before funding an AI project, and do it before the pilot instead of after it fails.
AI agents bring a newer risk. They create data as fast as they fix it. In January 2026, Bloomberg reported that Amazon's experimental Buy For Me tool had automatically generated listings for products from independent shops, including shops that never chose to sell on Amazon. Some listings pointed to the wrong item. One stationery store got refund demands after customers ordered a larger stress ball the shop doesn't stock and received the smaller one it does. Other merchants saw wrong product images and wholesale prices shown to shoppers. Amazon said sellers could opt out and that the program covered more than 500,000 items, up from about 65,000 at launch.
When a brand doesn't publish structured product data, someone else's AI will guess it. The brand still handles the complaints.
Vendor consolidation is a slower risk. Informatica now belongs to Salesforce, and Talend belongs to Qlik. Acquired tools can get merged, repriced, or deprioritized. If a vendor changed owners recently, put data export formats and exit terms in the contract.
The last risk sits in the business case. Many published data quality statistics come from vendor surveys that cite other vendor surveys, often without a date or sample size. Before a number goes into a budget request, find the source and check the year and who was asked.
How PIM Enforces Product Data Quality
Data quality tools in the warehouse watch pipelines. By the time one flags a wrong product weight, that weight may already sit in several distributor feeds. PIM sits earlier, where product data is created, enriched, and approved. That position lets it block a bad record before publishing.
Our customers often come to us after product data has outgrown the ERP. For a typical manufacturer, the ERP item master holds what production and logistics need. Marketing attributes, translations, and certificates live in spreadsheets maintained by whoever had time that week. Every distributor wants its own template, so the same product gets retyped for each one, and each retyping adds new mistakes.
In projects we implemented, the fix started with a shared data model. Then the checks moved into that model. These controls do most of the work:
- Typed attributes with fixed units and allowed values. A weight field accepts a number in kilograms, so "approx. 2 kg" never makes it in.
- Validation rules per channel and language. A marketplace may require a GTIN and package dimensions, while the German catalog requires a translated safety notice.
- Completeness and readiness scores per channel, so the team sees which products can go live where.
- Supplier onboarding through validated templates or a portal. Suppliers fix their own errors at upload, before anyone on your team touches the data.
- AI extraction and enrichment with human approval. The AI pulls attributes from datasheets and drafts descriptions, and a person approves the result.
- A full audit trail. Every change records who made it, when, and the previous value, including changes made by AI.
- Checks against each channel's specification before export, so a feed that breaks a retailer's rules never leaves the system.
Completeness Scores Can Lie
Completeness is the easiest number in PIM. It counts filled required fields and turns them into a percentage. It's also easy to game.
A field filled with the wrong value scores exactly like a correct one.
The dashboard stays green while distributors send products back. We regularly see catalogs that look near-complete, where a spot check against the original datasheets still finds wrong dimensions and expired certificates.
Readiness rules go further. They ask whether a product meets the needs of one specific consumer: a distributor's item specification, a marketplace's category rules, a regulator's data schema, or an AI shopping assistant that filters by attributes. Each consumer gets its own rule set.
So add accuracy sampling. Each month, pull 20 random records per product family and compare them against datasheets and certificates. The error rate from that sample tells you more than the completeness score does.
Treat AI Agents Like A New Supplier
PIM vendors now ship agents that map supplier data, write descriptions, translate, and classify products. They work fast. An agent that enriches 10,000 products in an hour can also get 10,000 products wrong in an hour.
Govern them like a new supplier whose data you don't trust yet. Agents propose changes and a person approves them. Permissions stay scoped, so an agent can edit descriptions but not prices or safety data. Every agent action gets a log entry. AI-generated text carries a label so reviewers and auditors can find it later.
Customers who start AI enrichment with us usually begin with low-risk fields such as marketing copy and translations. They keep approval on every change at first. Once sampled error rates stay low, they widen the scope to technical attributes. This is slower than switching everything on at once. It also avoids a cleanup project six months later.
Publish Structured Data Before Someone Else Guesses It
AI shopping assistants filter products by attributes, units, and categories. A product with a missing voltage or a free-text material field drops out of the candidate set before anyone reads the marketing copy. The Amazon case shows the other side. When structured data is missing, a third-party AI fills the gap with its own guesses.
So publish authoritative product data in formats machines can read: structured feeds and markup for channels, and APIs for partners and AI tools. PIM is the natural source, since its data has already passed validation.
The Digital Product Passport Turns This Into Compliance
The EU Ecodesign for Sustainable Products Regulation (ESPR) entered into force in July 2024. Its Digital Product Passport (DPP) is a digital record, usually accessed through a QR code, that follows a product through its value chain. According to an overview by the law firm Krogerus, it will hold information such as materials, origin, carbon footprint, durability, repairability, and recyclability.
The EU's central DPP registry went live in July 2026, but companies only need to register product data once the delegated act for their product group applies. Battery passports become mandatory on 18 February 2027 for electric vehicle, light-transport and industrial batteries above 2 kWh. The delegated act for textiles is expected in 2027, with compliance likely about 18 months later. Over time, the regime is expected to cover nearly all products placed on the EU market, which pulls in manufacturers, importers and distributors based outside the EU too.
Most passport data comes from suppliers and engineering. Manufacturers preparing for the DPP usually find that material composition exists only in supplier certificates and PDFs, in no system that can export it. In projects we implemented, we modeled passport attributes as a separate attribute group and collected them through the same validated supplier onboarding flow as other product data. That way, the data arrives in a checked format, with a record of who supplied it.
Exact DPP data requirements depend on delegated acts that aren't all published yet. Build the model so attributes can be added without a redesign.
Where PIM Ends, And Other Data Quality Software Begins
PIM manages product content for selling: attributes, descriptions, assets, translations, and channel requirements. MDM builds golden records across domains such as product, customer, and supplier, with matching and merging of duplicates. Data quality and observability tools monitor the pipelines that feed analytics and AI.
Many companies run all of them. A common setup uses the ERP or MDM as the source of technical truth and PIM as the enrichment and readiness layer, with observability on the pipelines. The boundaries blur, since several vendors now cover both PIM and MDM. Map the capabilities you need and ignore the category labels.
How To Choose Data Quality Software For Product Data
- Test with your own data. Load a messy supplier file and a real ERP export during the trial, since demo data is always clean.
- Check how the tool handles units, variants, and multilingual attributes. Tools built for analytics tables often treat product data as flat rows.
- Find out what the AI features can change without approval and whether each AI edit gets logged.
- Skip the star average in reviews and read the complaints. They show learning curves and integration limits faster than a demo.
- If the vendor changed owners recently, ask about the roadmap and get export formats and exit terms in writing.
- Ask for the source of every statistic in the vendor's business case.
Start with one product family and one channel. Define readiness for that pair and measure accuracy on a sample. Then expand.