The real starting point: a legacy PIM that must keep running
In a business, the PIM is not the vendor's latest demo. It is often an older, customised version connected to an ERP and publishing channels that cannot be interrupted. It can export and import files and may expose an API; that is already enough to add an enrichment workflow without changing its role as the system of record.
The first question is therefore not “what AI does the PIM have?”, but “how can data leave it and return cleanly?”. Only then should you consider where the source material lives, which controls are required and how frequently processing must run.
01 · Legacy PIM
Keep it as the system of record
Exports · imports · API if available
02 · Recent PIM
Use native features when they work
Attributes · descriptions · assets already present
03 · Missing data
Source it before feeding it back
Websites · scattered PDFs · catalogues · third-party datasets
A documented feature is not necessarily a deployed feature
Documentation for recent Akeneo versions describes broad capabilities: generation and translation from attributes, images and PDFs, automation rules, human review and even the use of web data in certain 2026 offers. That describes the vendor's current product — not automatically the PIM installed at a customer's premises.
In practice, the version, edition, customisations, permissions, hosting model and integration constraints can make those features unavailable or unsuitable. An Akeneo environment may work perfectly well as a PIM while being unable to connect the AI features shown in recent documentation. The audit must examine the real instance, never the vendor's product page.
Decision framework
| Real situation | Practical choice | Control point |
|---|---|---|
| Legacy PIM with file imports and exports | Process batches externally, then import them back into the existing schema. | Stable identifiers, formats, allowed-value lists and error reports. |
| PIM with a usable API but no useful AI | Connect an external layer without moving existing workflows. | Quotas, write permissions, error recovery and logging. |
| Native AI features genuinely available | Use them to generate, translate or reformat content already present. | Test quality, governance and cost on a representative batch. |
| Missing characteristics and scattered sources | Research and cross-check upstream, regardless of the PIM's age. | Source hierarchy and provenance for every value. |
| Matching two product catalogues | Use a matching engine or specialised component. | Justification for matches and handling of non-matches. |
| Catalogue publishing and governance | Keep the PIM as the system of record. | The external layer feeds the PIM; it does not replace its workflows. |
Four questions before choosing an architecture
- 1. How does the instance exchange data? Actual export, import, API, frequency, volume and validation rules.
- 2. How much information is already present? Separate structured data, copy, assets and material that still lives outside the PIM.
- 3. Do you need to generate or investigate? Generation transforms known context; investigation discovers and establishes a missing fact.
- 4. What evidence must return with the value? A reviewed suggestion does not carry the same guarantee as a value linked to a precise page in a manufacturer's document.
The legacy PIM can stay exactly where it is
A sound architecture keeps the PIM as the system of record for categories, attributes, locales, channels, permissions and publishing. An external component only handles the missing upstream work — collection, cross-checking, normalisation and matching — then returns values in the accepted format. A simple file exchange may be enough; an API improves throughput but is not a prerequisite. If native AI becomes genuinely available later, it can complement this setup.
Related reading: Akeneo integration · product data enrichment method · product matching
Sources and verification date
- Akeneo — AI-Enhanced Enrichment in the PIM;
- Akeneo — AI behaviour, limitations and realistic use cases;
- Akeneo — executing AI configurations manually or through rules;
- Akeneo — February 2026 Serenity updates.
Features checked against official documentation on 15 July 2026. These pages describe Akeneo's recent offering, not every production instance; availability and integration options vary with the customer's version, edition and architecture.