Main site

Retail AI agent · Catalogue and listings

Catalog-Enrichment & Listing AI Agent

Draft product copy and normalise attributes against the source documents on file, refuse the values no document supports, and leave certification, warning and origin fields to your compliance specialist.

Typical delivery
4–6 weeks
Deployment
Your stack
Compliance
Source-bound
After launch
Agent Care

What this agent does

An attribute is a specification, not copy

In

  1. 01

    Collect the product record, supplier declarations, certificates and the taxonomy from supported PIM.

  2. 02

    Map each incoming value onto the controlled vocabulary, and keep the document identifier that supports it.

Reason

  1. 03

    Compose the marketing description and bullets inside the claim vocabulary cleared for that category.

  2. 04

    Treat certification, warning, origin, fibre and ingredient fields as read-only to the agent.

  3. 05

    Return an explicit unknown wherever no source document establishes a value.

Decide

  1. 06

    Check the listing against recall feeds, certificate expiries and the destination-jurisdiction matrix.

  2. 07

    Hand regulated attributes and first publications to the category manager and the product-compliance specialist.

Out

  1. 08

    Store the prompt, the model version, the source document and the approver against each attribute version.

  2. 09

    Execute write actions only inside the approval boundaries agreed during implementation.

Product statement

The agent proposes attribute values with their sources; the compliance specialist sets the regulatory fields, and the category manager owns first publication in a regulated category.

Example workflow

One listing, source to publication

  1. 1Product record receivedPIM entry, supplier feed, seller onboarding form or syndicated content
  2. 2Sources gatheredSupplier declarations, certificates, test reports and the taxonomy in force
  3. 3Attributes proposedControlled values, drafted copy, unresolved fields and confidence
  4. 4Conformance checks runSource-binding checks, recall and expiry checks, jurisdiction checks and confidence threshold

No human action required

Stages 1 to 4 run unaided and nothing is published at any of them — the agent is proposing, and the approver lane opens at the confidence gate.

5DecisionBranches on the attribute gate
Sourced attribute

Goes to the category manager to approve.

Unsourced attribute

Adds a compliance countersignature first.

Category and compliance approval

The listing is held with its source documents, its unresolved fields and the confidence.

Approve · Correct · Send to compliance

Approved — released to publish

  1. 6Catalogue and channel systems updatedOnly where write access and approval policy allow it
  2. 7Listing evaluatedSource-binding rate, corrections made, marketplace rejections and post-publication suppressions

Retractions

Every correction made in approval is counted in the evaluation.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required

8 items

  • Setting a certification, compliance or test-result attribute.
    • Setting country of origin, fibre content or material composition.
    • Writing or placing any safety or state-mandated warning.
    • First publication in a regulated category or a new jurisdiction.

  • Automation boundary

    Agent acts unaided

    • ✓Draft marketing descriptions and bullets inside a cleared claim.
    • ✓Normalise and deduplicate attributes into approved.
    • ✓Generate non-regulatory imagery variants with.
    • ✓Reconcile the published catalogue against recall feeds for the named owner.

    Any write happens inside the boundaries agreed at implementation, never ahead of the approvers' sign-off.

    • Naming the manufacturer identity block or the EU responsible person.
    • Recall, stop-sale, suppression and republication decisions.
    • Clearing brand-owned copy, imagery or trademarks.
  • Changes to the taxonomy or its controlled vocabulary.

Example output

One attribute, annotated

Everything the agent proposes is attached to the document that supports it.

Enrichment output · single listing

Illustrative example
Listing
Outdoor jacket
Proposed value
Water-resistant to the ingress class stated on the test report
Field type
Marketing field
Document of record
Laboratory test report
Confidence
87%
Approver
Category manager

As received

Taken from the supplier declaration and the test report on file — nothing on this side is inferred.

Sources used

  • Supplier declaration
  • Laboratory test report
  • Certificate of compliance

Why this value

The report rates ingress, so the wording stays rated and the absolute stays out.

What the score decides

Below the configured threshold the listing picks up a compliance read before it reaches.

Action

ApproveCorrectSend to compliance

Value

Where AI adds value

The same four claims, placed at the point in the workflow where each one applies.

Where the value lands

Value 01 – 04

Every listing

From the record and its documents

03Enrichment

Propose against the source

Draw on supplier declarations, certificates and test reports, and cite the document behind each value.

01Approved path

Say what the product is

Descriptions, bullets and controlled attributes come back drafted and already cited.

02Human review

Put the regulated fields in front

Unsourced values, expiring certificates and regulated categories are flagged, so review starts where warranties attach.

04

Build an evidence trail

Each attribute keeps the source that supports it and who approved it.

Integrations

Typical integrations

Five system groups connect to the same agent. Which of them are in scope is decided in discovery.

  • Product information

    • Salsify
    • Akeneo

  • Channel syndication

    • Rithum
    • ChannelAdvisor
    • Feedonomics

  • Assets and imaging

    • Bynder
    • Cloudinary

Agent

Catalogue enrichment & listings

  • Reads the sources
  • Proposes attributes
  • Holds for approval

  • Marketplaces and compliance

    • Amazon SP-API
    • Walmart
    • Assent
    • Compliance & Risks

  • Observability & evaluation

    • OpenTelemetry
    • Langfuse
    • Supported monitoring/evaluation sources

Integration availability depends on the client's existing systems and API access.

Agent controls

Six layers between the model and the catalogue

Each control wraps the next. The map beneath names what escapes.

L6 · Outermost — last line of defence

Inward → L1 · closest to the model

L6Rollback / safe modeHold publication and fall back to proposals when evaluation or production signals degrade.Roll back
L5Version monitoringTrack model, prompt, taxonomy and controlled-vocabulary changes.Track
L4TraceabilityRecord the source document, the proposed value, the approver and the publication time.Record
L3Category approvalHold listings for the named approvers; it governs release, not whether an approved attribute is right.Gate
L2Policy guardrailsTest values against the controlled vocabulary, the source binding and the jurisdiction matrix; a failure returns the listing.Restrict
L1Confidence thresholdsRoute low-confidence listings to a compliance read before the category manager sees them.Require review
Model coreListing produced — controlled values, drafted copy, unresolved fields and confidence
L1 – L2
Test whether a listing may stand
L3
Puts the release in two named hands
L4 – L5
Hold the source behind every attribute
L6
Stops publishing and holds drafts

How Nestack evaluates it

Evaluate the whole enrichment path — not only the finished listing.

Coverage runs the whole depth of the workflow, and every layer is cut by slice.

Surface — the listing the shopper reads

  1. E1Final-output evaluationDid every regulated attribute trace to an unexpired document of record?
  2. E2Step-level evaluationDid the agent use the current declaration, certificate and taxonomy?
  3. E3Tool evaluationDid it read and write the correct product and the correct field?
  4. E4Confidence calibrationDo low-confidence attributes actually attract more corrections?
  5. E5Slice evaluationHow does performance change across specific product taxonomies?
  6. E6Business outcomeHow many published listings were corrected, suppressed or withdrawn?

Floor — the listing the market reads

Failure modes

Where each failure originates in the agent

Seven modes, from source to published listing.

Agent lifecycle

Direction of processing →

  1. 01 · Retrieval

    1 mode

    • EN-03

      Superseded spec sheet

      Fibre content and origin are written from a declaration the supplier has replaced.

    Stage gathers

    Product records, declarations, with the source each came from

  2. 02 · Reasoning

    2 modes

    • EN-04

      Rated value made absolute

      A rated ingress class is normalised into a waterproof flag.

    • EN-06

      Family-wide generalisation

      One certified variant's hypoallergenic claim spreads across the parent.

    Stage proposes

    Controlled values, drafted copy and confidence

  3. 03 · Tool / write

    2 modes

    • EN-02

      Fan-out before review

      An enriched set pushes through the PIM to every endpoint ahead of approval.

    • EN-05

      Suppressed listing restored

      A catalogue re-sync republishes a SKU sitting under an open recall.

    Stage writes

    Only where write access and approval policy allow it

  4. 04 · Output

    1 mode

    • EN-01

      Warning in the wrong language

      An EU offer carries a machine-translated warning, or no responsible person.

    Stage returns

    The listing the approvers sign and the shopper reads

  5. 05 · Change / Version

    1 mode

    • EN-07

      Attribute silently remapped

      A schema update moves a certified value into a field that does not mean the same thing.

    Stage tracks

    Model, prompt, taxonomy and vocabulary versions

  • Sev-1 · a held act performed by the agent
  • Sev-2 · a false attribute reaches the catalogue
  • Sev-3 · a source degrades
See the complete Retail failure-mode catalogue

Affected slices

One cohort carries most of the exposure

Newly onboarded sellers in regulated categories, listing into Europe and California at once Nestack reports it by slice rather than in aggregate The cohorts that carry it are named, not averaged away..

Slice performance — reported separately, not only in aggregate

Illustrative example
SliceFailure rateLiftLift vs. thresholdStatus
New sellers, regulated, EU and California9.1%3.8×Review
Imported goods with no declaration6.7%2.8×Review
Brand-authorised syndicated content3.6%1.5×Watch
Repeat SKUs, certified attribute set1.2%0.5×Normal

Bar: unsourced-attribute lift vs. the all-listing baseline · scale 0–4.0× · tick marks the 2.0× review threshold

2 of 4 slices over threshold

Evidence-linked improvement

A cycle ends as a case in the suite

The cycle is over once the miss is a permanent case rather than a remembered incident.

Improvement cycle · five stages

Switchback — the path turns at Improve and returns at Learn

  1. 01

    Detect

    Unsourced attributes concentrate in one taxonomy.

  2. 02

    Diagnose

    First the published values are read back against their documents, then the taxonomy version that carried them, until one of the two accounts for it.

  3. 03

    Improve

    Version-stamped, with the listings that failed attached to it.

  4. 04

    Verify

    Release is conditional on the affected cases passing.

  5. 05

    Learn

    It joins the suite and the attribute standard changes with it.

  6. Learn → Detect

    The return edge. Each pass begins with more cases in the suite than the one before.

Typical build scope

Twelve workstreams across six weeks

The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, sources, enrichment workflow, evaluation, integration, then production validation and handover.

Workstream

Week 1Week 2Week 3Week 4Week 5Week 6

  1. 01Catalogue workflow discovery and boundary definition.Week 1
  2. 02PIM and supplier-feed assessment.Week 1, Week 2
  3. 03Taxonomy, controlled-vocabulary and jurisdiction mapping.Week 1, Week 2
  4. 04Product-record and source-document.Week 2
  5. 05Enrichment logic and source binding.Week 2, Week 3
  6. 06Confidence scoring and approval routing.Week 3
  7. 07Category and compliance.Week 3, Week 4
  8. 08PIM, syndication and marketplace integration.Week 5
  9. 09Attribute regression cases.Week 4, Week 5
  10. 10Guardrails and publication.Week 4
  11. 11Source-trail instrumentation.Week 5, Week 6
  12. 12Deployment, documentation and Agent Care handover.Week 6

12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel

Final scope and sequence confirmed in discovery

Engagement tiers

What each tier includes

Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.

Capability✓ in scope · — not at this tierPilotOne catalogue, one channelProductionProduction catalogue volumeAdvancedMultiple channels / markets
Introduced at Pilot
Enrichment against your sources✓✓✓
Category and compliance approval✓✓✓
Attribute-accuracy baseline✓✓✓
Introduced at Production
Reporting by taxonomy—✓✓
Approval workflow in your systems—✓✓
Approved catalogue write-back—✓✓
PIM integration—✓✓
Introduced at Advanced
Multi-jurisdiction listing rules——✓
Multi-stage catalogue approvals——✓
High SKU volume——✓
Enterprise catalog controls——✓
Build priceFrom $5,000From $8,000Custom quote
Final build price
Confirmed after discovery based on integrations, workflow complexity, transaction volume, approval controls and deployment requirements.
Separate from build
Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.

What we need from you

What you bring, and what we build with it

Each input maps to a piece of build scope and a week in the delivery timeline.

  1. 01Your taxonomy and its controlled values

    Taxonomy and controlled-vocabulary mapping

    Week 1

  2. 02Representative published listings

    Enrichment baseline, normalisation and source binding

    Week 2

  3. 03Your supplier declarations and certificates

    Source-document mapping and automation-boundary definition

    Week 1

  4. 04Access to relevant APIs, feeds or exports

    PIM, syndication and compliance assessment, then integration setup

    Week 2

  5. 05Listings that had to be pulled

    Attribute cases and the evaluation suite

    Week 4

  6. 06Which attributes may never be generated

    Source thresholds, publication routing and controls

    Week 3

  7. 07A named category manager to approve attributes

    Category and compliance approval workflow, then pilot and production validation

    Weeks 5–6

Nothing else is required

Deployment, documentation and Agent Care handover are ours.

Delivery timeline

Four phases across six weeks

Each phase occupies its own weeks, with a genuine overlap in the fifth rather than a padded one.

Phase

W1

W2

W3

W4

W5

W6

  1. DiscoveryW1
  2. BuildW2 – W3
  3. EvaluateW4 – W5
  4. Pilot & LaunchW5 – W6

Week focus

W1Catalogue workflow discovery, taxonomy mapping and the automation boundary

W2Supplier content sources in place

W3Enrichment logic, confidence scoring and publication controls

W4Attribute testing and publication guardrails

W5Syndication integration, pilot listings and targeted corrections

W6One catalogue release approved by the category manager, then handover

Reading the band
Each bar is its named weeks. The fifth is doubled because the work is.
At the end of W6
Validated on a real release, then Agent Care carries it.
Duration
Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · Retail AI agent

Build a catalogue agent around the documents your attributes have to match.

Show us your taxonomy, your source documents and who signs a regulated field. If your certificates, declarations and test reports sit in systems that do not agree, that is the first thing we map, before any attribute is written.

Typical delivery
4–6 weeks
Deployment
Your PIM and channels
Compliance sign-off
Source-bound
After launch
Agent Care available

Nestack Agents · Catalogue enrichment

AGT-RT-14 · Agent Care available after launch

Build a catalogue agent

Four questions so the call starts from your actual setup — then pick a time that suits you.

About you
Your catalogue

No obligation. We’ll map the workflow, name the automation boundary and tell you if an agent is the wrong answer.