Retail and e-commerce · Playbook 23

I've worked in this sector

Product content at scale

New lines go from supplier data sheet to accurate, accessible listing on the same day, with every claim traced back to the spec. A person approves every listing before it goes live.

  • Document AI
  • Computer vision
  • Agents with tools
  • Human in the loop
  • DMCC Act 2024
  • CMA
  • Product safety (OPSS)
  • Energy labelling
  • Equality Act 2010
  • Copyright

At a glance

The problem
Every new line means reading a supplier's data sheet, editing photos and writing the listing by hand, and mistakes reach customers
Typical volume
Around 8,000 live SKUs, 400 new lines a month, and a steady stream of supplier spec changes
Who's involved
E-commerce, the content team, buying, product compliance, the studio, IT, customer service
Pilot
6–8 weeks on one category, after a 2–3 week discovery
Platforms
AWS, Azure, or a sovereign or on-prem stack

An illustrative scenario. Linked to a real, anonymised case study below; the numbers here are still typical examples. The cloud services and UK rules were checked on 9 November 2026.

The business problem

A new line arrives as a supplier's PDF data sheet, a spreadsheet row, a folder of photos and sometimes an energy label. Someone on the content team reads it all, works out which attributes the category needs, types them into the PIM, writes a title and description to the house style, crops and cleans the photos, writes alt text, and checks the safety warnings and energy class are right. In our example retailer, a straightforward line takes an hour or two. A complex one, like a TV or a washing machine, takes longer.

So new lines queue. Buyers chase the content team, stock sits in the warehouse before it can be sold, and the pressure shows up as shortcuts: copied descriptions, missing attributes, alt text that just repeats the product name, or a spec carried over from the previous model. Supplier data is often wrong or incomplete too, and nobody has time to check it against the photos.

Accuracy here is a legal requirement, not a matter of style. Since 6 April 2025, the DMCC Act 2024 lets the CMA decide for itself that consumer law has been broken and fine a retailer directly. A listing that overstates a feature, leaves out something a customer needs to know, or shows the wrong energy class is a misleading action or omission, whoever typed it. AI can write listings quickly. The design has to make sure it never writes anything the supplier's data doesn't support.

What AI changes

  • Supplier data sheets, spreadsheets and photos are read on arrival, and every attribute the category needs is extracted with a link to where it came from.
  • Photos are cropped, resized and put on a consistent background automatically, and checked against the spec: right colour, right model, right number of items in the box.
  • Titles, descriptions, bullet points and alt text are drafted in the house style, using only facts from the supplier's data. Anything the data doesn't say is left out and flagged, not guessed.
  • Consistency checks compare every draft with the source: numbers, units, energy class, warnings and compliance marks. Mismatches go to a person with the reason.
  • A content editor approves every listing before it's published, and listings in safety-critical categories also go to product compliance.
  • When a supplier sends an updated spec, the live listing is compared with it and any difference is raised for review, so listings don't drift out of date.

How success is measured

Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.

Time from goods-in to live listing

How long new stock waits before it can be sold.

ExampleFrom days to the same day

Attribute accuracy

Share of attributes that match the supplier's spec, measured on a weekly audit sample, not self-reported.

ExampleAbove 99%

First-pass approval rate

Share of drafts an editor approves with no or minor edits.

Example70–80% after the pilot

Edit and override rate

How much editors change. Very high means it isn't helping; near zero can mean drafts are being approved unread.

ExampleStable, and reviewed weekly

'Not as described' returns

Returns and complaints where the listing didn't match the product.

ExampleFalling, measured against last year

Images with meaningful alt text

Share of product images whose alt text describes what a customer would need to know.

ExampleEvery image on every new listing

UK regulation and controls

Cross-sector rules are on the UK regulation map.

  • CMADMCC Act 2024: misleading actions and omissions (in force 6 April 2025)

    A listing must not mislead and must not leave out information a customer needs to decide. The CMA can now enforce directly and fine up to 10% of global turnover. Every claim in a draft is traced to a field in the supplier's data, and the AI is blocked from adding features, performance claims or environmental claims the data doesn't contain.

  • CMABanned practices, including fake reviews, and the CMA's guidance (CMA207, CMA208)

    Fake and concealed incentivised reviews are banned outright, along with reviews presented in a misleading way. The AI never writes, edits or selects reviews. If review summaries are added later, they summarise all reviews fairly, and that's a separate decision with its own review.

  • OPSS and Trading StandardsProduct safety: the General Product Safety Regulations 2005, being replaced under the Product Regulation and Metrology Act 2025

    Safety warnings and instructions are copied exactly from the supplier's material, never reworded by the AI. The government consulted in 2026 on requiring listings to show producer details, a product identifier, images, warnings and compliance marks, so these are held as structured attributes now. If you sell into Northern Ireland or the EU, the EU General Product Safety Regulation already asks for much the same.

  • Trading StandardsEnergy labelling: the Ecodesign for Energy-Related Products and Energy Information Regulations 2021

    For labelled electricals, the online listing has to show the energy class arrow linked to the full label, and the product information sheet. These come from the supplier's label and database entry, never from the AI, and a listing without them can't be approved.

  • EHRCEquality Act 2010: reasonable adjustments

    Product pages need to work for disabled customers, and WCAG 2.2 AA is the usual benchmark. Alt text describes what matters about the product, not just its name. If you sell to EU consumers, the European Accessibility Act has applied to e-commerce since 28 June 2025.

  • UK lawCopyright, Designs and Patents Act 1988

    Supplier photos and copy are licensed, not owned. The design keeps a licence record for every image, limits edits to what the licence allows, and never uses supplier content to train a model without permission. Background removal and cropping are fine; generating new product detail isn't.

The solution

Showing AWSAzuresovereign / on-prem services. Switch with the toggle; every diagram and table on this page follows it.

Product content at scale: reference architecture

  1. 01Channels

    • Supplier data sheets and feeds
    • Supplier and studio photos
    • New-line requests from buyers

  2. 02Intake

    • API Gateway + AWS Transfer FamilyAPI Management + Blob Storage SFTPKong, or your existing API gatewaySupplier feeds and uploads
    • Amazon S3Blob StorageCeph, or IONOS S3 Object StorageFile and image store
    • AWS Lambda with open-source image toolsAzure Functions with open-source image toolsOpen-source image tools and segmentation models on KubernetesImage processing

  3. 03Intelligence

    • Bedrock Data AutomationContent UnderstandingDocling plus an open-weight vision modelRead spec sheets and photos
    • Agent
      Bedrock AgentCoreFoundry Agent ServiceLangGraph on Kubernetes, with open-weight models via vLLMListing agent
    • Agent
      Bedrock AgentCoreFoundry Agent ServiceLangGraph on KubernetesSpec change watcher

  4. 04Orchestration

    • Step FunctionsLogic Apps / Durable FunctionsTemporalListing workflow and checks
    • EventBridge + SQSEvent Grid + Service BusApache Kafka or RabbitMQEvents and queues

  5. 05Systems of record

    • PIM
    • E-commerce platform
    • ERP and purchasing

  6. 06People

    • Person
      Content editor: approves every listing
    • Person
      Product compliance: safety and energy
    • Person
      Buyers: chase supplier data gaps
  • Cloud or platform service (AWSAzuresovereign / on-prem)
  • AI agent
  • A person decides
  • Your existing systems
  • Outside your organisation

Components, side by side

Product content at scale: components on AWS, Azure and a sovereign or on-prem stack
LayerAWSAzureSovereign / on-prem
Intake: feeds, uploads, filesSuppliers send data in every shape: PDFs, spreadsheets, feeds and the occasional zip of photos. Accept them all and keep the original.API Gateway, AWS Transfer Family, S3API Management, Blob Storage (with SFTP), Event GridKong or your existing API gateway; Ceph or IONOS S3 Object Storage
Reading spec sheets and photosTest on real supplier sheets, including scanned ones and ones in other languages, not the tidy examples.Bedrock Data Automation (documents and images)Azure Content Understanding (includes Document Intelligence)Open-source parsing (e.g. Docling) plus an open-weight vision model
Image processingCrop, resize, straighten and remove backgrounds. Nothing is added to the product itself, so the photo still shows what the customer will get.AWS Lambda or ECS with open-source image toolsAzure Functions or Container Apps with open-source image toolsOpen-source image tools and segmentation models on Kubernetes
Drafting and reasoningBoth clouds offer a choice of models, including Claude. Drafts use only the extracted facts and the house style guide.Amazon BedrockMicrosoft FoundryOpen-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub
Category rules and style guide (retrieval)Which attributes each category needs, banned phrases, and how to write titles. Owned by the content team, versioned like code.Bedrock Knowledge BasesFoundry IQ, built on Azure AI SearchOpenSearch, or PostgreSQL with pgvector
Agent runtimeThe original Bedrock Agents closed to new customers on 30 July 2026; new builds use AgentCore.Amazon Bedrock AgentCoreFoundry Agent ServiceAn open-source agent framework (e.g. LangGraph) on Kubernetes
Consistency checksPlain rules do most of the work: every number, unit, energy class and warning in the draft must match the source exactly. The model-based check catches claims the rules can't see.Rules in Lambda, plus Bedrock Guardrails contextual grounding checksRules in Azure Functions, plus Foundry guardrails (Azure AI Content Safety) groundedness detectionA rules service on Kubernetes, plus open-source guardrails (e.g. NeMo Guardrails)
WorkflowAWS Step FunctionsLogic Apps or Durable FunctionsTemporal, or n8n
Middleware and eventsAmazon EventBridge, SQSEvent Grid, Service BusApache Kafka or RabbitMQ
Link to the PIM and e-commerce platformThe AI writes drafts into the PIM. Only the PIM's own approval step publishes to the site.Through their published APIs, behind API GatewayThrough their published APIs, behind API ManagementThrough their published APIs, behind Kong or your existing gateway
Human reviewEditors review where they already work, with the source document shown next to the draft.In the PIM's existing approval workflowIn the PIM's existing approval workflowIn the PIM's existing approval workflow
Identity, keys, data protectionThe agent's role can create drafts but has no permission to publish.IAM Identity Center, KMSEntra ID, Key VaultKeycloak or your existing Active Directory; OpenBao or HashiCorp Vault
Monitoring and auditKeep the source file, every draft, every edit and the approver for each listing. That's your evidence if a listing is ever challenged.CloudWatch, CloudTrailAzure Monitor, Log AnalyticsPrometheus, Grafana, OpenTelemetry
Data residencyProduct data is rarely personal data, so EU hosting (IONOS AI Model Hub runs in Germany) is often acceptable. Check supplier contracts for any restrictions.eu-west-2 (London)UK South / UK WestYour own UK data centre, or a UK or EU provider

On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →

Hybrid and on-prem

Many organisations keep core systems and sensitive data on their own servers. This is how the same design works across both.

Stays on-prem

The PIM, the ERP and the studio's master image files stay where they are, whether that's your own servers or a hosted service you already use. The approved, published listing always lives in your systems, not the AI's.

Uses cloud AI

Reading data sheets, image processing, drafting and the consistency checks run in AWS or Azure in UK regions, or on your own GPUs with open-weight models. Only product data and images are sent, which rarely include personal data.

How they connect

Supplier files land in cloud storage, and drafts go back to the PIM through its API over a private link (AWS Direct Connect or Azure ExpressRoute) or a secured internet connection, depending on where the PIM runs.

If the link drops

Supplier files queue safely, and editors keep writing listings by hand in the PIM. Nothing live changes, because the AI can't publish. The queue is drafted when the link returns.

The agents

What each agent may do through its tools, what it may not, and where a person signs off. The limits are enforced by the platform, not the prompt.

Listing agent

May do

  • Read the supplier's data sheet, spreadsheet and photos
  • Extract attributes and record where each one came from
  • Look up the category's required attributes and the style guide
  • Draft the title, description, bullets and alt text
  • Run the consistency checks and write the draft into the PIM
  • Raise a data gap to the buyer when the supplier's data is missing or contradicts itself

May not do

  • Publish, or change a live listing
  • Set or change a price or promotion
  • Reword safety warnings, or enter energy class or compliance marks from anywhere but the source
  • Add claims the supplier's data doesn't support, including environmental claims
  • Write, edit or choose customer reviews
  • Generate or alter what the product looks like in a photo

Where a person signs offA content editor approves every listing before it's published. Electricals, children's products and other safety-critical categories also need product compliance sign-off. The agent's permissions are enforced by the PIM's roles and the platform's identity controls, not by instructions in a prompt.

Spec change watcher

May do

  • Compare a supplier's updated data sheet with the live listing
  • Flag differences in specs, warnings, energy class or what's in the box
  • Draft the corrected listing for review
  • Raise a task to the content editor, with the differences highlighted

May not do

  • Change or unpublish a live listing
  • Contact the supplier or the customer
  • Decide whether a change matters for safety

Where a person signs offA content editor decides every change. Anything touching safety warnings or energy labelling goes to product compliance, who also decide whether stock already sold needs action.

Middleware and integration

The PIM stays the single source of truth for product content, and the e-commerce platform reads from it as it does today. The AI writes drafts into the PIM through its API and never talks to the website directly. If you already run an integration platform, the AI uses it like any other system does.

Each new supplier file becomes an event on a queue. If the PIM is slow or down, files wait safely and nothing is processed twice: each one carries the supplier, the SKU and a version, so a retry updates the same draft rather than creating a second.

The ERP and purchasing system supply the SKU, supplier and category when a buyer raises a new line, so the content work starts the moment a line is approved, not when the stock lands.

Switching the AI off doesn't stop trading. Editors carry on creating listings by hand in the PIM, exactly as before.

Process and people

  • A house style guide and a list of required attributes for each category, owned by the content team and changed under version control.
  • A banned-claims list agreed with product compliance, including environmental and performance claims that need evidence before they can be made.
  • A data-gap queue owned by buying, so missing or contradictory supplier data goes back to the supplier rather than being guessed.
  • A weekly audit of a sample of approved listings against the source documents, including every safety-critical category.
  • Editors trained to review against the source, not just to read the draft, and to record why they changed something.
  • A licence record for supplier images and copy, checked when a new supplier is onboarded.

Delivery plan

  1. 2–3 weeks

    Discovery

    Map today's listing process, time it, collect a month of real supplier files, and agree the categories, style guide and banned-claims list.

  2. Gate · must be true to passBaseline measured, PIM integration route agreed, product compliance and the content team involved.

  3. 4–6 weeks

    Pilot

    One category, such as small kitchen appliances, in shadow mode: the AI drafts, editors write as normal, and the two are compared against the source.

  4. Gate · must be true to passAttribute accuracy and approval rate within agreed limits, no unsupported claims in the sample, compliance sign-off.

  5. 3–4 weeks

    Production

    Drafts go live for the pilot category through the PIM's approval step, with monitoring, the weekly audit and a rollback plan.

  6. Gate · must be true to passMonitoring, audit, rollback plan and editor training in place.

  7. Ongoing

    Scale

    Add categories one at a time, with labelled electricals and children's products last, then the spec change watcher across the whole range. Food or drink would need allergen information shown before purchase and at delivery, so it gets its own gate.

  8. Gate · must be true to passEach new category passes the same pilot gate.

Roles and RACI

Who is responsible, accountable, consulted and informed, by activity
ActivityE-commerce directorHead of contentSolutions architectI can fill thisTechnical PMI can fill thisEngineersI can fill thisProduct complianceSecurityI can fill this
Business case and benefitsAccountableResponsibleConsultedConsultedInformedConsultedInformed
Target architecture and PIM integrationInformedConsultedAccountable and ResponsibleConsultedConsultedConsultedConsulted
Style guide, required attributes, banned claimsInformedAccountable and ResponsibleConsultedInformedConsultedResponsibleInformed
Build and integrationInformedConsultedAccountableResponsibleResponsibleInformedConsulted
Accuracy testing on real supplier filesInformedAccountableResponsibleConsultedResponsibleConsultedInformed
Go-live decisionAccountableResponsibleConsultedResponsibleInformedConsultedConsulted
Run, monitor and improveInformedAccountableConsultedConsultedResponsibleConsultedConsulted

Business case and benefits

E-commerce director
A · Accountable
Head of content
R · Responsible
Solutions architect
C · Consulted
Technical PM
C · Consulted
Engineers
I · Informed
Product compliance
C · Consulted
Security
I · Informed

Target architecture and PIM integration

E-commerce director
I · Informed
Head of content
C · Consulted
Solutions architect
A/R · Accountable and Responsible
Technical PM
C · Consulted
Engineers
C · Consulted
Product compliance
C · Consulted
Security
C · Consulted

Style guide, required attributes, banned claims

E-commerce director
I · Informed
Head of content
A/R · Accountable and Responsible
Solutions architect
C · Consulted
Technical PM
I · Informed
Engineers
C · Consulted
Product compliance
R · Responsible
Security
I · Informed

Build and integration

E-commerce director
I · Informed
Head of content
C · Consulted
Solutions architect
A · Accountable
Technical PM
R · Responsible
Engineers
R · Responsible
Product compliance
I · Informed
Security
C · Consulted

Accuracy testing on real supplier files

E-commerce director
I · Informed
Head of content
A · Accountable
Solutions architect
R · Responsible
Technical PM
C · Consulted
Engineers
R · Responsible
Product compliance
C · Consulted
Security
I · Informed

Go-live decision

E-commerce director
A · Accountable
Head of content
R · Responsible
Solutions architect
C · Consulted
Technical PM
R · Responsible
Engineers
I · Informed
Product compliance
C · Consulted
Security
C · Consulted

Run, monitor and improve

E-commerce director
I · Informed
Head of content
A · Accountable
Solutions architect
C · Consulted
Technical PM
C · Consulted
Engineers
R · Responsible
Product compliance
C · Consulted
Security
C · Consulted

R ResponsibleA AccountableC ConsultedI Informed· not involved

Where I'd fit

Option A

Lead the architecture

If there's no architect on the programme, I'd own the design end to end: the pipeline from supplier file to PIM draft, the image processing, the checks and the gates. I've built this kind of workflow before. For an online electronics retailer, I automated the product photography and specification uploads and trained the staff, and work that took a couple of hours came down to about ten minutes. The AI layer in this playbook sits on top of that same kind of pipeline. I'd report to your programme board.

Option B

Support your architect

If your partner, your PIM supplier or your own team already has an architect, I'd take on the document reading, the image pipeline and the consistency checks under their lead, and give them a second senior opinion on the PIM integration and on where the AI is and isn't allowed to write.

Where this goes wrong

  1. Goes wrong

    The model fills gaps in the supplier's data with plausible-sounding features.

    How the design avoids it

    Drafts use only extracted facts, rules check every number and unit against the source, and gaps go to the buyer, not into the copy.

  2. Goes wrong

    Safety warnings get paraphrased into the house style and lose their meaning.

    How the design avoids it

    Warnings, energy data and compliance marks are copied from the source as structured fields, and the AI can't edit them.

  3. Goes wrong

    Editors approve drafts because they read well, without checking the source.

    How the design avoids it

    The source is shown next to the draft, edit rates are tracked, and a weekly audit compares approved listings with the data sheet.

  4. Goes wrong

    Image clean-up drifts into image generation, and the photo no longer shows the real product.

    How the design avoids it

    Image processing is limited to crop, resize, straighten and background removal, and generated product imagery is switched off.

  5. Goes wrong

    Supplier specs change, and the live listing quietly goes out of date.

    How the design avoids it

    The spec change watcher compares every new supplier file with the live listing and raises the difference.

On the horizon

Changes already announced or under way that affect this design. More in the UK horizon scan.

  1. Late 2026–2027
    New product safety regulations under the Product Regulation and Metrology Act 2025. The government consulted from 31 March to 23 June 2026 on replacing the General Product Safety Regulations 2005, with duties for online marketplaces and set information in online listings.

    Hold producer details, product identifiers, warnings and compliance marks as structured attributes now, so a new listing rule is a template change, not a data clean-up.

  2. Now and ongoing
    The CMA's direct enforcement of consumer law under the DMCC Act. It opened its first investigations under the new regime in November 2025.

    Keep the trail for every listing: source file, draft, edits and approver. It's what you'd need to show the CMA or Trading Standards.

  3. 2026–2027
    Copyright and AI. The government's report of 18 March 2026 held back from a new text and data mining exception and pointed towards licensing and transparency, with more evidence to gather.

    Keep a licence record for every supplier image and text, and don't train or fine-tune on supplier content without written permission.

  4. By 28 June 2030
    The end of the European Accessibility Act's transition period for services already on the EU market before June 2025.

    If you sell to EU consumers, build alt text and accessible product pages into every new listing now, rather than retrofitting the range later.

Questions to ask your partner or vendor

Good answers are specific. Vague answers to any of these are worth following up.

Copy them into your next supplier meeting

  1. How does the system stop the model adding a feature that isn't in the supplier's data?
  2. Which fields can the AI never write, and where is that enforced?
  3. How will we measure attribute accuracy on our own supplier files, not a demo set?
  4. Can the AI publish anything, and if not, what stops it?
  5. What happens to live listings when a supplier changes a spec?
  6. What does it cost to run per thousand listings, including image processing?

Planning something like this?

Whether you need someone to lead the design or to support the architect you already have, let's talk it through.

Book a call →How I work alongside your team
Sam Naseem
Sam NaseemBirmingham, UK · UK & Europe
Available now