Adobe Commerce B2B Features: What Magento Gives You Natively, and Where Custom Work Starts
Category
Artificial Intelligence
What is generative AI in retail and eCommerce?
Generative AI in retail and eCommerce is a class of model that produces new content – text, images, structured data, or code – from a prompt plus your own business context. In a store, that context is your product catalog, your brand guidelines, your policy pages, and your customer history. The model doesn’t retrieve a stored answer. It composes one.
That distinction matters commercially, because it changes what you’re buying. You aren’t buying a feature. You’re buying a pipeline: data in, generation, human review, publish, measure.
Generative, predictive, and agentic AI – the split that decides your budget
Teams conflate these three and then wonder why quotes vary by 10x. They’re different builds with different risk profiles.
| Type | What it does | eCommerce example | Build effort |
|---|---|---|---|
| Predictive AI | Scores or forecasts from historical data | Demand forecasting, churn scoring, dynamic repricing | Low-medium |
| Generative AI | Creates new text, images, or structured output | Product descriptions, lifestyle imagery, guided-selling replies | Medium |
| Agentic AI | Plans and executes multi-step actions | Re-ordering stock, running a returns workflow end to end | High |
Predictive AI is where our eCommerce pricing intelligence work sits. Agentic AI is a separate discipline covered in our breakdown of agentic commerce. This guide stays on the generative layer in the middle – the one most brands can ship this quarter.
Why retail and eCommerce need different answers
Physical retail applies generative AI to store operations: planogram drafting, associate assistants, localized signage copy. eCommerce applies it to the catalog and the session – content at SKU volume, and conversation at traffic volume.
The difference is throughput. A retail chain generates a few thousand assets a season. A 40,000-SKU eCommerce catalog with six locales and three marketplace channels generates hundreds of thousands. Anything that requires manual review of every output collapses at that scale, which is why our builds always start with a review-sampling rule, not a review-everything rule. If you want the broader retail picture, our post on AI in retail covers the store-side use cases.
The Generative AI Value Ladder: four tiers, four budgets
We use this ladder in every discovery call because it stops the conversation from drifting into a science project. Each rung has a different integration depth, a different failure risk, and a different payback window.
| Tier | What you build | Integration depth | Where value shows up |
|---|---|---|---|
| 1 – Content Operations | Catalog copy, meta, alt text, ad variants, email blocks | Read from PIM, write back via API | Faster launches, SEO/AEO coverage |
| 2 – Merchandising & Discovery | Semantic search, attribute extraction, collection curation, comparison content | Storefront API + search layer | Findability, add-to-cart rate |
| 3 – Conversational Commerce | Guided selling, sizing and fit advisors, support deflection | Session state + order data + CRM | Conversion rate, support cost |
| 4 – Agentic Operations | Multi-step workflows across ERP, 3PL, marketing | Write access to systems of record | Operating margin, headcount efficiency |
Tier 1 – Content operations
This is where we recommend most brands start, because the failure mode is cheap. A bad product description gets caught in review. A bad agentic action ships a pallet to the wrong warehouse.
The build is unglamorous and specific: pull structured attributes from your PIM or platform, feed them into a prompt chain that carries your brand voice rules and category-level constraints, generate copy in your required schema, and write it back. The interesting engineering is in the constraints, not the generation.
If you want the tool-level version of this before committing to a build, our guide to ChatGPT for eCommerce covers what you can do manually first.
Tier 2 – Merchandising and discovery
Here the model reads rather than writes for the shopper. It extracts attributes your catalog never had, powers semantic search that understands “quiet compressor for a home garage,” and drafts comparison content between similar SKUs.
This tier is where AI product discovery starts influencing revenue directly, and where feed quality becomes non-negotiable. If your attributes are thin, incomplete data can reduce product visibility and recommendation quality on AI-driven surfaces, including Google Shopping AI Mode.
Tier 3 – Conversational commerce
Guided selling, fit and sizing advisors, and support deflection. The technology is well understood; the risk is that the model confidently states a spec that isn’t true. Retrieval-augmented generation against your actual product data is the standard guardrail. We go deeper on this in our guide to AI chatbots for eCommerce and in our AI chatbot development work.
Tier 4 – Agentic operations
Multi-step execution across systems – reordering, exception handling in returns, marketing workflow orchestration. High payoff, high blast radius. We only recommend this tier once a brand has clean integration plumbing and observable logs, which usually means the ERP and platform work is already done.
Which rung you belong on is usually obvious within an hour of looking at your catalog and your integration surface, and it is the first thing our AI consulting team establishes before scoping anything.
Generative AI use cases in eCommerce, and the foundations we’ve built
The generative AI use cases in eCommerce that survive contact with a real catalog tend to be narrow, integrated, and boring in the best way. Below are the five that dominate production deployments, alongside builds from our own engagements. Some of those builds are generative. Others are the structured foundations a generative layer has to sit on, and we have labeled which is which rather than blurring the two.
How is generative AI being used in eCommerce right now?
Right now, most production usage falls into five buckets: catalog content generation, visual asset pipelines, guided selling and recommendation, semantic search, and support deflection. Merchandising automation and agentic workflows are growing but sit behind the others in maturity. Below is what each looks like when it’s plugged into a real store rather than a demo.
Catalog content at scale
Generating 40,000 descriptions is trivial. Generating 40,000 descriptions that respect category compliance language, carry your voice, avoid near-duplicate output, and land in the right PIM field is a system.
The parts that matter: attribute completeness scoring before generation, template constraints per category, a similarity check against your existing corpus so you don’t self-cannibalize, and a sampling review. Our own rule of thumb is 5-10% human-reviewed, weighted toward high-revenue SKUs, though the right rate depends on category risk. Our eCommerce technical SEO team runs the similarity check, because near-duplicate or low-value copy is where visibility and canonicalization problems start.
Visual asset pipelines
This is where we’ve shipped some of our most concrete work. For Fine Art Canvas, we built a Shopify Plus custom app that automates Amazon-ready image creation and CSV generation, removing a manual production step that scaled badly with variant count. We paired that with a separate engagement covering Google Shopping feed automation, 3D product rendering, and variant mapping.
The lesson from both: generation is the small part of the work. In our experience the bulk of the effort goes into variant logic, naming conventions, and channel-specific export rules. That’s the part agencies skip in the pitch and discover in week three.
Guided selling and recommendation
For Snowy Owl Cove we built a custom skincare quiz with a dynamic recommendation engine – a structured guided-selling flow that maps customer inputs to product logic.
Worth being precise: that engine is rules-and-data driven, not LLM-generated. We’re citing it because it’s the exact substrate a generative layer plugs into. A conversational advisor with no structured recommendation logic underneath produces charming answers and wrong products. Build the logic first, add the language layer second.
Search that understands intent
For Outdoor Limited we delivered a BigCommerce transformation with custom faceted filtering, AI search, and review aggregation. For Anderson Pens we built custom ink and nib comparison tools so shoppers could evaluate technical products without leaving the site.
Semantic search is the highest-ROI Tier 2 build for catalogs where shoppers describe problems rather than products – industrial parts, technical apparel, specialty consumables.
Support, service, and post-purchase
Deflection is the obvious win, but the durable version is grounded in order data. For CareNovex we built an AI-powered patient intake portal on WordPress and WooCommerce – a regulated-environment build where accuracy wasn’t optional. For Inside Injuries we modernized an AI sports injury intelligence platform using OpenAI with a Node and React stack.
Neither is an eCommerce store, and we are citing them for the engineering lesson rather than the vertical. Both taught the same thing: model choice was the easy decision. Data contracts, latency budgets, and fallback behavior when the API times out took the engineering time, and that holds identically on a storefront.
What generative AI looks like on your platform
Platform matters more than vendors admit. In our own engagements the same use case has run to weeks on one stack and months on another, and the difference is almost always integration surface rather than anything to do with the model.
| Platform | Where generative AI plugs in | Practical constraint |
|---|---|---|
| Shopify / Shopify Plus | Custom or public apps, Shopify Functions for logic, Hydrogen for custom storefronts, metafields for generated attributes | Shopify Functions are broadly available, but major checkout UI customization is Plus-only, so generation belongs pre-checkout |
| BigCommerce | Custom apps, GraphQL Storefront API, Stencil or headless frontends, catalog API for write-back | Strong API surface for catalog automation – our most common Tier 1 and 2 platform |
| Adobe Commerce (Magento) | Custom modules, GraphQL, Hyva frontends for performance headroom | Heavier build. Our engineering rule on current 2.4.x releases is to run generation as an async, event-driven service rather than an in-request call |
| WooCommerce | REST API, custom plugins, WP-Cron or external queue for batch jobs | Cheapest to start, hardest to scale – batch generation needs an external worker |
If you’re deciding where to run this, our eCommerce development and eCommerce AI teams size the same use case across platforms during discovery so you see the cost delta before committing.
How to implement generative AI in eCommerce in 90 days
The question of how to implement generative AI in eCommerce usually gets answered with a maturity model. Here’s a sequence instead – the one we run.
Days 1-15: Data readiness, not model selection
Score your catalog for attribute completeness by category. Pick one category where completeness is high and revenue is meaningful. Our working threshold is 80%, but set your own based on how much your category tolerates gaps. Write down your brand voice rules as constraints a model can follow – banned words, sentence length, claim boundaries, required disclosures.
If your data fails this step, stop. No model fixes a catalog with large attribute gaps. That is a PIM project, and it is a better use of the budget.
Days 16-45: Ship one Tier 1 build end to end
One category. One output type. Full round trip: read from source, generate, review, write back, publish. Instrument it before you scale it.
The output that matters at day 45 isn’t quality – it’s throughput and cost per unit. Quality you can tune. An economically broken pipeline you can’t.
Days 46-75: Measure against a real control
Hold out a control set. Compare organic impressions, click-through, add-to-cart rate, and return rate between generated and existing content. Return rate is the one teams forget, and it’s the one that exposes descriptions that overpromise.
Days 76-90: Scale, or kill and move up the ladder
If unit economics work, expand to adjacent categories. If they don’t, the honest move is often to skip to Tier 2 – semantic search frequently outperforms content generation on catalogs where shoppers can’t find products in the first place.
We run this sequence through our AI consulting engagement, and the deliverable at day 90 is a decision document, not a slide deck.
The cost model nobody puts in the pitch deck
Model API pricing changes constantly, so here’s the arithmetic rather than a number that will be stale by the time you read it.
Formula – Per-SKU generation cost = (input tokens × input rate) + (output tokens × output rate). Input and output are billed at different rates by many major providers, so a single blended rate can misstate the cost of output-heavy tasks.
A typical product description prompt carries brand rules, category constraints, and product attributes – call it 1,200 input tokens – and returns roughly 400 output tokens. Across a 10,000-SKU catalog that is 12 million input tokens and 4 million output tokens on a single pass. Price those two figures separately, because output tokens usually cost several times more than input.
Plug in your provider’s current input and output rates and you’ll usually find inference is the smallest line item. The real costs sit elsewhere:
- Integration engineering – reading from and writing back to your platform, queue management, retry logic
- Review workflow – the human hours on your sampling rate, which is a permanent operating cost, not a one-time build
- Re-generation – you will re-run passes as prompts improve; we typically budget several passes in the first year
- Monitoring – drift detection, cost alerting, output logging
A useful rule from our engagements: if someone quotes you generative AI work and inference is the biggest line in the estimate, they haven’t built the integration yet.
Where generative AI breaks – and the guardrails that hold
| Failure mode | What it looks like | Guardrail |
|---|---|---|
| Spec hallucination | Model states a dimension, material, or compatibility that isn’t true | RAG against your PIM; hard-block generation on any field the source data doesn’t contain |
| Brand voice drift | Copy reads generic across thousands of SKUs | Voice rules as explicit constraints plus a similarity threshold against approved samples |
| Near-duplicate content at scale | Near-identical descriptions across variants reduce visibility and create canonicalization problems | Similarity scoring pre-publish; variant-aware templates; make each page earn its place |
| Data leakage | Customer or pricing data sent to a third-party API | Data classification before the pipeline; redaction layer; self-hosted models where policy requires |
| Silent cost blowout | A retry loop burns budget overnight | Per-job token ceilings, cost alerting, circuit breakers |
| Agentic overreach | A workflow takes an irreversible action on bad input | Human approval gates on any write to a system of record |
None of this is exotic. It’s the same discipline behind zero-downtime migrations and the 92% conversion rate increase we delivered for Parts Connexion after their BigCommerce migration and custom integrations – instrument it, gate it, measure it, then scale it.
Frequently Asked Questions
Generative AI in retail and eCommerce is technology that creates new content – product copy, images, structured attributes, or conversational replies – from your catalog and brand data rather than retrieving pre-written answers. In retail it supports store operations and merchandising. In eCommerce it operates at catalog and session scale, generating assets and interactions on demand.
Five uses dominate production deployments: catalog content generation, visual asset pipelines for marketplace and ad channels, guided selling and recommendation, semantic search, and support deflection grounded in order data. Merchandising automation and agentic workflows are growing but require cleaner system integration, which in our experience is why most brands reach them well after their first build rather than in their first quarter.
Small catalogs get more value from Tier 2 and Tier 3 work than from bulk content generation. With a few hundred SKUs, a guided-selling advisor or semantic search build usually moves conversion rate more than rewriting descriptions. As a rule of thumb, the economics flip somewhere in the low thousands of SKUs, where manual content production becomes the bottleneck.
Inference is rarely the main cost. Budget instead for integration engineering, the ongoing human review workflow, re-generation passes as prompts mature, and monitoring. A scoped Tier 1 pilot on a single category is a defined, contained engagement. Tier 4 agentic work is a materially larger investment and shouldn’t start until integration plumbing is proven.
Google evaluates quality and usefulness, not authorship method, and duplicate content on its own is not a penalty. The real risk is content mass-produced at scale without adding value for the reader, whether or not a human reviewed it. Near-identical variant descriptions and unsupported claims are the common failure. Run similarity scoring before publish and hold a control set.
BigCommerce and Shopify Plus give the cleanest API surface for catalog automation. Adobe Commerce offers the most control but is best paired with async, event-driven generation rather than in-request calls, particularly alongside a Hyvä frontend. WooCommerce is the cheapest to start and needs an external worker queue to scale past a few thousand SKUs.
A chatbot is one application of generative AI, focused on conversation. Generative AI covers a wider set of outputs including product copy, imagery, structured attributes, and code. A chatbot without grounded product data will invent specs, which is why the retrieval layer matters more than the conversational interface.
Most eCommerce brands never need a custom-trained model. Retrieval-augmented generation against your own data, combined with well-constructed prompt constraints, covers the large majority of use cases. Fine-tuning becomes worth it when brand voice is highly distinctive or when a narrow, repeated task justifies the training and maintenance overhead.
On the 90-day sequence above, which is our own framework rather than an industry standard, a single-category Tier 1 pilot produces throughput and cost data by around day 45 and comparative performance data by around day 75. Conversion-side builds take longer to read because they need traffic volume first.
Yes – voice agents are a generative build with stricter latency requirements and different failure handling, since a shopper can’t scroll back through a spoken answer. Our AI voice agent development team handles sales and support voice deployments, and the grounding requirements are the same as for chat.
Final Thoughts
Generative AI in eCommerce pays off when it is scoped to one tier, wired into the platform you actually run, and measured against a control group. It fails when it is bought as a feature and bolted on afterwards.
The pattern across our own engagements is consistent. The brands that see real returns are not the ones that picked the best model. They are the ones that fixed their product data first, shipped one narrow build end to end, and only scaled after the unit economics held up.
That order matters more than any tooling decision you will make this year.
If you are working out which tier fits your catalog and your integration surface, our AI development services team runs a readiness assessment as the first step of every generative AI development engagement – catalog quality, integration depth, and the tier your economics actually support.
