Skip to content

Yuvraj Raulji | AI Commerce

Most companies asking about AI have a process problem, not an AI problem. Automating a process nobody has fixed gets you the same bad outcome, faster and at higher cost. The applications that hold up in commerce are unglamorous and specific: catalogue enrichment, retrieval-based search, ranking fed by real behaviour, and operations automation.

View relevant work

Quick answer

AI commerce is the application of machine learning and language models to the parts of an online business that are repetitive, high-volume and low-judgement: enriching product data, understanding what a customer means when they search, ranking what to show them, and automating operational work like quotes, approvals and reconciliation. It matters because those tasks scale badly with people and well with software. The business problem AI commerce addresses is throughput on work nobody should be doing by hand. What it does not address is a process that should not exist.

Best suited for

  • Catalogues too large to enrich or QA by hand
  • Repeatable processes with high volume and low judgement
  • Search logs full of queries returning nothing
  • Operations with data reconciliation on a schedule

Business problems

The symptom as the person with the problem describes it, what it is costing, and where the technology actually helps.

  • The operation runs on manual work

    Quotes, approvals, catalogue QA and data reconciliation absorbing people who should be doing something else. The cost is capacity rather than payroll.

    OpportunityThis is where automation pays, provided the process underneath is worth keeping. On a B2B commerce platform this route took 90% of order and quote processing off people and cut approval cycle time by 40%, and the first version was deterministic workflow rather than AI.

  • On-site search does not understand the catalogue

    Customers searching in language the keyword index has never seen, and leaving with no results on products you actually stock.

    OpportunityRetrieval and semantic ranking fix this in a way another synonym list will not, because meaning does not have to be enumerated in advance.

  • Product data is too thin to rank or convert

    Tens of thousands of SKUs with supplier descriptions and no attributes. Filters do not work, search does not work, and the category pages have nothing to say.

    OpportunityEnrichment at that scale is either an AI pipeline with human review or it does not happen at all. Coverage is the metric, and it is countable from day one.

  • AI answers describe the business incorrectly

    Language models increasingly sit between a buyer and a brand, and what they say is now part of how the business is perceived.

    OpportunityContent structure and structured data that let an answer engine read the business correctly. This is a publishing and markup problem rather than a model problem.

Approach

Five stages, in this order, on every engagement. Skipping the first one is the most expensive habit in this industry.

  1. Understand

    Business modelCustomersProductsOperations

    Fix the process first, on paper. Map the current steps and remove the ones that exist because of a system limitation nobody has revisited. Most of the saving is here, and it costs nothing.

  2. Architect

    SystemsIntegrationsCustomer journey

    Choose applications with a measurable edge and attach the number before starting: coverage, no-result rate, conversion on search sessions, hours off the rota. An AI project without a baseline cannot be judged, which is convenient for the vendor and nobody else.

  3. Build

    TechnologyExperienceFunctionality

    Wire it into the real system: the live catalogue, the pricing rules and the permission model, through the platform APIs. An integration that runs beside the platform rather than inside it drifts within a quarter.

  4. Optimize

    PerformanceCROSEOCustomer experience

    Keep a human checkpoint wherever the output touches price, stock or a customer commitment, and review the failure cases on a schedule. Agentic systems that plan and execute need this more than generative ones, not less.

  5. Scale

    AutomationAIAnalyticsPersonalization

    Extend from the application that worked to the one next to it, in the same order every time: data quality, then retrieval, then ranking, then personalisation. Reversing that order personalises bad data, which is worse than doing nothing.

Capabilities

  • Catalogue

    • Attribute enrichment
    • Description drafting
    • Categorisation
    • Image tagging
    • Data QA
  • Discovery

    • Semantic search
    • Query understanding
    • Ranking
    • Recommendations
    • Merchandising rules
  • Operations

    • Quote and approval automation
    • Data reconciliation
    • Order exception handling
    • Supplier data processing
  • Customer experience

    • Pre-purchase questions
    • Order status
    • Returns guidance
    • Support deflection
  • Foundations

    • Retrieval and RAG
    • Structured data
    • Workflow automation
    • Evaluation and review

AI and technology

What each application does, how it works, where the value is, where a person still reviews it, and what it will not do.

  • Catalogue enrichment

    What it does
    Generating the attributes, descriptions and structured data a large catalogue needs and nobody has the staff to write.
    How it works
    A pipeline reads existing product data, supplier feeds and images, drafts missing fields against a fixed schema, and writes back through the platform API in reviewed batches.
    Where the value is
    Coverage. Filters, search relevance and category page quality all depend on attributes existing, and none of them can be fixed while half the fields are empty.
    Where a human reviews
    Review before publish. The failure mode is a confident, fluent, wrong specification, which on a product page is a returns problem and a trust problem.
    What it will not do
    It does not know your products. It works from what it is given, so a supplier feed with wrong dimensions produces enriched wrong dimensions at scale.
  • Retrieval-based search

    What it does
    Search that understands intent rather than matching strings against product titles.
    How it works
    Embeddings over product text and attributes, queried by meaning, with the keyword index kept underneath for SKUs and exact identifiers. The two are blended rather than swapped.
    Where the value is
    No-result rate and conversion on search sessions, both of which exist before the project starts. This is the application to fund first on almost any catalogue.
    Where a human reviews
    A merchandiser reads the top failing queries monthly. Relevance is a business opinion at the edges and the model does not hold one.
    What it will not do
    It cannot sell a product the catalogue describes badly. Retrieval surfaces thin data faster; it does not repair it.
  • Operations automation

    What it does
    Taking repetitive work out of quotes, approvals, catalogue QA and reconciliation.
    How it works
    Workflow automation wired into the platform APIs and the approval model, with retrieval added only where a step must read policy or contract terms rather than follow a fixed rule.
    Where the value is
    Hours off the rota and cycle time on approvals. The measured version of this on a B2B platform was 90% of order and quote processing automated and a 40% cut in approval cycle time.
    Where a human reviews
    Anything touching price, credit or a customer commitment keeps a named approver. The system prepares the decision, it does not make it.
    What it will not do
    It cannot fix a process that should not exist. Automating an unnecessary approval makes it permanent and much harder to remove later.
  • Agents, where they are actually warranted

    What it does
    Systems that plan across several steps and call tools, rather than answering a single question.
    How it works
    Explicit scope, defined tool access, a bounded set of actions, and a review step before anything commits. Most requests described as agents are better served by deterministic workflow.
    Where the value is
    Genuine multi-step tasks where the sequence varies by case: exception handling, supplier data with inconsistent shapes, reconciliation that needs judgement about which side is right.
    Where a human reviews
    A checkpoint before any action that touches money, stock or a customer. This is a design requirement, not a phase-one compromise to be removed later.
    What it will not do
    Reliability falls as the number of steps rises. An agent that is 95% reliable per step is roughly 60% reliable across ten, and commerce operations notice the difference.

Architecture

How the pieces sit together, top to bottom. Only technologies actually used appear here.

  1. Customer

    Searching, browsing and asking questions in their own words.

    StorefrontSupportAI assistants

  2. Experience layer

    Where AI output is actually seen: search results, product pages, recommendations.

    Search UIRecommendationsPDP content

  3. AI services

    Retrieval, enrichment and ranking as services, not as a feature inside one page.

    EmbeddingsVector indexLLM APIsRAG

  4. Commerce platform

    Owns catalogue, pricing, permissions and orders. Everything above must respect it.

    MagentoShopifyCustom platform

  5. Workflow

    Where operations automation runs, with the human checkpoints in it.

    n8nQueuesApproval model

  6. Business systems

    ERP, PIM and CRM, still the systems of record.

    ERPPIMCRM

  7. Evaluation

    The part most projects skip, and the reason they cannot tell whether it worked.

    BaselinesReview queuesCoverage metrics

When to use it

Good fit

  • A repeatable process with high volume and low judgement per instance
  • A catalogue too large to enrich or QA by hand
  • Search logs full of queries returning nothing
  • Data reconciliation that people do on a schedule
  • A defined place for a human to check the output before it matters

Think twice

  • A process nobody has fixed, where automation would make it permanent
  • No baseline measurement, so the result cannot be argued either way
  • Product data too thin for retrieval to have anything to work with
  • An expectation that a chat widget will lift conversion
  • Volume too low for a model to beat a well-chosen manual rule

Comparison

Most commerce automation should start deterministic and only reach for a model where the input is genuinely unstructured. The 90% automation figure on the B2B platform came from workflow and approval modelling, not from AI, and saying so is the point: choosing the more impressive technology when the simpler one fits is how these projects end up expensive and unreliable at the same time.

AI commerce, Deterministic automation, Doing nothing compared across eight criteria
CriterionAI commerceDeterministic automationDoing nothing
Best atAmbiguous, unstructured inputFixed rules at volumeNothing, but it is free
Setup costModerateLow to moderateNone
Running costPer call, plus review timeLowThe manual work continues
PredictabilityStatisticalExactHuman variable
Fails byBeing confidently wrongRefusing to proceedSlowly, and invisibly
Needs reviewAlways, somewhereAt exceptionsConstantly
Scales with volumeWellVery wellBadly
Right first choice forEnrichment, searchApprovals, routingLow volume

Relevant work

The automation and integration record this practice is built on. Both are workflow and catalogue engagements rather than AI deployments, and they are on this page because they are the honest foundation for it.

  • A B2B procurement platform, built on Magento 2
    Custom platform · B2B

    Procurement and approvals platform

    Scalable B2B procurement platform streamlining purchase requests and approvals.

    B2B workflowsApprovalsIntegrations

  • Multi-category marketplace Magento 2 marketplace
    Magento 2 · Marketplace

    Multi-category marketplace

    A scalable Magento 2 platform powering a wide multi-category retail catalogue.

    Magento 2Multi-categoryScale

Business outcomes

  • 90%

    Of B2B order and quote processing automated

    Approval chains and quote handling modelled explicitly, taking the repetitive half of the process off people entirely.

    Where this came fromFrom a high-traffic B2B commerce platform on Magento 2, with a 40% cut in approval cycle time alongside it. This was deterministic workflow automation, not AI. It is quoted here because it is the record this practice stands on, and mislabelling it would be exactly the thing this page argues against.

  • Attribute coverage

    The metric for enrichment, and the one everything downstream depends on. Countable before the project starts and countable every week afterwards.

  • No-result rate

    The clearest measure of whether retrieval improved discovery. Every no-result session is a customer who described what they wanted and got nothing back.

  • Hours off the rota

    The plainest business case for operations automation, and the one that survives contact with a finance review. Cycle time is the second number.

One measured figure, and it belongs to workflow automation rather than to AI. This is the newest practice area on the site and no AI engagement on the record has a published measured outcome yet. If you want a delivered AI case study with a number attached, I do not have one to show you, and writing one before it exists would be the fastest way to lose the argument this page is making.

FAQ

What is AI commerce?
The application of machine learning and language models to commerce operations: enriching catalogue data, understanding search intent, ranking products, and automating repetitive operational work. It is a set of specific applications with measurable outcomes, not a layer you add to a store.
What is the most reliable AI use case in eCommerce today?
Catalogue enrichment with human review, and retrieval-based on-site search. Both have a number attached before you start, both fail visibly rather than silently, and neither requires the customer to trust the model directly.
Where should a business start with AI in commerce?
With the process, on paper, before any technology. Map the steps, remove the ones that exist because of a system limitation nobody revisited, and then look at what is left. A surprising amount of the available saving is in that first pass and costs nothing.
Do you build AI agents for eCommerce?
Where the task genuinely needs planning across multiple steps, and with explicit scope, defined tool access and a human checkpoint before anything touches price, stock or a customer commitment. Most requests described as agents are better served by deterministic workflow, and I will usually say so.
Will AI replace on-site search as we know it?
It will change the ranking, not remove the requirement. Working implementations are hybrid: embeddings for meaning, keyword matching for identifiers like SKUs and part numbers. Replacing keyword search outright is the most common way these projects regress.
How do you measure whether an AI project worked?
By choosing the number before starting. Attribute coverage for enrichment, no-result rate and search-session conversion for retrieval, hours and cycle time for operations. If none of those is being recorded today, record them for a month first.
Is a chat widget worth adding to a store?
It is the most requested and the least valuable AI feature on most stores. It can genuinely deflect where-is-my-order questions if it is wired to real order data. It cannot be trusted to quote price, stock or delivery, and a confidently wrong delivery date costs more than the ticket it saved.
What does AI cost to run in commerce?
Per-call model cost is usually the smaller half. The larger half is keeping indexes in step with a catalogue that changes daily and the human review time that makes the output safe to publish. Projects budgeted only for the build are the ones abandoned in month four.
Does AI commerce work on any platform?
It works wherever there is a real API and a real permission model to respect. Magento, Shopify and custom platforms all qualify. What matters is that the integration runs inside the platform rules rather than beside them, because anything running beside them drifts within a quarter.
Is there an AI case study with published results?
Not yet. The automation and integration record behind this practice is real and documented in the experience timeline, but no AI engagement on the list has a published measured outcome. Writing one up before it exists would be the fastest way to lose the argument this page is trying to make.

Next step

Let us look at your business model, the systems you already run and where you are trying to get to, before choosing the technology. That conversation is usually shorter than people expect, and it occasionally ends with me saying you do not need the project.

View relevant work