Contents · 9 sections
- 01Most AI in commerce is operationally empty.
- 02Start with the business problem, not the AI tool.
- 03AI is an operational layer, not a product.
- 04AI runs on the data underneath it.
- 05Governed automation, with humans in the loop.
- 06Phased adoption beats a big-bang programme.
- 07WithPraxis.ai · the intelligence and governance layer.
- 08Why most AI initiatives in commerce never reach trading.
- 09Proof · AI-adjacent work shipped on commerce projects.
- 10Common questions.
01 · The gap
Most AI in commerce is operationally empty.
Most of it talks in abstractions, overpromises automation, lists generic use cases, and ignores the governance, data and commercial reality underneath.
The buyers we work with sit between curiosity and paralysis: under pressure to "do something with AI", unclear where the value actually is, uneasy about accuracy and brand risk, unsure who governs it. The honest position is that AI in commerce is an operational discipline, not a product feature.
What the market promises
What actually pays back
"AI will transform commerce."
Most pilots never reach trading. The ones that do are narrow, governed and well-instrumented.
"Autonomous agents replace teams."
Agents extend teams. Take the human out and small model errors become customer incidents across thousands of interactions.
"Plug in a model, get value."
Value comes from the data underneath, the workflow it sits in, and the rules that govern it.
"AI fixes bad product data."
AI reflects bad product data back at you, louder. Fix the PIM first.
"Generative AI everywhere."
Generation is the easy part. Accuracy, brand voice, audit and rollback are the hard parts.
"AI is the new platform."
AI sits as a layer across your platform. Treat it like an integration, not a replatform.
02 · Problems before tools
Start with the business problem, not the AI tool.
Most organisations approach AI backwards, picking the tool, then hunting for somewhere to apply it. We start at the other end: where the operational friction sits, what an hour of it costs, and whether AI is even the right answer. Often it isn't, and saying so saves a six-figure pilot.
How most teams start
01
ChatGPT / copilots
Pick the tool first.
02
Find a use case
Hunt for somewhere to apply it.
03
Run a pilot
Demo internally. Generate excitement.
04
Try to scale
Hit the data and governance wall.
05
Quietly stall
Pilot becomes a slide, not a system.
How we start
01
Operational friction
Where workflows actually break.
02
Commercial value
What an hour of that friction costs.
03
Data + workflow audit
What needs to be true for AI to work here.
04
Smallest useful slice
Narrow scope, governed, measurable.
05
Scale what worked
Expand only the parts that paid back.
The first hour of any AI conversation we have is operational. Where's the bottleneck? Who owns it? What does an hour of it cost? The tool comes last, and only when there's a problem worth solving with one.
03 · Where AI creates real value
AI is an operational layer, not a product.
The wins are unglamorous and they compound. Faster product onboarding. More consistent merchandising. Search that finally returns the right thing. Customer service that holds brand voice. Internal teams that stop doing the same manual task forty times a day. AI extends a commerce team, it doesn't replace one.
01High
Product enrichment
Attribute generation, gap-filling, taxonomy alignment across thousands of SKUs.
Owner · PIM
02High
Search & discovery
Vector + keyword hybrid, intent rewrites, merchandising rules that learn.
Owner · Storefront
03High
Merchandising assist
Category curation, promotion ideas, anomaly flags reviewed by a human.
Owner · Trading
04High
Customer service
Drafted responses with brand voice, retrieval from policy and order history.
Owner · Service
05Medium
Content generation
Category copy, PDP variants, email, drafted, human-approved, version-tracked.
Owner · Marketing
06Medium
B2B account assist
Quote summaries, account-specific recommendations, replenishment prompts.
Owner · B2B
07Medium
Internal knowledge
Q&A across runbooks, contracts, product info, for staff, never customers.
Owner · Operations
08Foundational
Data normalisation
Cleanup of legacy product data, supplier feeds, return reasons, taxonomy drift.
Owner · PIM
We don't pick favourites. The right starting point depends on where the friction is, what data is governable today, and which workflow can absorb a phased rollout without disturbing trading.
AI without operational structure creates noise. AI wired into commerce, ERP, PIM and the workflows around them earns its keep, one workflow at a time.
The next three sections cover what makes AI commercially credible: the data underneath it, the governance around it, and the order it gets adopted in.
8-24
Systems per project
ERP · OMS · WMS · PIM · CDP · search · payments · tax
600+
Commerce projects
Since 1995, 31 years on complex commerce
<5%
AI initiatives in production
Industry · per published surveys, 2024-25
Default
Governed AI work
Human review, audit trail and rollback
04 · Foundations
AI runs on the data underneath it.
AI pages that lead with autonomous agents and never mention PIM are a tell. Good AI in commerce depends on governed product information, connected systems and operational context, none of which is interesting until it's missing.
We have ground to stand on here because we've spent 31 years on the unglamorous half of commerce: Adobe Commerce, Akeneo, ERP integration, B2B trade complexity. The same engineering underwrites the AI work. You can't bolt intelligence onto a broken substrate.
AI use cases
Generation, agents, recommendations, search.
Layer 5
Workflow & governance
Approvals, thresholds, audit, rollback.
Layer 4
Operational context
Customer, order, inventory, fulfilment state.
Layer 3
Connected systems
ERP · OMS · WMS · PIM · CDP · search · payments.
Layer 2
Governed product data
Attributes, taxonomy, quality, ownership.
Layer 1
Each layer up depends on the one beneath. Pilots fail when teams skip layers.
05 · Responsible AI & governance
Governed automation, with humans in the loop.
The commerce buyers we work with are uncomfortable with uncontrolled AI, hallucinations, brand risk, customer trust, audit, compliance. They should be. AI should operate inside governed systems: approval thresholds, confidence floors, audit trails, named owners, reversible actions. The matrix below is how we decide who reviews what.
Model confidence
High
Model confidence
Low / uncertain
Customer impact
Internal · low
Oversight model
Auto-publish · audit log
High confidence, low impact. Run free, log everything, periodic spot-check.
Oversight model
Human review · before publish
Low confidence, low impact. Drafts queued for staff, brand voice checked.
Customer impact
Customer-facing · high
Oversight model
Human approval · per action
High confidence, customer-facing. Gated behind a named approver.
Oversight model
No autonomous action
Low confidence, customer-facing. AI assists, the human decides.
Every AI action is logged. Every customer-facing action is reversible. Confidence thresholds are set per use case, never globally.
Customer-facing automation gets the heaviest oversight. Internal assistance, staff drafting, search ranking, anomaly flags, can run lighter. The principle is simple: the further an AI output sits from a human, the more governance has to wrap it.
06 · Pragmatic adoption
Phased adoption beats a big-bang programme.
Most businesses don't want a sweeping AI programme. They want a sensible starting point, controlled experimentation, measurable value early, and a way to grow capability over time. The shape below is how a typical first AI engagement runs: twelve weeks, three phases, one production use case at the end.
Weeks 01-02
Discover
Operational friction, data audit, oversight model.
Weeks 03-04
Design
Smallest useful slice. Confidence floor. Rollback path.
Weeks 05-08
Build · pilot
Internal-only first. Instrumented. Human in every loop.
Weeks 09-10
Govern
Approval thresholds, audit, brand voice review.
Weeks 11-12
Production
Limited release. Measured against baseline. Owner named.
Kickoff────────── 12-week phased adoption ──────────Production · governed
By week twelve there is one production AI capability, with measurable value, governed by a documented oversight model. From there capability grows in increments, never as a single launch.
07 · Featured · AI inside production workflows
WithPraxis.ai
Strategic partner · Governance & decision support
The strategy and governance layer behind responsible AI in commerce.
AI work is rarely a tool problem. It's a judgement, prioritisation and governance problem. WithPraxis.ai works alongside us: they bring the operating models, decision support and AI strategy; we bring the commerce systems, integrations and engineering. Clients get strategy, governance and execution under one delivery plan.
Capability stack
The six things WithPraxis.ai brings to a project. Strategy, governance and decision support, surfaced as named outputs, not slide-deck advice.
AI readiness assessment
Where the organisation actually sits, data, governance, capability, risk appetite.
Operating model design
Who owns AI, who reviews it, who escalates, how it scales.
Governance frameworks
Approval thresholds, audit trails, confidence floors, brand voice rules.
Decision support · structured
Frameworks that surface evidence and trade-offs before AI decisions get made.
Workflow intelligence
Mapping where AI can amplify the team, and where it must not.
Responsible automation
Boundaries on autonomous action. What stays human. What never gets automated.
Where it fits
How the partnership runs day-to-day, and what WithPraxis.ai is deliberately not.
How the partnership works
On AI engagements WithPraxis.ai sits at the strategy table: readiness, operating model, governance framework, decision support. iWeb runs the engineering, data, integrations, the AI layer itself, the production rollout. Both sides are accountable to the same delivery plan. The client runs one project, not two consultancies.
What WithPraxis.ai is not
Not a generic AI consultancy. Not a buzzword shop. Not a vendor selling "autonomous agents". The position is calm, evidence-led and governance-first. These conversations are built for operations, finance and compliance, not the innovation lab.
Read further
How WithPraxis.ai articulates the same operating models, governance frameworks and AI readiness work, in their own words.
How WithPraxis.ai thinks about responsible AI.
Their site goes deeper on the operating models, AI readiness work and decision frameworks we run projects against, the same governance our clients see inside an iWeb engagement.
08 · AI failure map · Industry vs iWeb-led
Why most AI initiatives in commerce never reach trading.
Six failure modes account for most stalled AI work. Industry bars on the left. The right column shows the same modes across the AI work we've shipped: same problems, lower numbers, because the operational discipline came first.
Failure mode
Industry · share of stalled
iWeb · 2024-26
Poor product / operational data
68%
8%
No governance · uncontrolled output
57%
4%
Tool chosen before problem
51%
3%
Pilot doesn't survive integration
46%
6%
Hallucination · brand-voice drift
38%
2%
No named owner · pilot stalls
34%
0%
Industry bars · share of stalled AI initiatives attributable to each cause, weighted across published AI-in-enterprise surveys 2024-25 · iWeb bars · AI-adjacent work shipped on commerce projects 2024-26.
09 · Proof
10 · Questions we get asked
Common questions.
What does iWeb do on AI for commerce?
Practical, operational AI wired into the commerce stack: product enrichment, search relevance, support tooling, decision support and merchandising assistance. Every use case sits on governed data and stays under human oversight.
Where does AI actually earn its place in a commerce stack?
Inside operational workflows that already exist: enriching product data before it hits the PIM, ranking and recovering search results, triaging support tickets, summarising order and account history for agents. The value is in the workflow, not the model.
How do you keep AI outputs safe in a commerce context?
Governed data as the input, human review on anything customer-facing, and audit trails on every automated decision. Confidence thresholds decide what runs unattended and what routes to a person.
Do we need to be on a particular platform to adopt AI?
No. iWeb adopts AI on the platform already in production, whether that is Adobe Commerce, Magento, Shopify Plus or a composable stack. Data readiness matters more than the storefront choice.
Where do most AI-in-commerce projects fail?
They start with the model rather than the problem, run on poor product and order data, and skip the governance work needed to trust the output. iWeb starts with the operational problem and the data underneath it.