Not every AI use case in commerce is worth building. Here’s how to decide.
Last week I defined what agentic commerce actually means — systems that act, not just suggest. But knowing the definition doesn’t tell you where to invest. The real question for any engineering leader is: which commerce problems actually benefit from AI, and at what level of autonomy?
Most teams get this wrong. They either underinvest — adding a recommendation widget and calling it done — or overinvest, building autonomous agents for problems that a SQL query could solve. The difference between wasted budget and real ROI is understanding where each use case falls on the value curve.
The Three Levels of AI in Commerce
Think of commerce automation as three distinct levels. Each has different infrastructure requirements, different risk profiles, and different returns.
Level 1: Static Rules and Heuristics
This is where most Magento and Shopify stores operate today. It’s not AI — it’s logic.
- Cart rules: If cart total > $100, apply free shipping.
- Customer segmentation: Tag customers by purchase count, assign to groups.
- Inventory alerts: If stock < threshold, send reorder notification.
- Search synonyms: Map “sneakers” to “running shoes” manually.
These are deterministic, predictable, and cheap to maintain. They also hit a ceiling fast. When you have 50,000 SKUs, 200 customer segments, and pricing that varies by geography and contract — rules don’t scale. You spend more time maintaining rules than they save.
Value ceiling: Low marginal cost, but diminishing returns as complexity grows. You can’t write rules fast enough to keep up with a dynamic catalog.
Level 2: Predictive Models and AI Features
This is where most “AI-powered” commerce tools live. A model analyzes data and produces insights or ranked outputs. A human or a rule decides what to do with them.
- Product recommendations: “Customers who bought X also bought Y” — collaborative filtering.
- Search ranking: Reorder results by predicted conversion likelihood, not just keyword match.
- Demand forecasting: Predict which SKUs will spike next month based on historical patterns.
- Churn prediction: Flag customers likely to lapse based on behavior signals.
- Dynamic pricing suggestions: Model recommends price adjustments. A merchandiser approves.
These deliver measurable lift. A good recommendation engine adds 5-15% to AOV. Predictive search ranking can improve conversion by 10-20%. Demand forecasting reduces overstock costs.
But the model doesn’t act. It informs. Somebody still has to review the pricing suggestion, approve the reorder, curate the recommendation placement. That human bottleneck limits how fast you can respond and how many decisions you can optimize simultaneously.
Value ceiling: High ROI per use case, but linear scaling. Each new use case needs its own model, its own pipeline, its own approval workflow.
Level 3: Agentic Systems
Here the AI doesn’t suggest — it decides and acts within defined boundaries. It has a goal, observes outcomes, and adjusts its approach.
- Autonomous repricing: Agent monitors competitor prices, demand signals, and margin targets. Adjusts prices within guardrails. No human approves each change.
- Proactive customer retention: Agent detects churn signals, selects intervention strategy (discount, email sequence, loyalty reward), executes, and measures effectiveness.
- Intelligent inventory allocation: Agent rebalances stock across warehouses based on regional demand patterns, transfer costs, and delivery SLAs.
- Conversational commerce: Agent handles natural language queries (“I need to reorder the same packaging we bought for the Berlin office last quarter”) by searching order history, confirming quantities, and creating the cart.
The step change: agentic systems compound. A repricing agent that observes the outcome of its decisions gets better at pricing over time. A proactive retention agent learns which interventions work for which customer segments. The value curve bends upward.
Value ceiling: Theoretically unlimited, practically bounded by guardrails, data quality, and trust.
Mapping Commerce Use Cases to the Right Level
Here’s where most teams make mistakes: they apply Level 3 infrastructure to Level 1 problems, or stay at Level 1 when the problem clearly demands Level 2.
| Use Case | Level 1 (Rules) | Level 2 (Predictive) | Level 3 (Agentic) |
|---|---|---|---|
| Product recommendations | “Top sellers” widget | Collaborative filtering model | Agent personalizes per session in real time |
| Pricing | Tiered price rules | Demand-based price suggestions | Autonomous repricing within guardrails |
| Inventory management | Reorder point alerts | Demand forecasting | Autonomous rebalancing + purchasing |
| Customer support | FAQ page, macros | Ticket classification + routing | Agent resolves issues autonomously |
| Search | Keyword match + synonyms | ML-ranked results | Conversational, intent-aware agent |
| Cart recovery | Static email after 24h | Predicted-optimal timing/channel | Agent selects strategy per customer |
The pattern: Level 1 solves the problem when complexity is low and change is slow. Level 2 solves it when you have data and the problem is well-defined but too complex for manual rules. Level 3 solves it when the problem requires multi-step reasoning, adaptation, and speed that humans can’t match.
The Platform Lens
On Adobe Commerce
Adobe Commerce gives you the deepest hooks for Levels 2 and 3. You can capture events at every stage — catalog updates, cart changes, order status transitions, customer behavior — through observers, plugins, and the message queue.
For Level 2, Adobe Sensei provides a baseline recommendation engine. Most teams supplement it with third-party tools (Nosto, Algolia, Bloomreach) for search and personalization. These work and they’re relatively straightforward to integrate.
For Level 3, the pattern is external orchestration. An agentic service built in Python or Go subscribes to Magento events via RabbitMQ or webhooks, processes them through decision models, and pushes actions back via REST or GraphQL. Magento remains the commerce engine. The agent wraps around it.
The trap to avoid: building agentic logic inside Magento. The moment you put LLM calls in a Magento observer or a cron job, you’ve coupled AI latency to commerce performance. Keep the boundary clean.
On Shopify
Shopify’s more constrained model actually makes the level decision clearer. You don’t have the option to shove agent logic into the platform core.
For Level 2, Shopify’s ecosystem has mature options. Shopify’s built-in recommendations, plus apps like Rebuy, Klevu, and Searchspring handle most predictive use cases without custom infrastructure.
For Level 3, you build externally. A separate service receives Shopify webhooks, runs decision logic, and pushes changes back via the Admin API. Shopify Functions handle the commerce-side execution (discounts, validation). Hydrogen gives you frontend flexibility for agent-driven experiences.
Shopify’s advantage: the API-first architecture makes the boundary between commerce and agent layer naturally clean. The constraint is flexibility — you can’t override core checkout behavior the way you can in Magento, which limits some agentic patterns.
What You Actually Need to Build This
Level 1 needs nothing beyond what Magento or Shopify already gives you. If your rules work, keep them.
Level 2 requires: – Clean, accessible customer and product data (if your catalog is a mess, start here) – A third-party ML tool or a basic internal pipeline (Python + scikit-learn is enough for many cases) – A feedback mechanism to measure whether predictions actually improved outcomes – 1-2 engineers who understand data pipelines
Level 3 requires everything from Level 2, plus: – A separate runtime for the agentic service (not inside the commerce platform) – Event infrastructure (message queue or event bus for real-time signals) – LLM API access (OpenAI, Anthropic, or equivalent) for natural language reasoning – Guardrails: rate limits, approval gates for high-impact actions, audit trail – Monitoring for agent drift — is the agent still making good decisions after a month? – A team that understands both ML systems and commerce domain logic
If you’re not ready for Level 3 infrastructure, you’re not ready for Level 3 use cases. That’s not a failure — it’s a smart allocation of resources.
Decision Framework: Where to Invest
Start at Level 1 when: – You have fewer than 5,000 SKUs and 3 customer segments – The problem is well-understood and changes rarely – You need results this quarter with zero infrastructure investment
Move to Level 2 when: – Rule complexity is growing faster than your team can maintain – You have historical data but aren’t using it for decisions – A 10% improvement in search, recommendations, or pricing justifies the tooling cost – You can measure the outcome clearly (conversion rate, AOV, churn reduction)
Move to Level 3 when: – The decision loop needs to be faster than a human can manage – The problem requires multi-step reasoning (detect → decide → act → learn) – Outcomes compound — the system gets better with more data and more decisions – The ROI justifies dedicated infrastructure and monitoring – You’ve already proven value at Level 2 for this use case
Stay where you are when: – “AI” is the goal, not the outcome – You can’t measure the business impact of the automation – Your data foundations aren’t solid enough to trust model outputs – The team can’t maintain the system after the initial build
The Leadership Angle
The most expensive mistake in AI commerce isn’t building the wrong thing. It’s building at the wrong level.
I’ve seen teams spend six months building an autonomous pricing agent when a simple demand-based suggestion model would have delivered 80% of the value in six weeks. I’ve also seen teams stuck on static rules for search when a basic ML-ranked model would have paid for itself in a month.
The value curve isn’t a ladder you climb for prestige. It’s a map of trade-offs. Each level up costs more — more infrastructure, more complexity, more monitoring, more organizational trust. The return has to justify it.
For most commerce teams in 2026, the highest-ROI move is getting Level 2 right. Solid predictive models for search, recommendations, and pricing. Measurable, maintainable, and they compound over time through better data.
Level 3 — true agentic systems — is where the market is heading. But heading there doesn’t mean you have to arrive today. Build the data foundations. Prove value with predictions. Then let the agents take over, one use case at a time.
The teams that win aren’t the ones who go agentic first. They’re the ones who go agentic at the right time, for the right problems, with the right guardrails.
