How to Keep AI Recommendations On-Brand and On-Budget

The Short Answer: Keep the Server in Control, Not the AI

AI product recommendations go off-brand or over-budget for one reason: the AI was given too much authority over which products to surface. The fix is architectural. The server — not the language model — should decide which products are eligible. The AI's only job is to write convincing, brand-consistent copy for a shortlist it did not choose.

This separation of concerns is the foundation of every reliable recommendation system, and it is the design principle behind SmartBrain's conversational commerce engine.

What "On-Brand, On-Budget" Actually Means

Before diving into tactics, it helps to be precise about terms.

• On-brand means the recommendation matches your store's voice, aesthetic, and values — no recommending a fast-fashion accessory on a sustainable goods store, no formal language on a streetwear brand.
• On-budget means the recommendation respects both your margin floor (you do not discount below cost) and the customer's stated or inferred price ceiling.

Both constraints are data problems, not writing problems. They should be enforced before any text is generated.

Why AI Models Alone Cannot Guarantee These Constraints

Large language models are trained to produce plausible, helpful text. They are not trained to query your live inventory, check your current margin rules, or remember that you stopped selling a product line last Tuesday. If you ask a raw AI to "recommend something under €50," it may confidently suggest an item that is out of stock, discontinued, or priced at €68 after a recent update.

The model is not lying — it simply has no reliable access to ground truth. Plugging a product catalog into a prompt does not solve this; catalogs go stale in hours, and token limits make full-catalog injection impractical at scale.

How to Enforce Brand and Budget Constraints at the System Level

Step 1: Define Eligibility Rules on the Server

Your backend should maintain a rule engine that filters the catalog before the AI sees it. Typical rules include:

• In-stock only (quantity above a configurable threshold)
• Margin above a minimum percentage (prevents discounting into losses)
• Category tags matching the conversation context (a customer asking about skincare should never receive a kitchen appliance)
• Price within a range derived from the conversation (customer said "around €30" → serve products priced €20–€40)
• Brand-tier filter (premium customers see premium SKUs; value-seekers see entry-level)

The AI receives only the products that pass every filter. It cannot recommend what it cannot see.

Step 2: Use Brand Voice Templates, Not Free-Form Generation

Brand voice drift happens when the AI is told to "write a recommendation" with no structural constraints. Instead, define a template that locks the tone while leaving product-specific content flexible.

For example, a minimal template might look like: [Warm opener based on customer need] + [Product name and one key benefit] + [Call to action matching brand tone]. The AI fills the blanks — it does not choose the structure. A minimalist brand gets short sentences and no exclamation marks because the template rules it out, not because the AI guessed correctly.

Step 3: Validate Outputs Before Delivery

Even with a filtered shortlist and a constrained template, a post-generation check catches edge cases. Run a lightweight validation pass that confirms the product name in the AI output matches an item in the filtered shortlist, and that no price figures in the copy contradict your catalog data. This takes milliseconds and prevents the most common failure mode: the AI paraphrasing a product description and inadvertently citing a price from an old training example.

Catalog-First vs. Prompt-First: A Practical Comparison

Two common approaches illustrate why architecture matters more than prompt engineering.

• Prompt-first approach: You describe your catalog in the system prompt and ask the AI to recommend something appropriate. Simple to set up. Fails silently when the catalog changes, ignores real-time stock, and scales poorly beyond a few hundred SKUs.
• Catalog-first approach: The server queries live inventory, applies eligibility rules, and passes a shortlist (typically 1–5 products) to the AI for copy generation only. Requires more upfront engineering. Maintains accuracy as the catalog evolves and handles catalogs of any size.

SmartBrain is built on the catalog-first model. The recommendation decision happens in the server layer against real Shopify inventory data; the language model writes the message the customer reads.

Practical Examples by Store Type

Sustainable Apparel Brand

Brand rules: no synthetic materials, minimum 40% margin, stock above 10 units. A customer DMs asking for "a gift under €60 for someone who loves the outdoors." The server returns two eligible products — a merino wool beanie and a recycled-fiber tote. The AI writes two short, warm sentences for each, using the brand's first-person-plural voice. Nothing outside the rules ever surfaces.

Electronics Retailer

A customer says their budget is €200. The server filters to in-stock items priced €160–€220 (allowing slight upward flex), excludes clearance items below margin floor, and returns a shortlist of three products. The AI writes a comparison-style message in the store's direct, spec-focused tone. The customer sees no product outside their stated range.

DM Automation Agency

Agencies running conversational flows for multiple clients need per-client rule sets. With a system like SmartBrain, each client workspace carries its own eligibility configuration — brand tier, price bands, excluded categories — so the same underlying engine serves a luxury jeweler and a budget homewares brand without cross-contamination.

FAQ

Can I use this approach if my catalog changes daily?

Yes — and daily-changing catalogs are exactly where this approach proves its value. Because eligibility rules run against live inventory at query time, a product that sells out at 9 a.m. is automatically excluded from all recommendations by 9:01 a.m. No manual prompt updates required.

What if a customer's budget is lower than any eligible product?

Design your rule engine to return a graceful fallback: either the closest product above the stated budget with transparent pricing, or a "nothing available right now" message that does not fabricate alternatives. The AI should never be allowed to invent a product or price.

How do I maintain brand voice across multiple AI-generated messages in a long conversation?

Store the brand voice template and any approved tone parameters in your server configuration, not in the conversation history. Each message generation call pulls fresh constraints from the server, so tone stays consistent even if the conversation runs for dozens of turns.

Does this architecture work for agencies managing multiple stores?

It is particularly well suited to agencies. A multi-tenant setup — like SmartBrain supports — lets you configure separate eligibility rules, voice templates, and budget floors for each client from a single dashboard, while the AI layer remains shared infrastructure.

Is prompt engineering enough to enforce budget constraints?

Prompt engineering is not sufficient for hard constraints. A prompt instruction like "only recommend products under €50" will be followed most of the time, but it can fail on edge cases, especially if the model has seen conflicting examples in training. Server-side filtering is the only reliable enforcement mechanism for budget and inventory rules.