How does it work?

This chapter covers the evolution of recommendation systems, measurable AI objectives, and the architecture of top-tier solutions. Then it compares the Build approach (custom development) with the Managed approach (native cloud solution) and presents the diagram of a "privacy-first" hybrid implementation from our projects. Finally, we discuss the challenges of implementing AI for recommendations.

As a technological introduction, this is the funnel architecture diagram for a recommendation system:

Image Diagram

   FULL PRODUCT CATALOG (e.g., 1 million)
                    │
                    ▼
STEP 1: RETRIEVAL/GENERATION (High speed, wide coverage)
"Find 500 similar products"
(Vector search + lexical filters)
│
▼
CANDIDATES (e.g., 500 products)
│
▼
STEP 2.1: SCORING / RANKING (Medium speed)
"Top results by probability of purchase (pCVR)"
(Deep Learning Neural Network)
│
▼
STEP 2.2: RE-RANKING AND RULES (Fast)
"Remove out-of-stock and low margin"
(Post-processing via guardrails)
│
▼
FINAL RESULT (e.g., 3 recommended products)

                

For reading flow, we will illustrate technical concepts using the Google Cloud ecosystem. However, the architecture presented is universal: it can be replicated in other clouds, such as Microsoft Azure or Amazon Web Services, or with various SaaS solutions.

Guide: How to choose a Google Cloud implementation partner

The history of recommendation generations explains how we arrived at the present moment. We will see an evolution from human intuition to mathematical engineering, which is becoming accessible at scale through AI.

In many sales teams, recommendations are based on the expertise of senior agents and the company's know-how. The company knows that "X prefers brand Y" or that "Only this specific fitting works for this type of pump."

The major limitation is scalability and personnel risk. Beyond a certain point, the company:

• Can only grow its own know-how through rigorous training programs
• Can only apply commercial policy with detailed internal documentation and strict procedures

When a senior agent leaves the company, logical connections are lost. New agents sell basic products, missing upsell opportunities for 6-12 months ('ramp-up'). In fact, indicators cited in Ch. 1 show the maximum impact of AI in the case of junior sales agents.

With the emergence of e-commerce and ERP platforms, companies began to institutionalize their own know-how and rules.

The major limitation is maintenance:

• In a catalog with 50,000 items, it is impossible to maintain links manually.
• Successful companies rigorously define a hierarchical taxonomy that allows them to manage rules.
• Rules can become outdated: accessories go out of stock or are removed from the manufacturer's catalog.

Another limitation is rigidity. The system is "blind" to the customer, forcing B2B companies to operate a manual approval hierarchy or a loyalty system for discounts.

Well known for "Customers who bought this also bought..." sections, this is classic Amazon-style technology used by global and local e-commerce giants. This is why many popular plugins (e.g., WooCommerce) sell this logic under the generic "AI" label.

Today's standard in Predictive AI was accelerated by major platforms like YouTube, Google Shopping, and Netflix. It has been used in recent years by major players, including regional e-commerce giants, and AI progress is democratizing it for the mid-market segment; see the Managed approach in Ch. 2.4.

The same technique underlies hybrid search, which combines lexical search (e.g., SKU, close dimensions) with semantic search (e.g., wood screw = self-tapping screw). It brings the greatest benefits to B2B software; see applications in Guide #4.

What Gen 3 adds:

1. Understanding the Sales Sequence

The system understands the cyclicality of consumables. If a clinic bought an ultrasound machine today, it won't buy another one tomorrow. But it will usually need conductive gel. The algorithm predicts the probability of purchase, and a "Buy it again" widget can be generated.

With an added time-series modeling AI model, it can also predict the date the customer's stock will be depleted, activating the Predictive Replenishment deliverable from Ch. 3.

2. Content-Based Understanding (including Multimodal)

When a new product appears without history, predictive AI analyzes the attributes (complete and descriptive). It "understands" it and recommends it to relevant customers, solving the Cold Start problem from Gen 2.

See attribute extraction from images and PDFs in Guide #3: Multimodal Input

3. Context Processing

It takes into account not just "who you are," but also "what you are doing now." For example: the customer is on mobile, is in a hurry, is during working hours or on a weekend.

Already in 2026, AI agents based on LLMs are everywhere in the press, for example, after the launch of UCP for automatic negotiation between agents. LLM models are excellent as a conversational interface and as an orchestrator: they can explain the recommendations made, summarize the commercial context, and can quickly compose business rules, although they cannot apply them with guarantees (due to the risk of hallucinations) and operate at a slow speed.

For a scalable recommendation system with strict latency and predictable cost, the system usually remains Gen 3. LLM applications intervene selectively, for example through chatbots that explain and justify for the customer. See the Copilot deliverable in Ch. 3.3.1.

In brief

You can organize your sales engine with AI around fundamental notions that can be transformed into measurable indicators for the company. See also the deliverables in Ch. 3 .

1. Optimizing customer experience

The engine of sales performance. AI is used to link company data to the customer experience (usually frontend via e-commerce, but also in direct or multi-channel sales) through personalization and contextual intelligence.

2. Automating internal processes

What it is: Using AI to help the internal agent (acting as a copilot), with the goal of allowing people to concentrate on high-value activities, without repetitive tasks.

3. Models and tuning (personalization)

Adapting AI to the specific data, products, and customers of the company brings digital literacy to the company and the adoption of top technologies. Pre-trained models can be used, or you can customize (tune) foundation models (e.g., Google Gemini).

4. Data cleanliness and relevance

Recommendations based on the exact context of the customer depend on quality data from all company systems. Cleaning and controlling your own data are the keys to success.

5. Guardrails (safety rules)

Rules that protect the company and its image, avoid absurd recommendations, and prevent losses (e.g., selling below margin). These can be deterministic like expert systems, in line with the hybrid architecture idea. Google Cloud also includes safety controls in its AI products.

6. Explainable AI (xAI)

The ability to explain why a recommendation was made, to gain the trust of the agent, management, and ultimately the customer.

7. Human-AI Interaction

An unspoken fear in B2B sales is: "If the customer buys the product recommended by AI, do I get my commission?" To prevent resistance to new technology, the agent can use AI recommendations as an instant product sheet to win more customers.

Today's established model uses Deep Learning to achieve recommendations, upsell, cross-sell, and understanding of the purchase sequence, product content, and customer context.

This architecture is implemented from scratch by e-commerce giants, or included in fully managed solutions like Google Cloud's Vertex AI Search for commerce technology.

Case Study: Quoting time -68%, +14% average order value for distribution ERP

1. Top-tier architecture

Many e-commerce giants, as well as companies using a managed approach like Vertex AI Search for commerce, implement this architecture with two main stages: initial candidate generation (Retrieval) and their careful selection (Ranking).

Below we present the technical stages, but remember that in 2026 implementing from scratch is not necessary, see the Managed approach in Ch. 2.4.

2. Step 1: Candidate Generation (Retrieval)

From the millions of products in the catalog (possibly pre-processed with basic filters), a fast algorithm selects a few hundred relevant ones.

3. Step 2: Ranking (ordering / scoring and post-processing)

From the hundreds of candidate products, the model must calculate a numerical score for conversion probability. For example, pCVR: probability of conversion (purchase, offer acceptance, other success event). These candidates rank first.

Business rules (guardrails) applied at the end during post-processing are essential for AI control (see details in Ch. 2.5.1):

• Ensure alignment with commercial policy: some scores are adjusted, some results are removed.
• Express the priority of deterministic business rules over AI non-determinism.

4. Upsell, cross-sell, discounting

Image Diagram

APPLICATION OR SITE             AI ENGINE + RULES (CUSTOM)                    APPLICATION OR SITE

┌───────────────────────┐       ┌─────────────────────────────────┐       ┌───────────────────────────┐
│ Customer: Cart        │       │ 1. Estimate pCVR                │       │ Reload cart:              │
│ Products: [A, B, C]   │ ─────>│    (conversion probability)     │ ─────>│                           │
│ Total: 950 USD        │       │ 2. Check margin                 │       │ - Product A               │
└───────────────────────┘       │    (via guardrails)             │       │ - Product B               │
                                │ 3. Decide:                      │       │ - Product C               │
                                │    "Worth -4% to close now"     │       │ [DISCOUNT: -38 USD]       │
                                └─────────────────────────────────┘       │ (limited availability)    │
                                                                          └───────────────────────────┘

                

Upsell: upgrade

Based on product features and attributes. For a 1,200 USD laptop, the model can push laptops with similar but marginally better specs (and better margins) in the ranking because it learned: customers comparing specs tend to upgrade.

Cross-sell: accessories

Based on compatibility rules reinforced by behavioral data.

For example, the vector of an excavator is strongly associated transactionally with the vector of the specific hydraulic filter kit (for maintenance). This is the first foundation for deliverable #8: Technical Compatibility Engine.

The other foundation is a system of exact compatibilities between products, for which you need detailed technical sheets and to specify filters in the guardrails (e.g., which fields must be identical).

Discounting: price optimization

The architecture can treat the discount as a conversion optimization variable. In fact, AI models will estimate the conversion probability for each scenario in the background.

This is the foundation for the Dynamic Price Optimization deliverable from Ch. 3. In Managed solutions, this logic is often custom (a separate component).

The major difference between collaborative filtering (Gen 2) and Deep Learning recommendations (Gen 3 - current) is the ability to understand immediate intent, not just history.

1. Session Sequence Modeling

The current architecture doesn't just look at the global history, but at the current session.

                          Phone view  Add to cart  Search cases
                        

2. Device and time context

Hot data about the customer (session time, device used) is essential to recommend what the customer wants now.

In highly specialized e-commerce, context is also ensured through guardrails, manually defined to optimize the current shopping session.

The quality of recommendations varies directly with the quality and speed of input data (the GIGO principle: Garbage In, Garbage Out). For the AI model to be relevant, the data must be clean and pre-correlated.

For B2B, we recommend unifying data from all sources and continuous ingestion via API.

See in Guide #2 how we build the single source of truth to correctly feed the algorithms

1. Product catalog (Primary source: ERP, Secondary sources: WMS, PIM, e-commerce)

This is fundamental company data. AI can have access not just to the product name, but also to critical business metadata.

Retraining the recommendation model can take place daily, processing large amounts of data for all products.

2. Customer history (Primary source: CRM)

With the generalization of LLMs, ingesting unstructured data is also possible: sales calls and sent emails are permanently transcribed, then a sentiment or recommended action score (next best action) is obtained and can be written back into the CRM.

Case study: Financial sentiment analysis with GenAI & Google Cloud

3. Live events and clickstream (Source: e-commerce, other company applications)

These are immediate data and the strongest predictor of purchase intent, at least in e-commerce.

Live data ingestion flow (e-commerce)

For clickstream, systems like Apache Kafka or Spark are used in the Build version, or API (userEvents) in the Managed version.

On product click, an event is sent to a Feature Store or via API. When the next page loads, the system queries the Feature Store: "What are the last 5 clicks of this customer?" and the model takes the updated context into account.

This prevents stale recommendations. The application (UX) can pivot directly to cross-sell. For example, it will recommend a warranty extension or an accessory.

The neural networks used in custom architectures are largely black box. The human operator does not know exactly why product X was recommended to customer A now.

Risk reduction strategies are used.

In brief:

In business-to-business sales, the stakes of recommendations and upsell are not psychological impulse (as in the fashion industry), but the compatibility of supply with demand and the sales value. The challenge is managing a vast catalog, often supplemented with untested new products (Cold Start) and complying with strict business rules (e.g., do not recommend cheap products to premium customers).

We compare building your own models (in-house data science) with using pre-trained models, such as those in Vertex AI Search for commerce. How do we recommend a complementary product alongside a "24kW Heating Boiler" in the two approaches?

This is the classic data science approach. You will develop a Deep Learning model (Gen 3). For simplicity's sake, we illustrate via Gen 2 a simplified collaborative filtering workflow with open-source technologies:

Data flow:

Image Diagram

1. CUSTOMER EVENTS
(Order history, Data warehouse)
        │
        ▼
2. APACHE SPARK / PYTHON PANDAS
(Batch processing, e.g., nightly)
        │
        ▼
3. ALS MODEL / MATRIX FACTORIZATION
(Algorithm training: "Who bought X also bought Y")
        │
        ▼
4. REDIS / MEMCACHED
(Storing pre-calculated pairs.
E.g.: "product A" -> "recommend B, C, D")
        │
        ▼
5. API SERVER
        │
        ▼
6. DISPLAY in APPLICATION / E-COMMERCE

                

In 2026, the standard is to use vector bases (e.g., Milvus, Qdrant, pgvector) and Deep Learning models (e.g. Transformers) for semantic understanding (Gen 3). For example, "yellow, utility, reflective uniform" will sit next to "ISO 20471" (the protective equipment standard), even if the words do not match.

The advantage is (usually) ownership of the developed code and the possibility of implementing any type of learning.

There are specialized time-series models that can be used in the cloud (e.g., Vertex AI Forecast), and many types of models in Google's Model Garden (open-source as Llama / Mistral or commercial such as Gemini). These can be connected to the data and the rest of the logic developed internally.

As for infrastructure, the Build approach requires provisioning and maintaining servers (e.g., Kubernetes), either in-house or in a cloud like Google Cloud (IaaS), for which you will pay for the infrastructure (IaaS - Infrastructure as a Service).

As for limitations, because matrix calculation is intensive, it is usually done once a day (batch). If a product suddenly becomes popular at 10:00 AM, special logic (which is achievable) is needed to update the recommendations before the next batch; otherwise, the algorithm will not adjust until the following day.

• Team: Developer, data/ML engineer, DevOps engineer, business analyst
• Model development: 2-3 months (with iterations required for accuracy).
• Infrastructure (events, models, API): 1-2 months.
• Total until launch: from 4-6 months.

This is an approach where the AI model is fully managed externally. We use a service like Vertex AI Search for commerce, which includes Deep Learning models known in the market to be similar to Google Shopping or YouTube.

Data flow:

Image Diagram

1. CUSTOMER EVENTS                   2. ERP / E-COMMERCE CATALOG
  (Click, add to cart,                 (Metadata: category, price,
   view, real-time)                     attributes, description)
            └─────────────────┬─────────────────┘
                              │
                              ▼
               3. VERTEX AI SEARCH FOR COMMERCE
                   (Google-managed engine)
       ┌───────────────────────────────────────────────┐
       │ • Deep & Cross Network (DCN)                  │
       │ • Business rules (explicit filters)           │
       │ • Optimization setting: CTR (click-through    │
       │   rate) or Revenue                            │
       └───────────────────────────────────────────────┘
                              │
                              ▼
                4. API SERVER/MINI-APP (WIDGET)
                              │
                              ▼
                5. DISPLAY IN APP / E-COMMERCE
            ("For customer X, now, show product Y")

                

The advantage is the bundled resolution of the Cold Start problem and semantic comparison. If a new product appears in the ERP (e.g., "Boiler model 2026") with all its specifications, the algorithm indexes its attributes (e.g., power, brand, category) and even the image. It will recommend it to customers looking for similar products and, furthermore, Google algorithms are known in the market as being excellent for new customers without a history.

As for limitations, the product does not currently include some types of learning, such as time-series modeling for cadence or discounting logic. These require orchestrated implementation in the control middleware. A selected open-source model from Google Cloud's Model Garden can be used.

Secondly, you do not have access to the code of Managed models, but you can configure their objectives from the list available in the console (e.g., maximizing CTR or Revenue).

• Team: Developer, Data Engineer, business analyst
• Catalog synchronization (if data is clean): 2-3 weeks.
• Infrastructure (events, API): 1-2 weeks.
• Initial training: 1-2 weeks (automatic).
• Total until launch: from 2-3 months.

Building a recommendation system from scratch (the Build version) is an option for companies with a high volume of traffic and a need for code-level control.

Building with Vertex AI Search for commerce (the Managed version) is an option for fast implementation. Components that are not available, such as time-series forecasting, can be implemented custom.

In brief:

For companies with strict security requirements or on-prem ERP systems, we illustrate from OPTI Software projects what a centralized middleware for commercial truth looks like.

It can function in the cloud or on-prem and connects selectively to AI functions, whether managed by Google or custom via Model Garden.

Image Diagram

       INTERNAL DATA ZONE                    MIDDLEWARE ZONE             AI ZONE                        
 ┌─────────────────────────────┐     ┌─────────────────────────────┐     ┌─────────────────────────────┐
 │ 1. DATA SOURCES             │     │ 2. DATA HUB & SYNC          │     │ 3. Vertex AI (Search +      │
 │                             │     │                             │     │    Search for commerce,     │
 │ ┌───────────┐ ┌───────────┐ │     │ ┌─────────────────────────┐ │     │    Model Garden)            │
 │ │ ERP       │ │ FILE SRV  │ │     │ │ Fast database           │ │     │ ┌─────────────────────────┐ │
 │ │ (Master)  │ │ (PDF/Docs)│ │     │ │ (Cache)                 │ │     │ │ Vectors                 │ |
 │ └─────┬─────┘ └─────┬─────┘ │     │ └────────────┬────────────┘ │     │ │ Index                   │ |
 │       ▼             ▼       │     │              ▼              │     │ └────────────▲────────────┘ │
 │ [Secure unidirect. Sync]    │────>│ [Filtering & sanitization]  │     │              │              │
 └─────────────────────────────┘     │              │              │────>│ [Ingestion (public only)]   │
                                     │              ▼              │     │                             │
                                     │ ┌─────────────────────────┐ │     │ ┌─────────────────────────┐ │
                                     │ │ API Gateway             │ │<────│ | Gemini / LLM            | │
                                     │ │ (Combining AI + ERP)    │ │     │ | (RAG & Processing)      | │
                                     │ └────────────▲────────────┘ │     │ └─────────────────────────┘ │
                                     └──────────────┼──────────────┘     └─────────────────────────────┘
                                                    |                                                   
                                                    ▼                                                   
                                       ┌──────────────────────────┐                                     
                                       │ 4. Applications          │                                     
                                       │    (Quoting / Chatbot)   │                                     
                                       └──────────────────────────┘                                     

                

Comparative Architecture: Sales Quoting Software with AI from OPTI

In Vertex AI Search for commerce, integrating the API into the application and building deliverables requires programmers, but the sales strategy can be partially defined from the interface (low-code).

Low-code/no-code control flow:

Image Diagram

BUSINESS ANALYST               VERTEX AI DASHBOARD (UI)                E-COMMERCE (live)

┌─────────────────────────┐       ┌───────────────────────────┐       ┌───────────────────────────┐
│ Strategy:               │       │ 1. ServingConfig          │       │ "Recommended" Widget      │
│ "I want to push         │ ─────>│    [ Add Rule ]           │ ─────>│                           │
│  MAN trucks to          │       │ 2. Condition:             │       │ 1. MAN TGX 480            │
│  the front."            │       │    Brand == "MAN"         │       │    (promoted)             │
└─────────────────────────┘       │ 3. Action:                │       │ 2. Other products         │
                                  │    Boost score 0.5        │       │    (standard)             │
                                  └───────────────────────────┘       └───────────────────────────┘

                

For a commercial department with a technical analyst, display rules can be controlled from dashboards using visual controls. Training and communication with the technical team are essential to avoid misconfigurations, as the risks are considerable.

We do not want to sell below the minimum margin, recommend impossible / incompatible products, or leave the customer or sales agent without a suggestion when the AI integration fails.

The hybrid approach does not just mean AI + Cloud, but also predictive or agentic AI (Gen 3-4) controlled by explicit rules as in Gen 1. The purpose of using company data is to control the AI output. Guardrails are the safety net of the business.

1. Business Protection

Rules are invisible to the customer, but critical for the finance or marketing department:

Especially for upsell in B2B, rules are essential to avoid violating the contract, which may provide, for example, for only certain approved manufacturers.

Additionally, AI models can learn incorrect correlations, leading to repetitive or absurd recommendations. A high-volume customer who bought a fleet of Apple laptops should not be recommended Asus bags and accessories.

The solution is metadata-based exclusion rules that encode human expertise.

2. Fallback: Experience continuity

What happens if the API has latency, if the customer is new (no history), or if there is no reliable recommendation for an obscure product (e.g., score > 0.5 derived in the Build approach)?

3. Expert systems: Humans can know better

While AI is powerful, it can know too little (underfitting) about the complete commercial context. A manager can override AI recommendations.

Forcing campaigns:

For new campaigns where no relevant contextual data exists, promotion periods (e.g., Boost) can be defined with forced priority.

Seasonality:

The human calendar differs from pure statistics, for example:

• January (start of budgets): Prioritize high-value equipment considered investments.
• October (end of season in construction): Prioritize clearing stocks of seasonal or perishable materials.

In technical B2B, compatibility is binary: it works or it doesn't. Expert systems are used, defined on details from the specification sheet.

Technically, guardrails are classified as hard (filters and eliminations), soft (promotions and demotions in list order), and process (necessary for control and audit. E.g.: "request permission for excessively large discount" / "Log the application of the rule").

The "Garbage In, Garbage Out" principle states that without correct fresh data, AI recommendations will also yield wrong results. Additionally, any change in the fundamental source of company truth (e.g., ERP update, tax rate changes, or modification of a crucial category) can break integrations.

The solution is defining a Data Contract between the team managing the source of truth (ERP, CRM, e-commerce) and the team managing the AI integration.

1. Data preparation: from chaos to clarity

These procedures maximize the operational clarity of the company and minimize the accumulated technical debt from the past.

See the OPTI Software methodology in Ch. 5.3.2

2. Data schema: the structure

For the integration to "understand" the data source, three levels are defined. The Entities that will synchronize, how their data fields must look, and what the consequences for the integration are when the contract is violated.

3. Data semantics: the meaning

The AI application developers must understand what the data in the source systems means to keep the behavior of the entire system identical, especially toward the final customer.

Pay attention to everything related to the following:

The meaning of each field matters for the Build version, to avoid training a model on irrelevant correlations. It also matters in the Managed version, as with any SaaS integration.

4. Data SLA: latency and race conditions

Conditions are defined such as "Stock changes must reach the AI within a maximum of 5 minutes," "When the ERP is not working, the latest data is used, but no more discounts are given" to ensure system consistency and predictability.

Case Study: Building a micropayment platform for Playwing

5. Checklist: Data required for an AI implementation in sales

The data contract establishes the schema, semantics, data preparation, and SLA for the company data affected by the AI implementation.

In brief:

See code and configuration examples for guardrails and data ingestion in Ch. 4, for Build and Managed.

Here are two technical fine-tuning challenges encountered in real AI implementation for sales.

The most popular models work differently and must be adapted to the specific moment in the customer journey:

In the Build (custom) approach, these models are defined and trained separately, with the reuse of some common components.

In the Managed (Cloud Native) approach, you start by configuring predefined models by Google for the popular cases above. You get faster delivery and the ability to choose the maximized business goal, without being able to modify the base model.

To implement packages, we need data engineering or a hybrid approach.

1. Smart purchase discount (standard)

2. Bundles directly in the catalog

The first variant involves modifying the shopping cart and the ordering module. The second variant involves deep integration with automated commercial flows.

In addition to these challenges, there are others related to scalability, latency, security, and monitoring, depending on the complexity of the company's workflows. But AI technology for B2B sales exists, as shown in this chapter.

×

Not sure what architecture fits your company? OPTI Software offers a free technical audit.