Step-by-Step Seller: Next Best Action in B2B with Google Cloud, CRM, and Business Rules

1. Specific B2B Actions, Recommended with Precision

In e-commerce or B2C, a product recommendation may be sufficient, but in B2B, sales take into account order history, framework contracts with negotiated prices, credit limits, outstanding payments, logistical conditions (delivery from a specific warehouse within a certain time), technical compatibility, or seasonal budgets, which are not visible in a simple catalog.

Examples of step-by-step actions:

See all in section III

The challenge and opportunity lie in adapting to the company. A good Next Best Action system generates few but highly relevant actions and notifications. Otherwise human agents will prefer their own experience and it will not help them.

What does a recommended action look like in the CRM?

Next Best Action cards can be placed in the company's CRM next to each customer, lead, or opportunity. They must answer four questions for the sales agent:

• What do I need to do?
• Why now?
• What commercial offer should I propose?
• What contractual or commercial constraints must I respect?

The sales agent receives from the software a logical recommendation, with an explanation and explicit limits, which makes their work easier.

Which industries does it fit?

2. How Do You Ensure Quality? Through Hybrid Architecture: AI Proposes Based on Company Data, but Rules Decide

With secure access to company data, AI can identify opportunities, make predictions, and generate intelligent explanations. But the recommended action must always be verified through strict business rules (guardrails).

AI proposes. Business rules decide what reaches the agent

The essential rules for almost all B2B companies include respecting margin, stock, overdue payment policy, maximum discount, and contractual specifications.

For this reason, the recommended architecture looks like this, starting from data, passing through AI, and applying the company's rules:

Clean data ensures the precision of the recommended action

An NBA module translates data into actions, so quality starts with data and only then do LLM prompts and AI technologies in general start to matter.

It is important that data is interoperable and standardized. You need an intermediary layer in the data warehouse, which should include both a copy of data from different systems and data quality rules, so that the architecture holds up in the long term. It will look like this:

Success is ensured through this solid foundation on which AI will then operate.

3. From the Product Catalog to the AI Explanation for the Sales Team

Product recommendations: what the customer might buy

The first intelligent layer will be the product engine. The functions available for products in NBA software can include:

• products frequently bought together
• similar products
• complementary products
• products suitable for the customer's history
• products recommended based on customer segment
• products with available stock or priority rotation
• alternatives for unavailable products

In Google Cloud, the AI Commerce Search (Vertex AI Search for commerce) product can be used for commercial search and recommendations, especially where there is a catalog, customer interaction events, and a rich purchase history.

In B2B, however, this recommendation must be adapted. An industrial customer does not buy the same way as a final consumer in B2C, and above all, essential commercial rules apply in B2B.

For this reason, a Next Best Action is more than a product recommendation engine. The recommendation is just one of the system's inputs; the recommended actions will vary depending on the specifics of each company and its customers.

The right moment: when it makes sense to act

A good recommendation sent too early may seem aggressive. The same recommendation sent too late may lose the order. Therefore, a Next Best Action module must estimate the right moment from the signals available in the data sources and the recommendations of CRM systems like HubSpot. A few examples:

• The customer orders every 30-45 days
• Estimated consumption indicates stock depletion
• The sent quote has not received a response in 7 days
• The contract is approaching renewal and the customer has been recently active
• Internal stock must be rotated before a certain date
• Some products are seasonal or on promotion, including bundles
• The customer has started buying from a competitor (if you have data)

For such scenarios, technologies like BigQuery ML (with AI.GENERATE, AI.SEARCH, and AI.FORECAST functions) and AI functions in BigQuery can be used for classification, scoring, time forecasting, or time series analysis.

As the system receives feedback from the team and ultimately from customers through purchases, the rules can be refined in more detail or replaced by automatic predictive models.

The recommended action now: what the team must do based on the rules

Once we have the products and the analysis of the right moment, the system must choose the action. This correlative table of recommended action vs. ideal situation can help, which you can supplement based on company experience.

This list must be configured together with commercial management. AI can suggest patterns, but the company must decide which actions are acceptable. In particular, the action must cover at least the current commercial activity, so the module is an upgrade.

Here too, we must protect the company and its image through two levels of explicit rules (guardrails).

a) Deterministic business rules

These are applied in SQL, code, or a policy engine (rules engine). They must not be left to the interpretation of the LLM model. Examples:

• If there are overdue invoices over 30 days, recommend a payment clarification action
• If the estimated margin is below the minimum threshold, remove the recommended product
• If there is no deliverable stock, remove the product and propose a compatible alternative
• If the product is not approved for the customer's segment, send a request for inclusion in the segment

These rules are auditable and under commercial management's control. They are reflected directly in the NBA with explanations, so the team knows why a product was proposed or removed.

b) Guardrails for the generative LLM model

To avoid sensitive data leakage, block prompt injection attacks or unauthorized use (e.g.: scheduling a website in the company's chatbot), and control the company's data sent to AI, it is recommended to filter the input and output of LLMs.

See how Model Armor and agentic AI governance mechanisms can be used to control LLM interactions.

The action explanation: how the agent knows what to discuss with the customer

AI can explain, even if it does not decide on its own, why a customer receives a discount, if a product is being proposed, or if a payment overdue is being deferred. These decisions are controlled through data and deterministic rules, as shown above, but the agent must know what to discuss with the customer. AI is excellent at explaining a situation in easy-to-understand language.

The role of an LLM like Gemini is best suited for synthesis, explaining to the agent why the recommendation appears. It is also excellent for formulation, proposing a message, email, or short phone script, and for context: extracting information from documents, CRM notes, technical datasheets, or commercial policies, for example through a chatbot attached to the sales software.

Human feedback: how the system improves

The sales agent must be able to indicate: whether the recommendation was accepted, whether it was wrong and especially why (unsuitable product, wrong timing, customer already contacted). Additionally, in the CRM you will track whether the sale was ultimately recorded or the opportunity was lost. This feedback is essential for the system to permanently adapt to the reality of the sales cycle.

4. The Secret to Success: CRM Integration and Control of the Entire Operational Flow

Principle 1: The recommendation will be delivered where you already work

From experience, a Next Best Action module has value when the recommendation is delivered where the agent already works:

For example in HubSpot, as an App Card / UI extension on a company, contact, or deal. Or in Salesforce, as a component on the account or opportunity page.

If urgent alerts are needed, they can appear in Slack or Microsoft Teams. And when the team is disconnected or in the field, the solution can be integrated into the internal portal, the SFA solution used, Slack, or even WhatsApp for field teams.

Principle 2: Control of the entire operational flow, at IT and Management level

A simplified flow for all the aspects presented above is:

1. The system updates commercial data in the data warehouse (BigQuery) from data sources
2. Signals are calculated: recurrence, churn, stock, margin, opportunity
3. The recommendation engine identifies relevant products
4. The action engine decides the type of recommendation
5. Business rules eliminate impermissible recommendations
6. An LLM like Gemini generates the explanation and phone script
7. The card appears in the CRM
8. The agent accepts, modifies, or rejects the recommendation
9. The result is saved for system improvement (feedback)

For a mature system, this flow runs daily, every few hours, or in near real-time, depending on the company's rhythm. Management and the IT department have control over it, even though it runs automatically, so they can diagnose issues at any time.

Of course, not all companies need real-time synchronization. In B2B distribution, for example, a daily batch may be sufficient. In enterprise or medical sales, where commercial cycles are long, recommendations can be generated based on events. But in principle, all actions are linked to each other in this complete flow.

Principle 3: avoid risks through planning and flow control

Here are some development risks and things to watch out for in a Next Best Action module for B2B sales.

Why OPTI?

OPTI approaches Next Best Action projects based on our experience as a Google Cloud partner for AI technologies and HubSpot CRM partner for internal workflows. Our B2B software development experience leads us to emphasize the value of procedures already existing in companies. We integrate AI into already existing workflows, without replacing the software the company runs on.

A "step-by-step seller" system saves the time lost searching for context, provides relevant actions for the employee, and connects data from IT systems with day-to-day commercial results.

Contact us now for a free audit.