New AI technologies for B2B
In 2025, AI technologies matured that are not yet widely in production:
- Agentic AI: agents that negotiate, configure, make decisions, and take actions.
- GraphRAG: combines vectors with knowledge graphs for verified reasoning, e.g., for product compatibility.
- Composite AI and generative configuration: generating a complete configuration (e.g., BOM), not just choosing a single product.
A company that creates an AI core will be able to adopt agentic automation in 2027 or even 2026. The core means clean data and predictive AI integration plus LLM applications (e.g., chatbots). Even agentic AI will not lead to the abandonment of the central software where the company's prices, stocks, and commercial policy reside.
What you build today is the foundation for using tomorrow's AI agents.
What should we use in B2B sales?
We recommend predictive AI as the primary sales engine in B2B for two reasons. First, stability (response times in milliseconds in optimized scenarios, not seconds) and auditability (behavior that is mainly repeatable and explainable).
Although extremely powerful, agentic AI is still characterized by:
- High latency: several seconds, especially when launching chains of actions.
- Higher inference costs: at least compared to an optimized recommendation system.
- Behaviors difficult to guarantee 100%: hallucinations, loops, unforeseen actions.
The hybrid architecture described in this series of guides:
- Allows the development of an AI sales engine, consisting of Deep Learning, RAG, business rules, plus LLM applications (chatbots).
- Requires a solid data layer: clean and normalized data, a data warehouse (e.g., BigQuery), functional APIs to ERP, WMS, CRM, and e-commerce.
- Guarantees that the data layer can be used by AI agents. Without it, neither Deep Learning nor newer technologies can function.
Negotiation and sales with AI: the launch of UCP
Today, AI suggests a product or a package. With agents, it can negotiate terms (price, discount, payment terms) on behalf of the seller or the buyer.
Since December 2024, there has been MCP (Model Context Protocol), an open standard for connecting AI agents to data systems to understand company data within established limits.
In January 2026, Google and Shopify launched UCP (Universal Commerce Protocol), a dedicated standard for the commercialization of products and services through AI agents (Source). It allows agents to search, compare, negotiate, and perform transactions without human intervention.
What will the future look like?
Mainly for B2C:
Agents that offer a dynamic discount range within a chatbot, depending on customer history, the current basket, and purchase intent.
Frameworks such as AutoGPT, LangChain Agents, and Autogen allow for the definition of objectives. E.g.: "maximize margin" / "move old stock".
Mainly for B2B:
The agentic flow is more ambitious in B2B.
AI agents understand company data (via MCP), for example, to use complex product configurators or any information the company wishes to provide. Agents order products and services (via UCP), negotiating with each other.
Merchants who do not adopt AI can rely on classic APIs, but in a world where buying is repetitive and standardized, as AI agents can work 24/7, competitive pressure will arise.
| Step | What does the AI agent do? |
|---|---|
|
1 |
Prospects the market by connecting to:
|
|
2 |
Queries suppliers via MCP or UCP (it does not visit the website) and analyzes:
|
|
3 |
Negotiates with a list (short or long) of sellers. |
|
4 |
Sends the purchase decision via API or UCP. |
|
5 |
Pays, tracks, and confirms delivery, without human intervention. |
In brief:
- Companies that do not have functional APIs or access to their own data cannot adopt AI agents.
- For buyers, AI agents need clear ceiling rules, a complete decision log, and the possibility of override by human management, to avoid risks.
- For sellers, product recommendations can no longer be just visual (carousel and photos). Structured data must be exposed: margin, TCO, SLA, delivery times, certifications.
- For both parties, if internal systems (e.g., ERP, WMS) do not have APIs, and data is not standardized and normalized, the buyer AI agent either cannot query at all or receives incomplete data and penalizes the seller.
From similarity to reasoning
Semantic search says "these two products are similar". Knowledge graphs can say "these products are linked: if you take X, you also need Y".
GraphRAG combines vector search with a knowledge graph (e.g., Neo4j, GraphDB, graphs in BigQuery) that models relationships of compatibility, dependencies, or valid substitution between products.
Example: 110mm Pipe 110mm Elbow Compatible Gasket
Connections are through deterministic (classic) rules, which is why AI hallucinations can decrease.
Integration between an LLM (e.g., ChatGPT) and graphs will become native in many AI products. This will lead to increased accuracy when AI needs to work with a real catalog of B2B products or services.
Quoting a complete solution: Bill of Materials (BOM)
What happens if we combine several types of technologies?
LLM technologies, semantic search, graphs, and deterministic constraint engines are ingredients in generative configuration.
- Not just one product, but a complete valid configuration.
- Bill of Materials (BOM) are complex offers of tens, hundreds or thousands of components logically linked to a project plan.
- Ideal industries include distributors of complete equipment-based solutions: HVAC, electrical, building materials, IT infrastructure.
- There are already successful projects for professional services as well. E.g.: IT maintenance offers.
Example: installations
An architect customer (or contractor) has a PDF building plan, attaches it, and writes: "I need a complete underfloor heating system for a 500sqm warehouse, with medium-level insulation".
The request reaches a quoting engineer who now works for 48 hours to create a BOM.
In the future, a generative configuration agent sketches the offer in minutes.
It uses structured data from the ERP and technical installation rules from graphs (see above) to guarantee that the project can be built and the individual products are in stock.
| Step | What does the Composite AI agent do? |
|---|---|
|
1 |
Concept decomposition with LLM: the needs are pipes (approx. 3000m), manifolds, automation, elbows, clips. |
|
2 |
Searching and validation with RAG and graphs: which products are in stock and are compatible. |
|
3 |
BOM generation with Constraint Engine verification: Assembles the product list with 150 lines of different quantities, verifying constraints noted in the Constraint Engine. E.g.: "Does the pump handle the flow required for 3000m of pipe?". |
|
4 |
Quoting in the company template: generates the offer with the total price plus the assembly diagram. |
Distributors selling instantly generated solutions will have an advantage over those selling items piece by piece.
Roadmap 2026-2027
Companies that have not yet adopted Gen 3 AI can follow this action plan based on hybrid architecture:
|
2026 |
The Foundation: Gen 3 via hybrid architecture |
|
|
2026–2027 |
Activating agentic and composite components (Gen 4) |
|
|
Continuously |
Gradual expansion |
|
This plan is designed for companies wishing to use AI opportunities for growth and scaling, with technological updates justified by results.
Data cleansing and creating a hybrid architecture that brings the benefits of AI is an infrastructure investment. If company data remains locked in closed software, not only do human customers have a poor experience, but AI agents will not be able to negotiate with it either.
See the sales deliverables built on this basis in Ch. 3
How ready is your company for the future? OPTI Software offers a free technical audit
Continue Exploring
Quick Questions
What is the difference between predictive AI and agentic AI?
Predictive AI recommends, agentic AI acts. The key difference is the need for control, auditability, and explicit permissions.
What is GraphRAG and why is it relevant?
GraphRAG combines knowledge graphs with LLMs, enabling answers grounded in deterministic relationships rather than text alone.
What does UI-less commerce mean?
Scenarios where AI agents negotiate, check stock, and generate offers without users navigating a traditional interface.
What are the risks of agentic AI in sales?
Unauthorized decisions, lack of traceability, and commercial errors. Agentic systems without guardrails are a liability, not an advantage.
How should companies prepare for 2027?
By building the right foundation in 2026: clean data, clear APIs, explicit rules, and strong governance.
What is the TLDR (conclusion)?
This chapter avoids futurism and focuses on implications: agents increase autonomy and risk; without guardrails, audit trails and clean data, gains turn into incidents.
What technologies and methodologies are involved?
Technologies: Gemini, Vertex AI, Vector Search, Knowledge Graph, BigQuery, APIs, IAM, KMS, Cloud Logging, Monitoring
Methodologies: risk assessment, control plane, policy enforcement, human-in-the-loop, audit trail, retrieval-augmented generation, graph constraints, secure tool-use, red teaming
![[UPDATE] The Romanian startup awards will be presented on March 12. Mr. Benny won 3rd place at the Romania Startup Awards 2026.](/images/new-post/small_long-romania-startup-awards-2026-mrbenny.jpg "[UPDATE] The Romanian startup awards will be presented on March 12. Mr. Benny won 3rd place at the Romania Startup Awards 2026.")
