Track · AI in Product4 min read

Semantic Knowledge Graphs Earn Their Keep in AI Agent Work

Cosentino reduced credit-related order delays from 22 days to one, while a human-governed semantic layer gave AI agents the context to flag operational and financial risks.

What is a semantic knowledge graph, and does your AI need one?
What is a semantic knowledge graph, and does your AI need one?3844328 / pixabay

Context

  1. Cosentino reduced average credit-related order delays from 22 days to a maximum of one day.

  2. Cosentino reported up to $10 million in first-year savings and expects $100 million by year-end.

  3. The company began its semantic-layer project three years ago and targets completion by year-end.

  4. Fifteen departments previously used separate dashboards; an AI agent can now predict whether an order will arrive on time and identify the holdup.

  5. A paper released at the end of 2025 by UC Santa Barbara and JP Morgan AI Research found accuracy improvements of 26% or more over standard chain-of-thought baselines.

Cosentino expects $100 million in savings by year-end from its semantic-layer project, after reducing credit-related order delays from an average of 22 days to a maximum of one day.

Cosentino had a data lake, lakehouse, ERP and CRM. The semantic layer explains what the data means; collecting more cannot explain why a KPI turned red. “Agentic AI is AI that can do things, not just tell you things,” said Rafael Domene, Cosentino’s global CIO. Agents need business knowledge, he added.

A graph database stores entities as nodes and relationships as edges, not records in fixed columns. Any node can connect to another, letting teams trace supply-chain dependencies, detect fraud or map preferences for recommendations. Queries follow edges to related entities.

A knowledge graph sits atop that database and describes its relationship map; a semantic graph adds enterprise meaning, such as whether a company is a client or prospect, a person a sales contact or a title a signal of importance.

Building this layer historically required subject-matter experts and continual updates as operations changed. It rarely justified the effort when analysts already held the context. Agents inherit neither general knowledge of company operations nor specific knowledge of the organization they must serve.

What changed at Cosentino?

Cosentino started the layer three years ago and aims to finish by year-end. Main processes and flows are in place; remaining work covers the financial side and logistics, manufacturing, operations, customer service and sales, procurement and HR.

Fifteen departments each maintained dashboards, and Cosentino often learned late that an order would miss service. An AI agent can now inspect any order, predict whether it will arrive on time and identify the holdup. Missing raw materials and factory failures are cited causes.

Early warning lets a salesperson contact the customer, explain the problem and propose solutions. The agent must also distinguish this order from a previous one and determine the next action, which is the semantic layer’s value beyond retrieval.

Blocked orders caused by customers exceeding credit limits previously averaged 22 days; the maximum delay is now one day. Agents also took routine case handling away from the finance team, freeing staff for customer work, Domene said.

The project produced up to $10 million in first-year savings, with revenue effects beyond cost savings. Sales employees previously spent more than one-third of their time on administration.

Cosentino began building sales agents last year, tested them this year and expects them to be fully operational soon. Not all saved time will become revenue, Domene cautioned, though he sees substantial potential.

The semantic layer began as manual work by Cosentino experts and engineers from Celonis, the company’s selected process-mining vendor. Domene said SAP’s fixed flows and context make it relatively easy to model, but people still did the work. AI later augmented that process; it has not replaced it.

David Hoelzer, a SANS Institute fellow and chief of operations at Enclave Forensics, warns that people do not always create documents in ways AI can interpret as semantic relationships. AI can accelerate construction and maintenance, but autonomous work may encode nonsensical links, so specialist review remains necessary.

Why keep vector databases alongside knowledge graphs?

About a year ago, the standard route for adding information to a large language model was retrieval-augmented generation, or RAG, with a vector database. It organizes unstructured material so similar topics sit together, allowing prompts to retrieve related context. RAG remains useful for unstructured data in conventional pipelines.

RAG can also retrieve too much, including dated or confusing passages. More text consumes more tokens, raising cost, and agents often make repeated model calls. A knowledge graph finds exact information; the semantic layer explains what it means and how to use it.

New York Life Group Benefits Solutions uses both. Its knowledge graph is built on top of LangGraph, the AI-agent framework from LangChain. Matt Marze, the company’s CIO, assigns knowledge graphs to multi-agent orchestration and the vector database to generative-AI RAG applications.

The graph begins with business information, data, signals and broker-community data. Its semantic layer adds product and plan definitions, relationships, current context, history, workflow and priorities. Marze said these inputs drive insights, predictions, recommendations and agent coordination.

James Kaplan, a distinguished partner at McKinsey & Company, reported rising adoption: 18 months ago his interest met disbelief; the past year brought uptake among institutions using graphs for complex decisions and stronger agents.

Where does the semantic layer fail?

The approach still inherits the weaknesses of probabilistic models. Hallucinations can compound across multiple calls to a large language model, especially for company-specific work.

Zakir Hussain, EY Americas’ data and AI leader for industrials and energy, said: “All the models are probabilistic ones.” Enterprise questions expose missing or misinterpreted context, he said, while generic training often fails when a system must answer company-specific questions.

A paper released at the end of 2025 by UC Santa Barbara and JP Morgan AI Research found that knowledge graphs improved accuracy by 26% or more over standard chain-of-thought baselines. Hussain called the work important because growth, efficiency and productivity require companies to codify context.

The likely near-term practice pairs human-governed semantic layers with agents, with product teams judging them on decision accuracy and token cost rather than retrieval volume.

via cio.com (Source)

More from Priya Raman

Priya Raman

Show full bio

News editor covering money, funding and market shifts around product management at Roadmap File.

78 articles