Genie Ontology: Achieving 1.6x Accuracy with Half the Tokens for Enterprise Data Agents

Genie and Turing launch an enterprise data agent benchmark where Ontology delivers 1.6x accuracy with half the tokens.
The article introduces a benchmark for enterprise data agents developed by Genie and Turing, and Genie Ontology's performance on it: 1.6x accuracy improvement with less than half the token consumption of comparison solutions. Enterprise data is scattered, table structures are complex, and semantics are implicit—traditional approaches that let LLMs figure things out from massive raw contexts are both expensive and unstable. Genie Ontology's value lies in pre-structuring entity relationships and business semantics, shortening reasoning paths and reducing ambiguity, thereby improving both accuracy and cost simultaneously. The article also notes these figures come from specific test conditions, ontology construction itself requires domain knowledge investment, and enterprises with weak data governance foundations need to factor in upfront costs. It recommends prioritizing semantic layer construction for core data domains and evaluating solutions using both accuracy and token efficiency metrics.
Enterprise data agents are becoming a key direction for AI adoption, but maintaining high accuracy while controlling token consumption in complex enterprise data environments has long been an industry challenge. Recently, the Genie team, in collaboration with Turing, developed a new benchmark for enterprise data agents and delivered a compelling answer based on Genie Ontology: accuracy improved to 1.6x while token consumption dropped to less than half.

The Core Tension of Enterprise Data Agents: Accuracy vs. Cost
Enterprise data scenarios differ fundamentally from general conversational scenarios. Enterprise data is often scattered across multiple systems and formats, with complex table structures, non-standard field naming, and implicit business logic. Agents need to understand the semantic relationships behind this "dirty data" to provide accurate answers.
The traditional approach lets large models "figure things out" from large volumes of raw data context, which creates two problems: first, enormous token consumption makes long-context reasoning expensive; second, models are easily distracted by irrelevant information, leading to unstable accuracy. The benchmark developed by Genie and Turing was designed precisely to quantify this tension—measuring both agent accuracy and token efficiency on real enterprise data tasks.
According to the source material, Genie with Genie Ontology achieved a 1.6x accuracy improvement on this benchmark, while using less than half the tokens of comparison solutions. This combination means significantly higher effective information output per token, with direct implications for the cost structure of enterprise deployments.
Why token consumption becomes a critical cost factor in enterprise scenarios
LLM pricing is directly tied to context length—input-side long contexts (table schemas, sample rows, conversation history) are billed per token, as is the output side. Enterprise data tasks often require multiple rounds of tool calls and schema exploration, with single Q&A contexts reaching tens of thousands of tokens; if daily requests number in the thousands, costs escalate rapidly. More critically, "effective information density" matters: irrelevant field samples and redundant table schemas not only cost money but also dilute attention and induce hallucinations. Therefore, token efficiency is not just a cost-saving metric but also an indirect safeguard for accuracy. The industry typically measures the cost-effectiveness of such solutions using "token cost per correct answer," which is exactly why the Genie-Turing benchmark measures both metrics simultaneously.
What Genie Ontology Does Differently
Ontology is not a new concept in data management—it describes a formal model of data entities, attributes, and their relationships. Genie Ontology's innovation lies in introducing this semantic structure directly into the agent's reasoning process, rather than having the model understand raw data from scratch.
Think of it this way: faced with an enterprise database containing dozens of tables, a typical agent needs to read large amounts of table schemas and sample data to "guess" field meanings; with Ontology support, the agent receives pre-organized entity relationships and business semantics, resulting in shorter, more focused reasoning paths. This explains why token consumption can drop below half—the model doesn't waste compute on redundant information.
At the same time, structured ontology information reduces ambiguity, making the model less error-prone on key judgments, which is why accuracy improves to 1.6x. Both metric improvements stem from the same mechanism, not independent optimizations.
How ontology differs from knowledge graphs and database schemas
Ontology is often conflated with knowledge graphs, but they differ in emphasis: schemas only constrain data storage structure (tables, fields, types) without expressing business semantics; knowledge graphs focus on instance-level facts ("Company A is a customer of Company B"); ontology sits between the two, formally defining concept classes, attributes, relationships, and constraint rules within a domain (e.g., "an order must be associated with a unique customer")—it is a reasoning-capable semantic layer. Knowledge graph instances can be built on top of ontology, and ontology can also map to relational database schemas. Genie Ontology belongs to this layer of domain ontology, mapping table structures scattered across multiple systems into a unified entity and relationship model, so agents only need to face a single semantic contract when querying rather than reading DDL and sample data table by table.
The Significance and Boundaries of the Benchmark
The value of this benchmark developed with Turing lies in targeting real enterprise data agent tasks rather than academic toy datasets. Evaluation of enterprise data agents has long lacked unified standards, making cross-comparison of different solutions difficult; a benchmark close to production environments helps the industry build consensus.
It's important to view objectively: the 1.6x accuracy and half-token figures come from specific test conditions of this benchmark. Enterprise data environments vary enormously, and building the Ontology itself requires investment—organizing entities and relationships depends on domain knowledge, and this cost is not reflected in the token ledger. For enterprises with relatively standardized data assets or existing ontology accumulation, the benefits will be more direct; for those with weak data governance foundations, upfront investment needs to be factored into evaluation.
What enterprise data agent benchmarks typically evaluate
Production-oriented benchmarks generally include three types of tasks: single-table/multi-table aggregation queries (testing schema understanding and join reasoning), fuzzy business questions to SQL (testing semantic disambiguation), and complex analyses requiring cross-system data retrieval (testing tool orchestration). Evaluation dimensions include not only execution result accuracy but also token consumption, number of call rounds, and latency. The construction challenge is that enterprise data cannot be made public, so evaluation typically uses desensitized synthetic databases or controlled private environments—which also means scores across different benchmarks cannot be directly compared. Turing has accumulated expertise in enterprise data annotation and evaluation infrastructure, which is the practical foundation for its collaboration with Genie on this benchmark.
Practical Implications for Enterprises
The directional significance of this work exceeds any single number. It suggests that enterprises should prioritize building the semantic layer of their data rather than simply pursuing larger context windows or more powerful base models. Ontology is essentially a form of "front-loaded intelligence"—crystallizing human understanding of the business so agents can take fewer detours.
On the implementation path, enterprises can consider prioritizing ontology construction for core data domains with high-frequency queries, then expanding gradually; simultaneously, they should evaluate agent solutions using both token efficiency and accuracy metrics, avoiding the trap of focusing only on accuracy while ignoring reasoning costs. This collaboration between Genie and Turing provides a reference sample for this evaluation approach.
Related articles

Gemini 4 Neon Review: Benchmark Scores Rival Carbon, API Pricing and Quota Breakdown
Google Gemini 4 Neon is now live, benchmark scores closely tracking Carbon, API pricing 3/15 at twice Hydrogen's rate. This article compiles Bilibili creator testing: Neon can replace 90% of Carbon scenarios, AntiGravity quota consumption is only one-third of Carbon, plus subscription channel restrictions and 2API ban risks.

Gemini 4 Helium Coming Soon? Signals of Google Loosening Subscription Restrictions
Google team member hints releases will continue; community speculates Gemini 4 Helium (Flash Lite) may be next. Subscription ban controversy gets a response as product lead says "Soon," suggesting restrictions may be lifted. This article covers Gemini developments, OpenAI Coding Agent and risk management, Doubao real-time audio-video API, and compute business dynamics.

AI-Assisted React Development in Practice: Optimizing a Video App and Handling Errors
A 14-year veteran developer demonstrates AI-assisted React video app optimization: compressing video from 80MB to 5MB (~94%) via ImageKit, plus reliable error handling and retry mechanisms. Includes planning, code understanding, and verification methods for working with AI.