The Role of Contextual Information in Enhancing Agentic Systems

Authors: Alexon Bell, Chief Product Officer, FinCrime & KYC, Quantexa and Cat Mackay, Product Manager, Product Innovation, Quantexa

The potential of agentic AI is currently limited not by the models themselves, but by the quality of the data they consume. Our research demonstrates that augmenting agents with contextualized data via the Quantexa Agent Gateway delivers significant improvements in accuracy and the ability to perform complex research. 

The industry is buzzing with excitement at the prospect of new ways of working powered by agentic platforms. We hear multiple announcements of rapid progress building agents to solve different problems, which are fantastic. The world will benefit from the ability to rapidly create and manage multiple agents, which in turn solve problems and transform processes in multiple sectors.  This will continue to be an area of ongoing innovation and exciting progress. 

At Quantexa, we leverage these agents and want to share some insight into how to uplift their performance. In our experiments, we find that better contextualized data delivers ~40% better risk mitigation than un-contextualized data. Whilst un-contextualized data can still help agents catch “obvious” fraud, only contextualized agents detect more complex operations.  

In our experiments, we find that better contextualized data delivers ~40% better risk mitigation than un-contextualized data.

Only 4% of Gen AI projects went into production in 2025, with one of our clients running 750+ PoCs, indicating the scale of innovation. But the success rate is too low.  “…At the heart of good AI is good data…” and while many companies are focusing on the “good AI” part, we have been focusing on the “good data” part, and it has been a truly enlightening journey.   

Not many people can say that working with data is exciting, but it is the nourishment for good AI, and we are starting to see some clear steps that can deliver orders of magnitude improvements to AI model performance.   

Many readers will know that 80% of data science is preparing the data for modeling.  Even today, with the substantive investments in data infrastructure, processing and quality, what is evident is that the real world is messy, and the humans that enter the data into systems are at times sloppy.  These small data issues at entry are compounded by the aggregators, which exacerbates the problem downstream into an inaccurate agentic response.  

Why we tested AI agents

Many agentic AI projects fail to show ROI or create productivity improvements with real value. Only 14% of senior leaders report that agentic AI technology has been fully implemented in their organizations, showing a divide between senior leaders’ commitment to agentic AI and its full-scale adoption.  

At Quantexa, we saw an opportunity to “show, not tell,” taking a methodical approach to assessing the value of better data for AI agents. According to the often-cited 80/20 rule of data science, 80% of time is usually spent finding, cleaning, and organizing data, leaving only 20% of the time to actually perform analysis. Our objective was to show that contextual data empowered AI agents to deliver better results, without the wasted time of preparing faulty data.  

Only 14% of senior leaders report that agentic AI technology has been fully implemented in their organizations, showing a divide between senior leaders’ commitment to agentic AI and its full-scale adoption.

The experiments  

We conducted the AI agent testing within Q Labs, where we can model real-world “what-if” scenarios. In this case, our hypothesis was: Can contextualized data increase the accuracy of AI agents, since the agents can pull information from a more accurate graph?  

To test this hypothesis in Q Labs, we used three AI agents, each running with and without Quantexa Agent Gateway, which manages distributed agent networks. The tests were structured to compare the accuracy and usefulness of each agent response. The responses were judged by a range of factors, including the impact of adding Agent Gateway.  

The chart below details the purpose of the three agents and the data inputs they access. (Note that the agents are prototypes within Q Labs and are not available on the Quantexa platform.) 

 For each agent, we simulated a series of controlled variations. The primary variable was the agent's access to Quantexa's contextual data. The performance of each agent was measured by comparing the outputs of the agent when it queried Quantexa, to the outputs when the agent did not query Quantexa. 

For example, we gave the Insurance Agent a known suspicious insurance claim. We then compared the agent’s risk rating for that claim, with access to Quantexa contextual data and without this data. This approach provided a clear metric to assess whether the agent with Quantexa provided a more complete view of the claimant and their claim, than the test without Quantexa contextual data. 

The Results 

We first examined test results by agent. 

The Simple Agent  

This AI agent was tested by searching the web for various companies and people, with and without Quantexa Agent Gateway. We repeated the experiment multiple times with multiple companies and people to gather results. 

The prompt for this experiment was, “Please provide information about the COMPANY/PERSON XYZ.” Although it’s difficult to measure how much more content the agent gathered while using Quantexa versus not using it, we did gather many details to compare results.  

Below is one of the comparisons of the two sets of results for the Simple Agent, highlighting the richer and more numerous details captured with AI agents accessing Quantexa’s contextual data: 

Key difference summary: 

The Insurance Agent  

The Insurance Agent was tested using four variations, featuring the prompt, “Please triage this insurance claim” for each test variant. 

Variation A (control: with Quantexa, known suspicious claim): The AI agent, with access to Quantexa, had to triage a known suspicious claim. The agent correctly identified that the claimant was connected to known fraudsters and found suspicious evidence from various sources.  

Result: Risk score of HIGH (9.5/10) 

Variation B (without Quantexa, known suspicious claim): The agent was given the same task as in Variation A, but without access to Quantexa. The agent still identified a high risk but did so with less certainty.  

Result: Risk score of HIGH (8.5/10) 

Variation C (with Quantexa, less-suspicious evidence): The agent, with access to Quantexa, was given a claim with a smaller amount of suspicious evidence.  

Result: Risk score of HIGH (9/10), due to detection of a network of fraudulent activity and individuals via Quantexa’s graph. 

Variation D (without Quantexa, less-suspicious evidence): The agent was given the same less-suspicious claim without access to Quantexa.  

Result: Risk score of MEDIUM (7/10) 

The chart below summarizes the Insurance Agent test results. Results shown are the exact agent outputs. 

Another point to note is the subtle variation in formatting that occurs every time a test is run. These differences are due to the non-deterministic nature of an LLM and therefore have not been analyzed in this experiment.    

The Banking Agent  

Like the Insurance Agent experiment, the Banking Agent experiment tested the agent's ability to perform a triage investigation. The prompt for these tests was, “You are a financial crime analyst – please do a triage investigation for the latest alert.” The alert, target entity, and evidence of risk was varied multiple times, and as always, was tested with and without the use of Quantexa.  

The results were very similar to those of the Simple Agent and Insurance Agent. For example, in the tests using limited risk indicators but for a company that’s connected in its Quantexa graph analysis to risky entities, the alerts were more likely to be detected with the use of Quantexa.  

Since this was not a quantitative test, we collected the specific responses in the chart below, explaining what Quantexa provided to the alert-triaging test as well as some of the critical findings attributed to Quantexa. 

What Quantexa provided: 

Critical findings attribution:

What we learned 

The experiments proved that the efficacy of enterprise AI agents is determined by the quality and context of the data they access. Across all agents and experiments, the integration of Agent Gateway consistently and significantly elevated performance and improved results. 

The tests showed not just incremental results. For example, the Simple Agent tests uncovered a wealth of interconnected information that was otherwise invisible in the “without Quantexa” tests. In the Insurance and Banking agent scenarios, the impact was even more critical. The ability to see a holistic, contextualized graph of relationships allowed the agents to correctly identify high-risk situations that might otherwise have been assessed as moderate or overlooked entirely.  

Why context matters  

The Insurance Agent's re-evaluation of a claim from a medium (7/10) to a high (9/10) risk score, based solely on contextual network data, demonstrates the value of this data to analysts, and to teams looking for more ROI from their agentic AI investments. 

Contextual data is the catalyst for accurate, trustworthy AI agents.

In a world where data quality can be highly inconsistent, better decision making comes from organizations with better data. Poor data will generate only limited insights from AI agents, or worse, miss key indicators and leave critical information hidden.  

This is why a contextual data layer is not just a “nice-to-have.” Contextual data is essential for getting the most value from AI agents, especially for organizations that want to operationalize agents at scale.  

Bridge the agentic AI hype gap 

Our findings prove that contextual data is the catalyst for accurate, trustworthy AI agents. As enterprises look to successfully gain returns on their investments in AI, they must shift focus to the underlying data infrastructure that feeds critical decisions. A robust “data glue,” like that provided by Quantexa Agent Gateway, unlocks the true transformative potential of agentic AI in the enterprise. 

Learn more about Agent Gateway.