Hi, I’m Sean, Senior AI Product Manager at JSI. Thanks again for joining us for part five of our AI Essentials for Investigative Intelligence video series, with this episode focused on emerging AI systems. This episode will look at how investigative technology as a whole is evolving to support law enforcement and intelligence teams at scale. Let’s dive in.

Up to this point, we’ve looked at individual AI capabilities, including natural language processing, computer vision, and generative AI, largely in isolation. That’s useful for understanding what each technique can do, but it’s not how real investigative data actually shows up. In practice, data is messy, fragmented, and spread across many formats and sources, including text, images, transactions, communications, and metadata. All tell part of the story, but rarely in one place.

For investigations, the real value comes from connecting those dots, linking outputs from different AI processes into a shared representation that supports reasoning, correlation, and even hypothesis testing. This episode is about stepping back and looking at how all of these pieces fit together, and how processed data flows into a coherent system that transforms raw information into integrated, actionable investigative insights.

Understanding the complete pipeline for collected data is just as important as understanding any given component within it. As an example, you may be facing large amounts of text, audio, or visual content in your investigation. At this point, tools like natural language processing, speech to text, computer vision, and translation can enrich the raw data into something more useful.

Building on these outputs, and depending on them as well, are further downstream processes, such as classifying the data into important categories or risks that need to be flagged to end users for further analysis. This is especially important in triage workflows. Finally, as the outputs of all of these tools are generated and stored within the system, searching and analyzing that data is also dependent on those outputs, making it critically important to understand the transformations and outputs realized throughout the life cycle of a given record.

At the end of all of this, you’ll see tools like Q&A or analysis with large language models, or applications built on large language models.

At JSI, we’ve been developing a new AI-powered conversational search assistant called Iris, which has been built primarily to help streamline the identification of relevant content, insights, and trends from data collected within our platform called 4Sight.

Given the natural and approachable interaction with Iris through conversation, users can achieve their objectives much faster than with traditional methods, without requiring full expertise in all of the 4Sight functionality that exists today. To ensure transparency in the answers provided, all responses from Iris include interactive citations, which provide traceability back to the source data. This allows users to fact-check responses and further explore referenced records that could help provide an investigative lead.

Tools like Iris are far more complicated than they may seem at first glance. For example, if we take the prompt, “Does Kenzie make any references to rock salt?”, this is a valid question that you might ask of your investigative data. You might get back an answer saying that yes, she does, along with details about when she had that conversation. However, it’s important to look at the different steps this type of tool needs to take in order to produce that answer.

There is a lot more to the iceberg underneath. First, the system needs to understand the question. What language is it in? In this case, it’s English. Who are the main entities it needs to search on? What are the things it needs to run within the system? What tasks does it need to create and follow in order to generate an answer?

Then it needs to figure out who Kenzie is. Do we already have a representation of a Kenzie entity within the system, or will this require additional searching across the database to find references to this individual? From there, the system may need to apply other filters, if necessary, to specific kinds of content or specific date ranges.

Lastly, the system needs to run different text searches against the content, including both full-text matches and semantic similarity searches, to find all possible relevant records that need to be considered to generate the final answer. That answer is then summarized and produced for the user to review, complete with inline citations. This is far more complex than it may seem when a user simply enters a prompt and receives an answer back.

It is important to understand all of the steps needed for that answer to be created and provided. We also use another technique behind the scenes when retrieving information and searching across the database to help mitigate the possibility of hallucinations, while ensuring users have full traceability into the source data. This technique is called retrieval augmented generation, or RAG for short.

RAG helps ensure that any large language models being used to analyze and summarize information are grounding their responses in the real data that users have access to and permission to access. It also allows the system to intercept when the user does not have permission to results, or when there are no search results that meet the search criteria.

For instance, a user might ask about rock salt, and there may be no records. Instead of letting the large language model make something up, the system can say that the search results yielded zero results and return that to the user. This makes it clear that there was nothing in the database, rather than allowing the model to fabricate information.

Another key design pattern used within Iris is the use of agents, tools, and workflows, which help provide more agency to the system to solve tasks at hand. Specifically, agents are instances of AI models, typically large language models, that have been provided the autonomy to solve a given task through reasoning, planning, and tools.

The main pattern is that agents can decide which tools to use, when to use them, what to do after receiving the output of a tool, and when the task provided to them has been completed. Agents can have one or more tools assigned to them depending on the type of tasks they were designed for.

Tools are specific functions that allow the agent to do things, whether that means searching for data, interacting with various systems, or carrying out actions. The tools assigned to a given agent are typically focused on specific tasks or objectives. For example, a geospatial agent may have tools for geocoding or proximity analysis.

Agents do not necessarily need to be granted full autonomy from start to finish for tasks. They can be used strategically within a workflow, which is a predefined code path that incorporates agents for key decision points while also incorporating hard-coded logic to guide other decisions and outputs once the agent has finished its work. Workflows are especially useful for providing more predictability in complex projects.

To help bring this concept home, here is a demo from the NVIDIA website showing an agent that has access to tools. In this case, there are three tools, one of which is called “Get the current weather.” We enter a prompt asking for the current weather in Ottawa, and the agent has to make an explicit decision that it has this tool at its disposal. It then calls that tool to fetch the current weather.

The tool itself is executing normal computer code. It runs an API request to a weather service and retrieves the current temperature for the location that was requested. The answer provided in the user interface is accurate relative to what a quick Google search would provide as well.

The agent itself is not all-knowing. It does not know the weather in Ottawa today on its own, but it can use the tool to delegate that task, retrieve the current weather based on an API, and send that back to the user with factual, grounded information.

As we wrap up Part 5, I hope this has helped underline the importance of whole-systems thinking when applying AI to investigative work. The real value is not necessarily in isolated models or individual tools, but in how these systems work together to reduce manual effort, surface relevant insights earlier, and ultimately give investigators a more complete and coherent picture.

The final part of this video series is next, where we’ll talk about a topic that underpins all of this: compliance and ethics. Thank you for watching. I really hope to see you in the next part. If you want to learn more about JSI, please visit our website at jsitelecom.com. Thanks.