//Door one: agents doing real work inside the firm
AI Agents for Investigation and Security Firms
An AI agent is software that carries out several steps of real work on its own. The question that decides whether it is an asset or a liability is not how to build one. It is how you check what it did on a live client file, and who has to approve before it goes any further.
By Matt Aubin, Founder, Southern Recon Agency and E3 Legacy Intel. CDFE, FBCI. Investigating since 2009. Updated .
The whole problem, stated once
An agent that nobody can check is a rumor with formatting. An agent with the approvals in the right places and a record you can read is deployable labor.
Everything on this page is about the second one, and it is written for the owner or the operations leader, not for a developer.
What is an AI agent, in plain terms?
An AI agent is a program that uses an AI model to carry out several steps toward a goal on its own, rather than answering one question and stopping. It can read files, search sources, call other software, hand work to itself and keep going until the task is finished or it gets stuck. Agentic simply describes systems built that way. The difference that matters to a firm owner is that a chatbot produces an answer somebody reads, while an agent produces work somebody else in the firm may act on without ever seeing how it got there.
How do you build AI agents for a firm?
Privately, for one firm, on the part of the work that is genuinely yours. A custom private build starts with the workflow you would most like off your people's desks, states what the agent is allowed to touch and what it is not, and puts the approval points in before the first line of it is written. It is built and maintained for one client and nobody else runs it. Most firms should buy the common parts and build one thing that is actually theirs, and working out which one is usually the first conversation rather than the last.
How a build is scoped, delivered and maintained is set out on the firms page. A build stands on its own and does not need an advisory seat beside it, though the two often run together.
How do you check what an agent actually did?
You read its record, not its output. An agent that cannot tell you what it looked at, what it retrieved, what it discarded and what it concluded from each source has not done work, it has produced a paragraph. So the build carries the record as a feature: every claim traces to a source that exists and says what the agent says it says, the run is logged step by step, and a named human signs that they checked it before it leaves the firm. On a live client file the check is done against the source document, word against word, not against the agent's own summary of the source.
Those checks are the Evidence-Grade AI Standard, and the signed AI evidence audit is the document that states a firm actually runs them.
Where should a human have to approve?
Anywhere the next step is hard to take back. Before anything reaches a client, opposing counsel, an insurer or a court file. Before money moves. Before a record is created that the firm will later have to stand behind. Before the agent acts on a conclusion it reached rather than a fact it retrieved. Everything else can run unattended, and the point of putting the gates in the right places is that you get the speed on everything reversible instead of paying a person to watch the whole thing.
Which job would you take off your people first?
Say the workflow in a couple of lines. You get a straight answer on whether it is a build, an advisory seat, or something you should not automate at all.
Do you need to write code?
No. An owner or an operations leader needs to be able to say what the agent is for, what it is allowed to touch, where the approvals sit and how its work gets checked before anyone relies on it. That is judgment about the work, not software knowledge, and it is the part nobody can do for you. The building itself is the part that gets hired out, which is what a custom private build is.
What breaks first?
Confident wrong output that nobody caught, because it reads exactly like the right output. After that, scope creep, where an agent built for one narrow job quietly gets pointed at work it was never checked for. Then the record, when a run cannot be reproduced months later and nobody can say what the agent saw at the time. Then handling, when client material moves through a tool nobody approved. Every one of those is a design decision made early, which is why the approvals and the logging go in before the capability does.
What does this have to do with what you run yourself?
He built a proprietary AI investigation platform and the investigation agents around it, and his own firms run on it on live cases. That is where the answers on this page come from. The platform itself is never sold or licensed and never handed over. What a client buys is a private build of their own, designed for their work, with the same discipline about records and approvals that his own firms have to live with when a finding gets argued with.
//Why him
He runs the cases the agents would be working on.
Investigating since 2009. Founded his first firm in 2010 and built it into Southern Recon Agency in 2014. He is an AI investigation expert and cybercrime specialist, a Certified Digital Forensics Examiner and one of fewer than fifty Florida Board Certified Investigators, and he built a proprietary AI investigation platform his own firms run on. The agents he builds have to survive somebody arguing with their output, which is a different design brief from a demo.
He also teaches it. He taught at World Orlando, the State Department sponsored conference where he trained a delegation of Brazilian prosecutors in Orlando, and on the training circuit he delivered "A-I for the Modern Private Investigator" at Intellenet's 2026 conference on April 15, 2026. If the need is to bring your own people up rather than to have something built, that is prompting training.
The firm owner's wider page is AI for investigation and security firms, the security operations version is AI for security operations, and the standing seat that usually decides what to build is the monthly advisory seat.
Tell me the workflow you would hand to an agent tomorrow.
The intake, the record pulls, the first draft of the report, the monitoring nobody has time for. You get a straight answer and a scope in writing.