Artificial intelligence (AI) is opening new opportunities for federal agencies to automate workflows, improve mission operations, and respond to threats at unprecedented speed and scale. Turning promising AI pilots into production capabilities, however, requires agencies to address security, identity, governance, and operational realities from the start.
In a preview to MeriTalk’s Aug. 27 Mission Room forum, “Agentic AI and the New Federal Workflow,” MeriTalk sat down with John Sourk, director of federal sales at Abnormal AI, to discuss what makes AI production-ready, where agentic AI can deliver the greatest impact, and how agencies can govern AI agents as rigorously as human employees. The conversation also explores how purpose-built technologies and continuously learning systems can help federal teams modernize without creating unnecessary technical debt, application sprawl, or operational risk.
MeriTalk: As agencies move AI-enabled workflows from pilot to production, how should they weigh functionality, security, identity, governance, and operational readiness – and what does “production-ready” mean in practice?
Sourk: Production-ready has to mean more than “it works in a lab.” Federal production environments are nothing like the sandbox or lab environments where AI use cases are often piloted. You’re layering a new capability on top of years, sometimes decades, of established processes, legacy systems, and associated technical debt.
The real test is whether it works in that reality. Does it respect the access and identity controls already in place? Does it fit into workflows teams have built and run for years without disrupting them? Can it be governed the way you would govern a new hire, including adjusting controls if its behavior changes?
That’s where limited-use AI has an advantage over broad, general-purpose systems. Capabilities built to solve a specific problem are easier to test against the realities of a governed environment and easier to trust.
MeriTalk: How can agencies apply agentic AI to improve service delivery, case routing, compliance, and mission operations, and where are federal teams seeing the strongest potential?
Sourk: There’s real potential across all of those use cases. In service delivery and case routing, for example, agencies can use agentic AI to operate more efficiently and accurately internally. That can deliver better outcomes to constituents much faster and, ultimately, at a lower cost to the taxpayer.
But I’d argue security operations may have the greatest potential. Agencies aren’t just figuring out how to use AI internally; they’re racing against adversaries that are already using it. Threat actors are applying AI across the kill chain at a speed and scale that federal teams using traditional tools simply can’t match.
AI-scale defense isn’t a nice-to-have. It’s how agencies can begin closing a gap that adversaries have already opened. In security, I think the risk of moving too slowly is greater than the risk of moving too fast.
MeriTalk: As agencies deploy their own AI agents, how can they apply the same rigor used to govern human employees while using behavioral AI to build trust, accountability, and policy enforcement directly into the workflow?
Sourk: Agencies should look at every new AI agent as a new employee with system access. When somebody joins an organization, you have a basic understanding of the systems they’ll need, but you start smaller and use an approval process as they need access to more systems. AI agents should be treated the same way.
Behavioral AI makes that approach practical. You can continuously baseline how an identity – whether human or machine – normally operates, understand what it’s supposed to do, and automatically flag deviations instead of relying on someone to review audit logs after the fact.
Trust and policy enforcement should be built into how these systems behave, not treated as a compliance exercise that gets bolted on afterward. Agencies that get this right will govern AI agents with the same rigor they govern people: clear roles, clear boundaries, and a way to immediately detect when something accesses the wrong system or drifts from expected behavior.
MeriTalk: What architectural and development choices can help agencies expand AI-enabled workflows while limiting risks, such as technical debt, vendor lock-in, and application sprawl?
Sourk: This relates to a question the government has been asking for years around commercial off-the-shelf versus government off-the-shelf solutions. For technologies touching identity or security, I think the strongest approach is a vetted, authorized commercial platform built to solve one problem well, rather than a custom government solution or a general-purpose tool configured into numerous different workflows.
Purpose-built AI doesn’t create sprawl in the same way a broad, low-code platform potentially can, because it isn’t inviting every team to build its own version of a capability. Every homegrown integration can become future technical debt that the agency owns indefinitely. People leave, teams change, and there may be no vendor roadmap to keep that custom integration current.
FedRAMP also gives agencies an important vetting mechanism. Pairing a small number of purpose-built, authorized commercial AI platforms with clear boundaries can help agencies reduce technical debt and vendor risk rather than adding to them.
MeriTalk: As policy and mission needs evolve, how can federal teams build security into AI-enabled capabilities so trust and compliance keep pace automatically?
Sourk: The capabilities that will keep pace best are built on models that continuously learn instead of static rule sets that need to be manually reconfigured when a policy changes or the threat landscape shifts.
The architecture question agencies should ask is: Does this system maintain trust and compliance continuously, or will the team need to recertify and retune it whenever something changes? The first approach scales with the mission. The second becomes very difficult to maintain.
This is an area where the government needs to be more forward-leaning. Systems that improve and adapt without requiring human intervention at every step can be difficult to trust initially. But many leading private-sector organizations have adopted these technologies and accepted some of the risk that comes with moving early.
Federal missions are every bit as critical as those in the private sector, and agencies deserve the same caliber of technology. That means accepting an appropriate level of risk while looking toward platforms that can continuously learn, improve, and adapt as the mission changes.