DLA pilots autonomous AI agents for logistics operations
CIO Adarryl Roberts calls them "digital employees" and says cultural trust is the prerequisite, but DLA hasn't said how it'll assure security when unattended agents run across sensitive data domains.
TL;DR
The Defense Logistics Agency is moving beyond robotic process automation (which saved an estimated 300,000 hours in 2025) into agentic AI that operates unattended around the clock. CIO Adarryl Roberts told reporters at DLA's Industry Collider Day that roughly 90-95% of DLA's 185-190 current bots run unattended. The agency's AI Center of Excellence, founded in 2024, will govern the shift. Roberts framed workforce trust as the gating condition, pointing to "GenAI 101 training" and persona-based access controls that restrict agents to the data domains matching their role. He didn't address how security assurance works when agents operate autonomously across CUI and sensitive-but-unclassified environments over months-long runs.
The progression inside DLA is deliberate: from attended robotic process automation to unattended bots, and now toward agents that operate without a human watching. Roberts put the current state plainly: "We really cracked the nut to have those unattended." The agency's 190-ish bots run ~90-95% unattended, handling procurement, cataloging, distribution, and disposal workflows across the combat support agency's global footprint.
That isn't small. DLA manages the military's supply chain, food, fuel, medical supplies, clothing, weapon-system parts. When a procurement agent runs unattended at 3 a.m. moving financial and inventory data, the operational surface is real.
Trust as a proxy, not a control
Roberts's framing of the problem centers on culture. "There has to be a level (a cultural level) of trust in those outcomes," he said, arguing that GenAI 101 training and exposure to tools like Gemini and ChatGPT build workforce comfort before agentic systems go live. Trust is necessary, but it isn't a security control. A workforce that trusts agent outputs isn't a workforce protected against an agent making bad decisions at scale, or an agent whose access tokens get compromised and reused.
Persona-based access and the data-quality problem
DLA is applying a personnel-style governance model to agents: persona-based access that mirrors human role permissions. A procurement agent gets procurement and financial data domains, same as its human counterpart. Roberts also flagged the data-cleanup effort underway, "removing the duplicity, errors, etc., so the data is in a usable state."
The logic is sound but assumes clean boundaries between data domains. In practice, logistics data touches personnel data, financial data, and operational readiness data. A procurement agent working autonomously across six months will cross those boundaries routinely. DLA hasn't described how it detects or contains that drift.
What isn't in the briefing
Roberts acknowledged the "complex universe" of managing digital employees, performance evaluation, possible termination, supervision models where every human becomes a supervisor. He didn't address the security assurance model for autonomous agents operating on CUI and sensitive-but-unclassified data. DLA has implemented a zero-trust framework, which Roberts called "foundational," but zero trust operates on session-level enforcement. An authenticated agent session running continuously for weeks or months doesn't re-authenticate; it just keeps executing.
The AI Center of Excellence, stood up in 2024, carries the governance burden here. What it's actually inspecting, how often, and against what standard, those aren't public yet.
Published ·Deep Fathom