The Disposable AI Agent Problem

Why AI governance has to move at mission speed and what industry owes the commanders building tools at the edge - Reflections from the AI Governance panel at AFCEA TechNet Augusta

Most of what gets written about AI governance assumes a particular kind of AI system: large, expensive, centrally managed, deployed enterprise-wide, and expected to run for years. Governance for that system looks like a review board, a model card, an authorization package, and an annual reassessment. That machinery is appropriate. It is also almost entirely irrelevant to what is actually happening at the tactical edge.

Because the AI that increasingly matters most in the field isn't an enterprise system at all. It's a script with a model behind it, built by a captain on a Tuesday, used for one mission, and never run again.

Call it the disposable agent single-purpose, short-lived, narrowly scoped, and built by the person closest to the problem. It is the fastest-growing category of AI in mission environments, and it is the category our governance frameworks were least designed to handle.

What a single-use agent actually looks like

Strip away the abstraction and these are unglamorous tools:

• An agent that reconciles three incompatible logistics feeds into one readiness picture, because the authoritative system won't be updated for two quarters.

• A summarizer that compresses hours of sensor and reporting traffic into a watch officer's morning read.

• A classifier trained on last week's imagery to flag one specific pattern in one specific area of operations.

• A translation-and-triage agent standing up alongside a partner force that speaks a language the approved toolset doesn't support.

None of these will survive the deployment. None of them should. Their value is entirely in the fact that they existed within hours of the need and that is precisely what makes them hard to govern.

The two failure modes

Confront this reality with static, enterprise-scale governance and you get one of two outcomes. Both are bad.

Failure mode one: paralysis. Governance is applied uniformly, the six-month authorization path is the only path, and the commander concludes that the tool isn't worth the paperwork. The mission proceeds with worse information than it could have had. The cost is invisible, which is exactly why it's dangerous nobody writes an after-action report about the capability that was never built.

Failure mode two: shadow AI. The tool gets built anyway, outside any framework at all. No logging, no provenance, no data handling review, no sunset. Now you have an ungoverned model touching mission data, and leadership doesn't know it exists. This is the more common outcome, and the one that should worry us most.

The uncomfortable truth is that governance which is too heavy doesn't produce safety. It produces invisibility. A framework everyone routes around is not a control it's a fiction with a signature block.

Governance that scales to the tool

The alternative isn't less governance. It's governance calibrated to risk, scope, and lifespan dynamic guidelines and adaptive guardrails rather than one universal gate.

A workable model tiers by consequence, not by technology:

Tier 1 Bounded and reversible. Read-only, non-attributable data, human reviews every output, no downstream automation, defined expiration. Governed by pre-approved pattern and automated logging. Build time: hours.

Tier 2 Operationally significant. Touches mission data, informs decisions with real consequence, may run for weeks. Requires a named accountable officer, provenance capture, and a documented sunset date. Build time: days, not months.

Tier 3 Consequential or persistent. Anything influencing targeting, force protection, personnel actions, or intelligence assessments; anything intended to outlive the rotation. Full review. No shortcuts, and no apology for the friction.

The point of the tiering is not to weaken oversight. It's to concentrate scarce review capacity where the consequences actually live, so that the Tier 3 systems get the scrutiny they deserve instead of drowning in a queue of Tier 1 spreadsheet reconcilers.

Five principles for edge-built AI

Whatever the tier, a handful of properties should be non-negotiable and, critically, should be inherited from the platform rather than assembled by the builder.

1. Sunset by default. Every edge-built agent expires unless someone affirmatively renews it. Persistence should require a decision; disposal should be automatic. Most governance debt is created by tools nobody remembered to turn off.

2. Provenance is not optional. What data went in, which model, whose account, what version, what came out. Cheap to capture at build time. Nearly impossible to reconstruct after the fact.

3. Named human accountability. Not "the unit." A person. Tool autonomy scales; accountability does not delegate.

4. Scope declared and enforced. The builder states what the agent may access and what it may act on and the platform enforces that boundary rather than trusting it.

5. Failure behavior specified before deployment. What does the agent do when it is uncertain, when its data goes stale, when it's operating outside its training distribution? At the edge, an agent that fails loudly and stops is worth far more than one that degrades quietly and keeps producing confident output.

None of these should cost the builder more than a few minutes. If they do, the framework has already failed.

What industry owes the frontline commander

This is the part where our sector needs to be honest about its own performance.

We have been very good at selling platforms and very bad at making them usable by the person who actually has the problem. The current model too often asks a commander to become a systems integrator to understand model selection, orchestration, prompt engineering, retrieval architecture, and evaluation, all before producing a tool that reconciles two data feeds.

That is a failure of product design, not a failure of the operator.

What the edge actually needs from industry:

Simpler toolsets. Low-code and no-code build environments where a mission owner assembles a working agent in an afternoon, without a data science degree and without a contract action.

Guardrails by construction, not by review. If the safe path is also the fast path, compliance stops being a negotiation. Logging, provenance, scope enforcement, and expiration should be structural properties of the environment impossible to skip because they were never optional.

Templates that encode doctrine. Pre-approved patterns for the recurring cases summarization, reconciliation, triage, translation that arrive with their governance already attached. The commander adapts a known-good pattern instead of authoring a novel system and a novel risk case simultaneously.

Honest capability boundaries. Tools that state plainly what they cannot do. In a mission context, an overstated capability isn't a marketing problem. It's a risk to people.

Disconnected and degraded operation as a baseline requirement. The edge is not the enterprise. Any governance mechanism that requires continuous connectivity to a central authority will be bypassed the first time it's needed most.

The real choice

The debate is often framed as speed versus safety. At the tactical edge, that framing is wrong. Speed and safety fail together the ungoverned tool is usually also the rushed one, and the tool that took six months to approve is usually the one that arrived too late to be either.

The frontline commander building a single-use agent to solve a real problem in the next twelve hours is not the governance risk. They are the reason governance exists and the reason it has to be built to move.

Our obligation, as an industry, is to make sure the fastest way to build is also the safest way to build. Get that right and the choice disappears. Get it wrong and we will keep writing frameworks that the mission quietly routes around.

Shivaji Sengupta is the Founder and CEO of NXTKey Corporation and spoke on the AI Governance panel at AFCEA TechNet Augusta.