What Governance Controls Will DoD Require for Agentic AI Targeting Tools in 2026?

According to DoD Directive 3000.09, the core question is not whether the system can generate recommendations; the real issue is whether a human can still make a lawful, informed, and reviewable decision about force. Agentic AI targeting tools sit in the highest-risk category because they can rank, filter, suggest, or re-prioritize targets faster than a human staff process, but speed does not remove accountability. Under OMB M-25-21 and M-25-22, agencies must document governance, acquisition risk, and public-trust safeguards before relying on AI in mission workflows, and those expectations will spill into DoD buys through acquisition language and test artifacts. According to GSA acquisition practice, any vendor delivering AI-enabled decision support should expect a contract file that preserves model versioning, data lineage, and acceptance evidence. The SBA matters because small businesses supporting these programs still need the same compliance package as the prime. If the tool touches CUI or is hosted in the cloud, FedRAMP and DoD CMMC controls are part of the gate, not an afterthought.

According to DoD Directive 3000.09, autonomy in weapon systems is governed as a safety and command-authority issue, not just a software issue. That matters because agentic AI targeting tools can compress the find-fix-track-target cycle into seconds, which creates pressure to automate judgment before the command chain has validated the output. The CDAO Responsible AI toolkit and the 2024 GenAI toolkit both push the same answer: the system must remain bounded by defined use cases, human review, and reproducible evidence. In practice, DoD will ask whether the tool merely recommends or whether it meaningfully shapes a target list, a prioritization queue, or a strike option package. If it shapes the operational choice, then the governance burden rises sharply. The DODIG 2025-039 report also reinforced that AI governance and acquisition processes need stronger controls, clearer roles, and better documentation before fielding. For contractors, that means the real deliverable is not only the model; it is the evidence chain that proves who approved it, what data trained it, what changed, and when the operator overrode it.

Under OMB M-25-21, agencies are expected to pair AI adoption with public-trust governance, which in DoD terms means a clear approval authority, a risk assessment, and an auditable record of testing before deployment. According to the CDAO toolkit, the fastest way to fail review is to show up with a powerful model and no boundary conditions. DoD will want to know the exact mission purpose, the confidence thresholds, the escalation path for uncertain outputs, and the kill switch for when the tool drifts outside its approved role. That is especially true for targeting tools because errors can propagate into rules-of-engagement questions, collateral-risk analysis, and time-sensitive engagement decisions. Contractors should expect acquisition teams to ask for model cards, system cards, red-team results, prompt and output retention, and operator training records. According to GSA-style file discipline, every material software revision should be traceable to a contract modification, test event, or formal approval. If the record cannot be reconstructed, the capability is not ready for operational trust.

According to DoD’s RAI toolkit, the first control is boundary control. Contractors should define exactly what the tool may do: recommend, rank, summarize, or flag targets, and what it may not do, such as auto-approve a strike option or suppress dissenting human review. Per FAR Part 7 acquisition planning and FAR Part 39 systems acquisition discipline, that scope should appear in the statement of work, test plan, and acceptance criteria. The approval chain must name the mission owner, the technical authority, the legal reviewer, and the operational user who has final say. For targeting-support tools, DoD will likely require a documented human-judgment checkpoint at every material step, not just at final execution. The platform also needs a change-control rule that freezes the model version during test and fielding so operators are not comparing one build to another. If the tool is provided by a small business, the SBA does not reduce the evidence burden; it only changes the teaming structure. The audit standard stays the same.

According to the GenAI Toolkit published by CDAO, contractors should expect a control stack built around traceability, monitoring, and rapid rollback. That means immutable logs for prompts, inputs, outputs, overrides, and operator acknowledgments; data provenance for every source set; and a documented method for reproducing each recommendation. DoD will also want evidence of adversarial testing, bias testing, and failure-mode analysis before any operational trust decision. Under OMB M-25-22, acquisition teams are pushed to buy AI efficiently, but efficiency does not replace due diligence: the evaluation must still prove that the model is safe, useful, and fit for mission. If the system is hosted in a cloud environment, FedRAMP authorization becomes part of the due-diligence stack, and if the data includes CUI, DoD’s CMMC controls become the baseline for the contractor’s environment. A vendor that cannot show continuous monitoring, incident reporting, and configuration management should assume it will not pass a DoD gate review.

According to GSA acquisition discipline and DoD’s RAI guidance, the best first move is to create a compliance matrix that maps every mission function to a control, an owner, a test, and a record. That matrix should include a human-approval step, a rollback trigger, a monitoring interval, and a retention rule for logs and artifacts. Contractors often lose time because engineering teams assume the government only wants performance metrics, while acquisition teams need proof that the software was fielded safely. The SBA angle matters for small firms: if you are an 8(a), HUBZone, SDVOSB, or WOSB prime or subcontractor, your small-business status does not soften the evidentiary standard for AI systems that influence targeting or force decisions. If anything, it raises the importance of clean file hygiene because small vendors are often evaluated on trust and responsiveness. According to OMB M-25-21, agencies should align governance, innovation, and public trust; for contractors, that translates into a single package that combines technical validation, cyber evidence, and decision authority in one place.

Per FAR 4.8 records administration and FAR Part 39 acquisition of information technology, the government will want a file that survives personnel turnover, software updates, and mission tempo changes. That means the contractor should store the exact prompt templates, the approved data sets, the evaluation thresholds, and the version history used in each demonstration or operational trial. According to DoD CDAO, the point is not to generate more paperwork; the point is to make the system defensible when a commander, reviewer, or investigator asks why the tool recommended one target over another. A mature vendor will also show how the tool handles uncertainty, how often it escalates to a human, and what happens when sensor data is incomplete or conflicting. If the company cannot answer those questions in one sitting, it is not ready for DoD’s trust threshold. The winning vendors in 2026 will be the ones that can show operational usefulness and governance maturity in the same package.