The U.S. Department of Energy (DOE) 2022 UAS presentation identifies five risk-management stages: supply chain, data acquisition, model development, deployment and performance. It is a useful source for supplier questions, not a compulsory drone certification. NIST AI RMF 1.0 adds four risk-management functions—Govern, Map, Measure and Manage—which are distinct from DOE's five stages. Buyers should seek evidence for the exact aircraft, software version, mission and operational conditions rather than treating AI marketing, a laboratory benchmark, a contract or a flight authorisation as interchangeable proof.
What the DOE source actually establishes
DOE's Artificial Intelligence and Technology Office titled its 2022 presentation Trustworthy AI & UAS Technology. Slides numbered 2 and 3 define the five-stage lifecycle and propose questions about chip and firmware origins, training-data custody, model development, deployment protection and output monitoring. Slides 6–7 discuss edge-device capture and adversarial-model threats. These are risk considerations, not proof that a specific unmanned aircraft has undergone certification, attack testing or independent acceptance.
The DOE slides use a historical Level 0–5 autonomy illustration on slide 5, citing external drone-industry sources. It is not a universal, legally recognized DOE autonomy-rating system. The slide's historical indications about what was widely available in 2013–2020 must not be presented as audited 2026 deployment or market-share data. Inspect original DOE slides 2, 3, 5, 6 and 7.
Five DOE lifecycle stages translated into supplier-assurance evidence
Method boundary: The stage names are DOE's; the supplier requests and evaluation rules below are UXV Searcher's procurement-oriented interpretation, not a government-mandated assessment checklist or a claimed industry standard.
| DOE stage (2022) | Supplier evidence to request (UXV Searcher interpretation) | What is still NOT established by a document alone |
|---|---|---|
| 1. AI supply chain | Aircraft and computing variant; processor and sensor makers; firmware and software component inventory; AI model supplier; third-party integrations; signed build/update history; known vulnerability and patch handling. | A vendor logo or generic bill of materials does not prove the delivered serial-number configuration, source integrity or exploit resistance. |
| 2. Data acquisition | Dataset origin, collection and labelling responsibilities; rights and access; chain of custody; versioning; contamination checks; representation of weather, light, terrain and targets; training/test separation. | A large image count is not proof of representative training, lawful use of data or unbiased performance in the target mission. |
| 3. Model development | Model and training-run IDs; independent test split and raw denominators; mission-specific precision/recall, false alarms and misses; robustness and adversarial assessments; model limitations and change log. | A single benchmark or vendor-provided accuracy percentage does not establish in-flight safety, repeatability or cross-environment generalisation. |
| 4. Model deployment | Whether inference is onboard, ground-based or cloud-hosted; version and configuration traceability; mission latency, thermal and power limits; communications-loss and GNSS-loss behaviour; human intervention, update and rollback controls; integration flight tests. | Integration, permission to operate and an independent AI assurance assessment are different evidence types. |
| 5. Model performance | Field telemetry with scenario denominators; model drift and incident logs; false-positive/false-negative rates under the intended conditions; security response, revalidation triggers, operator feedback and corrective-action records. | A successful demonstration or a contract award does not prove sustained field reliability or production-wide conformity. |
For a credible assessment, tie every supplied document to the named OEM, variant, firmware, model build, operational design domain (ODD), operator and test date. If a source covers only a family, laboratory simulator or partner integration, mark applicability as NOT ESTABLISHED until supported. DOE does not specify procurement scoring thresholds. Source: DOE 2022, slides 2–3.
DOE's five stages versus NIST's four AI RMF functions
NIST AI 100-1, issued 26 January 2023, is a voluntary, non-sector-specific risk framework. Its functions are GOVERN (accountability and policies), MAP (context and potential effects), MEASURE (test, evaluation, verification and validation) and MANAGE (responding to and monitoring risks). They run across the AI lifecycle and must not be recast as a different five-stage DOE pipeline. NIST's live framework page describes revision work in progress; do not assert a completed new RMF version without its actual publication.
In particular, NIST AI RMF 1.0 MEASURE 2.1–2.7 emphasizes documented test methods, conditions close to deployment, known limitations, safe failure and resilience. The NIST MEASURE Playbook suggests testing false positives, false negatives, prediction latency, drift and security responses. The Playbook is voluntary guidance, not a universal pass/fail checklist; it cannot substitute for aviation approval or an engineering test report.
Edge AI, cybersecurity and provenance: verify the configuration
DOE slide 6 warns that a captured or compromised edge device may expose models and datasets; slide 7 describes adversarial manipulation. These are plausible threat classes, not documented breaches of a named drone manufacturer. Ask suppliers to demonstrate device identity, access controls, security-update signing, remote-model disablement where feasible, and test results for the relevant operational threat model. A claim that onboard AI is automatically secure, or that a GPS-denied feature proves resilience to all interference, is unsupported.
CISA and G7 partners' 12 May 2026 AI software-bill-of-materials guidance proposes minimum elements for improved component transparency. The guidance is non-mandatory and a software/AI inventory is not an aircraft structural bill of materials or proof of component provenance in an as-built unit. Vendor assurances should be matched to delivered software versions and support agreements, not transferred indiscriminately from a company or product family.
Regulatory boundaries: assurance is not flight approval
United States: The FAA's Part 107 waiver guidance describes case-specific operational permissions for flights that would otherwise exceed Part 107 limits—for example, certain beyond-visual-line-of-sight or multiple-aircraft operations. A waiver is not a DOE or NIST endorsement of a supplier's AI model. Applicability depends on the actual operation and aircraft.
European Union: EASA's June 2026 UAS rules compilation discusses operational authorisations and technical risk evidence. EASA also published proposed Issue 03 of its AI Concept Paper on 3 June 2026 for consultation, extending technical discussion to Level 3 AI. A proposed concept paper is not an issued aircraft certificate, and EASA's AI-level terminology must not be mapped automatically onto DOE's illustrative Level 0–5 chart.
Other requirements: Applicable safety, airspace, privacy, procurement, cybersecurity, export-control and defence obligations depend on jurisdiction, end use and the product's exact configuration. This analysis does not establish legal compliance for a named supplier, and military procurement requirements cannot be inferred from the civil guidance alone.
How UXV Searcher would grade supplier evidence without overstating maturity
| Evidence dimension | What to record | Decision rule |
|---|---|---|
| Technical validation | Vendor assertion → dated demonstration → replicable test report → independent evaluation → mission-representative field observation. | Preserve test conditions, reviewer independence and denominators; do not infer universal performance from one environment. |
| Commercial maturity | Research, prototype, named contract, delivered unit, accepted unit and operating fleet (if documented). | Contracts and deployments require their own dated source. A purchase is not a performance certification. |
| Configuration scope | Platform family, specific variant, software/model version, sensor set, operator and geography. | Family-level evidence is never silently applied to another variant. |
| Assurance coverage | Verified, partially documented, vendor-claimed or NOT ESTABLISHED for each DOE stage and NIST function. | Missing evidence remains a gap—not a zero-risk score and not an invented pass. |
A practical buyer request is a configuration-specific evidence pack: (i) subsystem and AI-software inventory, (ii) data/model lineage, (iii) independent and vendor test summaries with raw case denominators, (iv) flight and failure-mode validation, and (v) incident, update and contractual support processes. Buyers can use NIST's testing questions to examine an assurance pack while separately confirming regulatory permissions. UXV Searcher has not independently tested or certified any supplier in this article; the proposed evidence rubric is an analytical method rather than a finding of vendor performance.
Market intelligence implications—and what cannot be inferred
Supplier assurance creates research questions about testing services, embedded security, third-party validation and software provenance. The sources do not quantify 2026 AI-drone revenues, annual unit demand, compliance rates, incident rates, acceptance success or the size of an assurance-services market. Any commercial forecast would need a separately sourced buyer population, spend/price assumptions, units, timing and double-count controls. An AI label alone is not an attributable autonomous-software revenue stream.
For adjacent supplier and technology research, see AI drone suppliers and integration roles, edge perception and GNSS-denied evidence and counter-UAS claims and procurement stages.
Questions readers ask
What are the five DOE AI lifecycle stages for drone assurance?
The DOE's 2022 UAS presentation lists AI supply chain, data acquisition, model development, model deployment and model performance. UXV Searcher adapts those headings into supplier due-diligence questions; the DOE does not prescribe a universal drone certification.
Is NIST AI RMF mandatory or a drone certification?
No. NIST AI RMF 1.0, issued on 26 January 2023, is a voluntary, cross-sector risk-management framework using Govern, Map, Measure and Manage. Regulatory approvals and flight permissions must be assessed separately.
Does a Level 4 or Level 5 autonomous drone have an official DOE approval?
No. The DOE presentation reproduces an illustrative Level 0–5 drone-autonomy taxonomy, not a DOE licensing or certification process. A supplier must identify the actual function, operational limits, human oversight and applicable aviation permissions.
What evidence should buyers request from AI drone suppliers?
Request configuration-specific hardware and software inventories, data and model provenance, representative test results with denominators, integration and failure-mode evidence, change controls and post-deployment incident monitoring. Classify independently verified performance separately from contracts, demonstrations and vendor claims.
Original sources and evidence dates
Primary government publications and official guidance below support the specific dated statements. Supplier-evaluation tables are UXV Searcher's interpretations, not language mandated by these organisations.
- U.S. DOE AITO — Trustworthy AI & UAS Technology · 2022 (June-hosted PDF); slides 2–3, 5–7. Exact day of release not established.
- NIST AI 100-1 — AI Risk Management Framework 1.0 · January 2023; voluntary framework and core functions.
- NIST official publication record — AI RMF 1.0 · 26 January 2023 publication date.
- NIST AI RMF official framework page · checked 8 November 2026; revision activity, not a certified new version.
- NIST AI RMF Playbook — MEASURE · voluntary suggested testing and monitoring actions.
- CISA / G7 — SBOM for AI minimum-elements guidance · 12 May 2026; non-mandatory supply-chain guidance.
- EASA — proposed Issue 03 AI Concept Paper announcement · 3 June 2026; consultation-stage guidance.
- EASA — Easy Access Rules for UAS · June 2026 revision; operational authorisation context.
- FAA — Part 107 waivers · official operational-waiver information; accessed 8 November 2026.