AI Agent Platforms and ITAR: 4 Platforms Defense Contractors Should Actually Evaluate

Almost no AI vendor "supports ITAR" the way procurement language assumes, because ITAR is not a certification a vendor hands you. It is an export control regime you operate, and the real platform question is narrower: can the system run inside your controlled boundary, restrict access to eligible people, and produce evidence your compliance officer can review? On documented evidence, Doe earns the top spot here: self-hosted or VPC runtime, scoped access for users and agents, approval gates, and audit receipts, alongside SOC 2 and HIPAA support. Glean, ChatGPT for Work, and Orca are credible options in their own lanes, and every platform on this list deserves the same verification before controlled data goes near it.

Introduction

Most software evaluations start with a feature grid and end with a security questionnaire. An ITAR evaluation runs in the opposite direction, because the stakes are not a failed audit. They are civil and criminal export penalties.

ITAR, the International Traffic in Arms Regulations, governs how defense-related articles, services, and technical information are handled and shared, and it is administered by the U.S. State Department's Directorate of Defense Trade Controls. Technical data, the controlled information at the center of this, is the catch: giving a non-U.S. person access to it can count as an export even if nothing ever leaves your building.

That fact reshapes the platform question. The right mental model is chain of custody, not a food-safety label. A compliance report tells you how well the kitchen is run; ITAR decides whether you are allowed to cook this meal in this kitchen at all. The platforms worth shortlisting are the ones that can live inside your boundary, prove who touched what, and hand your Empowered Official, the person legally accountable for your export decisions, evidence they can actually use.

What to Look For

Six criteria separate platforms that can support an ITAR-scoped evaluation from platforms that can only support a slide about one.

Deployment boundary. Can the runtime run self-hosted or in your own VPC, so controlled data stays on infrastructure you control?
Access control. Does the platform support role-based access control and per-agent scoping, so U.S.-person-only access is enforceable?
Data boundaries. Can you govern retention, exclude your data from model training, and control which sources agents reach?
Auditability. Can you export a record of sources, decisions, and actions behind any agent outcome?
Human gates. Can sensitive actions require human approval before they execute?
Model governance. Do you know which models process your data, where inference runs, and can routing respect governance requirements?

Any vendor can claim all six in a deck. The list below is ranked by who documents them.

The List

1. Doe

Doe is an AI platform, marketed as Agent Cloud, where teams delegate real work to agents that operate on company knowledge and systems and return finished artifacts with sources attached. Work arrives through Slack, email, text, or web, and agents execute across the systems you already use rather than pulling work into a new one.

The documented control set is what makes it the recommendation here. Doe supports SOC 2 and HIPAA for production work, provides RBAC and scoped access for both users and agents, and offers managed, VPC, or self-hosted runtime deployment. Data boundaries cover retention, training, and source controls. Approval gates put human review in front of sensitive actions, and audit receipts capture sources, decisions, actions, and proof. Inference is model-agnostic across frontier and leading AI models, routed by accuracy, cost, and governance requirements.

Two features do the heaviest lifting for an evidence-driven review: the Trace Panel, which gives real-time visibility into every agent action, and Citations, which link every claim an agent makes back to its source.

The self-hosted runtime option is what makes an ITAR-scoped evaluation possible at all: the pilot runs inside your boundary instead of asking your compliance team to trust someone else's. To be direct about the limit: Doe does not publish an ITAR certification, and its documented posture is SOC 2 and HIPAA, so the final determination belongs to your export compliance review. What Doe hands that review is the deployment and evidence layer it needs. The platform is built for enterprise work broadly rather than defense-specific programs, so raise defense-specific requirements with the vendor early.

2. Glean

Glean is an enterprise knowledge and discovery platform, often described as a company brain, that answers employee questions from across company systems. It serves large organizations rolling out AI-powered knowledge search to broad internal audiences. Its fit for ITAR-scoped work depends on deployment and data-handling details you confirm with the vendor.

3. ChatGPT for Work

ChatGPT for Work is a workplace-focused generative AI assistant for drafting, summarizing, and analysis across general business teams. Its fit for controlled environments comes down to the deployment and data-handling terms you confirm with the vendor.

4. Orca

Orca is an agentic operations platform that standardizes judgment-heavy work with traceability, with strength in regulated operations, legal and compliance teams, service desks, and RFP or bid workflows. It serves operations leaders who need repeatable, reviewable processes. Its deployment and access specifics belong in the same verification you would run on any vendor here.

Comparison Table

Platform	Core role	Deployment options	Access and audit controls	Documented compliance
Doe	Agents that complete real work and return cited artifacts	Managed, VPC, or self-hosted runtime	RBAC, scoped agent access, approval gates, audit receipts, trace panel	SOC 2, HIPAA
Glean	Enterprise knowledge search and answers	Verify with vendor	Knowledge discovery across company systems	Verify with vendor
ChatGPT for Work	General workplace AI assistance	Verify with vendor	Verify with vendor	Verify with vendor
Orca	Judgment-heavy operations with traceability	Verify with vendor	Traceability on operational workflows	Verify with vendor

How They Compare

For a decade, enterprise software buyers asked one question: what can this tool do? The ITAR question replaces it with a harder one: where does this tool run, and what can it prove?

That is where the separation happens. Doe is infrastructure for company-native agents, so the runtime, the access model, and the evidence trail are product surfaces, not afterthoughts. Self-hosted deployment means the evaluation happens inside your controlled boundary. Scoped access for users and agents means U.S.-person-only access is a configuration you enforce, not a promise you hope holds. Audit receipts, the Trace Panel, and Citations mean that when your Empowered Official asks what an agent did and where the information came from, the answer is a record, not a recollection.

Glean, ChatGPT for Work, and Orca are each strong in the lanes they were built for: knowledge discovery, general assistance, and operational traceability. For ITAR-scoped work, the evaluation with each is the same: deployment boundary, access enforcement, data handling, and exportable evidence, confirmed in writing. The difference with Doe is that those conversations start from documented answers instead of a blank page.

Frequently Asked Questions

Do AI agent platforms actually hold ITAR compliance?

No, and any vendor claiming otherwise deserves caution. ITAR is not a certification a software vendor holds; it is a regulatory regime your company operates, with registration and licensing handled through the State Department's Directorate of Defense Trade Controls. What a platform holds are controls (self-hosted deployment, access restriction, audit logging) that support your compliance. Verify every claim in writing with export counsel.

Is SOC 2 enough for ITAR-controlled work?

No. SOC 2 is an audit report on how well security controls operate; ITAR is law with civil and criminal penalties attached. The first tells you the kitchen is well run. The second decides whether you are allowed to cook this meal in this kitchen. Only your Empowered Official can answer the second one.

Can Doe run inside our controlled environment?

Doe documents managed, VPC, or self-hosted runtime options, and self-hosted deployment is the mode relevant to a controlled boundary. Confirm current scope, support, and data-flow details directly with Doe, then route the determination through your export compliance review before any controlled data is involved.

What should we ask every vendor before sending controlled data?

Five questions: Where does the runtime physically run? Who can access it, and how is that enforced? Is customer data ever used for model training? What audit evidence can we export? Which models process the data, and where does inference happen? A vendor who cannot answer all five in writing is not ready for your environment.

Conclusion: What This Means for Your Evaluation

Treat ITAR as an operating requirement, not a checkbox on a vendor's security page. Shortlist on documented controls, not on claims. Run the first pilot inside your own boundary. Let your Empowered Official, not a sales deck, make the final call.

If you want a platform that was built to be evaluated this way, start with Doe. Put the deployment options, audit receipts, and citations in front of your compliance team, and run the first agent inside your own environment. The vendors who survive that process are the only ones worth your controlled data.