Your veteran-facing AI is deployed.
Can you prove — independently — that it works?
The mandate already asks for this. Post-deployment monitoring and independent review aren't our pitch — they're the buyer's requirement. We're the outside party structurally able to provide the read.
The requirement is already on the books
M-25-21
OMB requires post-deployment monitoring and independent review — explicitly including evaluating the effectiveness of vendor offerings.
GAO gap
GAO has flagged continuous, post-deployment monitoring as a recurring gap between policy and practice.
~64%
of federal high-impact AI use cases sit at the VA — the largest concentration of citizen-facing AI in government.
This makes the need official, not our opinion. The question isn't whether deployed AI gets reviewed — it's who is structurally allowed to review it.
An outside read on whether the AI lowered true cost-to-resolve
Resolved on first contact
~71%
would be the rate that actually held — not the rate marked "closed."
True cost-to-resolve
$41 vs $18
reported vs. what reopened cases would push it to.
Where the cost would leak
Handoff
AI→human escalations that wouldn't carry context forward.
On a read like this, the verdict would separate cases the AI truly closed from cases it only marked closed — and would attribute the leak to a specific party (model, data, integration, or human queue), not to "the AI" in general.
Figures are illustrative and conditional. On your engagement they're produced from your de-identified data and labeled accordingly.
Why we can grade it when the integrator can't
The company that builds or operates the AI cannot be the company that certifies it worked. That's not a slogan — it's the structural test the EU AI Act (Art. 31) and decades of audit doctrine already apply to conformity assessment.
Buyer-paid, not vendor-paid
The agency engages us, not the AI vendor. Our incentive is an accurate read, not a renewal.
No build, no operate
We don't sell the model, the platform, or the integration. There's nothing for us to defend by inflating the result.
Reproducible verdict
Same inputs, same method, same answer — one your team or a third party can re-run, not take on faith.
Most incumbents serving federal AI assurance are operator-run or operator-paid — they advise on, build, or sit beside the systems they would grade. We answer that structurally, at the capability level. We don't publish a competitor roster on a public federal page.
I built this measuring AI economics as a senior analyst, and I care for a disabled veteran — my father — so a veteran-facing system that marks a problem solved when it isn't isn't abstract to me. The point of an independent read isn't to embarrass a program. It's to give the people accountable for it a number they can stand behind in front of leadership.
— Brandon Burdin, Founder, MarginSignal OS
Currently in dialogue with federal AI leadership on independent, post-deployment impact measurement.
Scope one use case
A 30–60 day independent Impact read on a single deployed use case, run on de-identified data. You get an outside verdict on whether it's working — and where it isn't.
Request the one-pager or scope a working session. The federal lane has its own track on our design-partner page — not a marketing form.
Core methods patent-pending. Public data and de-identified inputs only.