Veteran-Owned · Virginia LLC · US-domiciled inference
Beauty for ashes.

Your AI agents are making
promises nobody approved.

A chatbot invents a refund policy and a tribunal holds you to it. A voice agent quietly runs on a model you'd never clear for public work. Something goes wrong — and there's no record of why. We put a governance layer in front of your AI: every use case classified, gated, decided by a human, and documented.

24–48 hour brief  ·  no pitch deck  ·  no commitment

// GOVERNANCE LAYER

classify → gate → decide → log
Foreign-adversary model in the path ⛔ HARD STOP · Va. EO 46
Risk classification EA-225 tiers
Control mapping NIST AI RMF
Rights-affecting decision → HUMAN REVIEW
Every decision, recorded ● AUDIT TRAIL
// What happens with no guardrails

Deployed an AI agent. Didn't govern it.

The technology works. That's the problem — it acts. When an AI agent talks to your customers, your residents, or your staff with no governance layer in front of it, these are the failure modes that show up in the headlines and the lawsuits.

Liability

Promises you're held to

A support chatbot states a policy you never offered. Courts and tribunals have already ruled the organization is bound by what its bot said.

Injection

Hijacked by its own input

Hidden instructions buried in a user message, an email, or a document override the system prompt — and the agent leaks data or takes an action no one authorized.

Data

Sensitive data, gone

PII and privileged records flow into a third-party model that retains them. You can't recall it, and you may not even know it left.

Provenance

A model you'd never clear

The agent is quietly running on a foreign-adversary model. For public-sector and regulated work that's a hard stop — and most teams don't check.

Accountability

No record of why

When it goes wrong, there's no trail: no record of what the AI was allowed to do, who approved it, or why it decided what it did. You can't answer the auditor.

Scope

Quietly deciding on its own

An agent meant to draft starts deciding — eligibility, pricing, who gets a service — with no human in the loop and no one who chose to let it.

// The governance layer

Four steps in front of every AI use case.

We don't slow your AI down — we put a thin, inspectable layer in front of it. Each use case runs the same path. The engine does the sorting and the paperwork; the call stays with a person.

01

Classify

Sort the use case into a risk tier from its purpose, data, and who it touches — mapped to a recognized risk framework, not a gut feel.

02

Gate

Run it through policy checks: foreign-model hard stop, data-residency, human-handoff, retention. Anything unresolved becomes a question, not a silent pass.

03

Decide

Approve, approve-with-conditions, or hold for review. High-risk and rights-affecting use cases are routed up to a person — never settled at the tooling layer.

04

Document

Every run produces a decision card and an audit trail: the tier, the gates, the conditions, the authorities cited, and who signed off.

The engine classifies, routes, and records — your people decide. It surfaces the risk and writes the paper trail so a human can make an informed call and prove they made it. It does not make the decision for you.
// The rules it carries

Policy, turned into checks that run.

The gates aren't a slideshow of principles — they're concrete checks the engine applies to every use case, mapped to the authorities your reviewers and auditors already answer to.

HARD STOP

Foreign-AI prohibition

A foreign-adversary model anywhere in the path stops the use case cold. Aligned with Virginia EO 46, grounded in the federal 15 CFR 7.4 adversary-nation list — the check most teams skip.

MAPS TO

NIST AI RMF

Each classification and gate maps to the Govern / Map / Measure / Manage functions, so the output speaks the framework your stakeholders expect.

MAPS TO

State & sector authority

For public-sector work, tiers align with Virginia's enterprise risk model (VITA EA-225) and the relevant executive orders — cited inline, not assumed.

REQUIRES

Human in the loop

Rights-affecting and high-risk use cases require a documented human decision before production. No autonomous sign-off.

LOGS

Audit trail

Every stage is persisted — inputs, gates, conditions, escalation, sign-off — so the record exists before the question is ever asked.

DISCLOSES

Data & provenance

Data sensitivity and model provenance are first-class inputs. The method is model-agnostic; for regulated and public-sector work it runs on US-domiciled, open-weight models on GovCloud-eligible infrastructure — no foreign-adversary models, scoped to the deployment we configure with you.

// The free governance brief

What you get back, in 24–48 hours.

Tell us where your AI is running today. We run up to three of your real use cases through the governance layer and send back a plain-English brief — what passes, what holds, and what you'd want to fix before an auditor or a customer finds it first.

  • A risk tier for up to three representative AI use cases you're running or planning
  • The gates each one passes — and the ones it doesn't
  • Foreign-model and data-residency exposure flagged explicitly
  • Where a human decision is required before you go to production
  • A sample decision card and audit-trail entry for one of your use cases
  • A short, prioritized list of what to govern first
// Inspectable, not a black box

The method is public. The governance framework these gates are built on is open for review on GitHub — read it before you ever talk to us.

github.com/Obsidian757/ai-governance-framework →

use_case  voice agent · resident intake
tier  HIGH
gate  foreign-AI  ⛔ fail
gate  human-handoff  required
decision  hold → committee review
logged  decision-card + audit trail

Illustrative output. A method and framework, openly documented — not a certified product.

// After the brief

The brief is the front door.

No obligation — the brief stands on its own. If you want to act on it, there's a clear path, sized to where you are.

STEP 1

Governance audit

A full pass across your AI use cases — every one classified, gated, and documented, with a prioritized remediation plan and decision cards you can hand to leadership or an auditor.

STEP 2

Working session

A workshop with your team to stand the governance layer up inside your own process — intake, gates, and the human-decision checkpoint, mapped to NIST AI RMF.

STEP 3

Ongoing assurance

Recurring review as you ship new AI — each use case governed before it goes live, with an audit trail that stays current for the next audit or board question.

// Start here

Request your free governance brief

No pitch deck. No commitment. Tell us what you're running and we'll send back where you stand.

What you share stays confidential — never sold, NDA available for paid work · info@12thhouseai.com

The governance brief is an informational, method-based assessment — not legal advice, a compliance certification, or a guarantee of any audit outcome. Briefs are prepared with AI assistance and human review, with sources cited. See Terms & Privacy.

● Received

Your governance brief is on its way.

Expect it within 24–48 hours, built around the AI you're actually running. Want to talk sooner? Book a 30-minute walkthrough →

Something went wrong sending that. Email us directly at info@12thhouseai.com and we'll take it from there.