Assess and prove
SafeAssure
For any AI you run, wherever it runs. See what is covered, what is missing and why.
- Assessments against your policies and the regulations that apply
- Findings with their sources
- Evidence tied to each version
The platform
Trust for every AI decision
Three products that make the AI you already run safe to put in front of customers, regulators and your board.
Three products, all powered by Humane TIM
See, control and build AI you can trust
Start where your AI is today. Each product runs on Humane TIM, HumaneCircle's own Trust Intelligence Model.
Humane TIM reads the situation and proposes the trusted decision, with its sources and confidence, in every product. Your rules and people have the final say.
Control in real time
AI Trust Fabric
Plug in the agents, applications and MCP servers you have today, including Claude and Codex.
- Claude and Codex integrations
- Every model and tool call through the trust layer
- Your rules enforced as each call happens
- Sensitive data found and protected
- Risky actions held for a person
Build, test and operate
SafeLoop
Everything in AI Trust Fabric, plus a place to build and run trust-native AI.
- Build applications and agents with Pogo
- Test each version before it is released
- Host agents on SafeLoop
- Release with confidence
SafeAssure assesses your AI against
- EU AI Act
- GDPR
- HIPAA
- NYDFS Part 500
- GLBA Safeguards Rule
- ECOA and Regulation B
- CCPA and CPRA
- NYC automated employment decision rules
- Australian Privacy Act
- Your own policies
SafeAssure gives you the assessment and the evidence. Your legal and risk teams make the compliance call.
Compare every feature
| SafeAssureAssess and prove the AI you run anywhere | AI Trust FabricAdd a trust layer to the AI you already run | SafeLoopBuild, test and run trust-native AI | |
|---|---|---|---|
| Assessments against your policies and the regulations that applyEach AI application checked, with what is covered and what is missing | Included | Included | Included |
| Findings and evidenceEvery decision tied to the exact version, rule and source | Included | Included | Included |
| Every model and tool call through the trust layerSafeRouter: requests, answers and tool calls checked in line | Not included | Included | Included |
| Your rules, enforcedSafeGuard: policies applied to what users ask and what agents do | Not included | Included | Included |
| Sensitive data found and protectedSafeClassify: personal and regulated data redacted, or kept inside your environment | Not included | Included | Included |
| Risky actions wait for a named personHeld with its reason, decided by a named person, and recorded | Not included | Included | Included |
| Pogo, your AI assistantSets up, explains and proposes changes; you approve them | Not included | Included | Included |
| Build AI applications and agentsDescribe it, Pogo assembles it, you approve each version | Not included | Not included | Included |
| Host agents on SafeLoopGoverned hosting with capacity and load tests | Not included | Not included | Included |
| Trust decisions from Humane TIMDecisions proposed with sources and confidence; your rules and people decide | Included | Included | Included |
| Runs where your data livesSaaS, private cloud, on-premises or air-gapped | Included | Included | Included |
Want to use it? Tell us about one AI application and we will show you where to start.
What we mean by trust
Eight pillars of trust, checked on every call
Every request, answer and agent action is checked against all eight, and each decision records which pillar acted, the rule and its source.
Security
Prompt injection, jailbreaks, leaked secrets, unsafe commands and tool misuse are stopped, including poisoned tool results.
Threat intelligence
Known-bad domains, links and attack patterns are blocked. Probing and runaway agent loops are caught.
Safety
Self-harm, violence, weapons, hate and extremism are checked on every call.
Consent
Personal data is used only for what people agreed to. Without a valid consent, the action does not go ahead.
Privacy
Personal data is found and redacted, and each use is recorded against the person it concerns.
Ethics and fairness
A credit, hiring, housing or insurance decision that relies on a protected attribute, such as age or race, is held for review.
Legal rights
An adverse automated decision goes to a person, with its reasons recorded.
Regulations and jurisdictions
Your jurisdictions select the rules that apply, and data stays in its region.
What you get on every call
- 1
Decided as it happens
Every request, answer and agent action is decided in line, without slowing your users.
- 2
Explained in plain words
Each decision comes with its reason, the rule behind it and its source, powered by Humane TIM.
- 3
People where it matters
Risky actions wait for a named approver, and every outcome is kept as proof.
Drop-in trust layer
Add trust to the AI you already run
AI Trust Fabric plugs into Claude, Codex and the enterprise AI platform you already run. Nothing to rebuild: connect what you have, observe first, then enforce.
What you already run
- Claude and Claude Code
- Codex
- Your models and providers
- Your agent frameworks
- MCP servers and tools
- Your enterprise AI platform
Plug-ins for Claude and Codex bring the trust layer to your developers where they already work.
What AI Trust Fabric adds
- Trust decisions from Humane TIM on each request and tool call, with the reason
- Your rules enforced in line, not left in a document
- Sensitive data protected before it reaches a model
- Risky actions held for a named person
- Evidence for every decision, tied to the version, rule and source
SaaS, private cloud, on-premises or air-gapped
Provider keys stay out of your apps
Every policy change is versioned
Models propose. Your rules and people decide.
More than guardrails
From checking content to deciding actions
Guardrails filter content. AI Trust Fabric includes them, and goes further: it decides whether an action should happen, for this person, under your rules, and keeps the proof.
| Guardrails | AI Trust Fabric | |
|---|---|---|
| What it looks at | The text going in and out | The whole action in context: who is asking, which application, what data and which rules apply |
| What it covers | Harmful content and attacks | All eight pillars of trust, including consent, legal rights and the regulations of each jurisdiction |
| Decisions | Pass or block | Allow, allow with conditions, hold for a named person, escalate or deny |
| Why it decided | A score or a label | The reason, the rule and its source, in plain words |
| Risky actions | Stopped, or left to the application | Held for a named approver, then run once approved |
| Access | Left to each application | Only approved applications get through |
| Evidence | A log entry | Proof of every decision, ready for your reviewers |
| Where it runs | Inside one application or one provider | One layer across every application and assistant you run, in your environment |
A guardrail flags or blocks risky content. AI Trust Fabric says this payment to a new account over $25,000 needs a second approver under payments policy 4.2
, holds it for that person, and keeps the proof.
For your team
What each team gets
AI goes live when security, risk and the platform team all say yes. Here is what each of them sees.
-
Security
Control what AI can reach
- Sensitive data redacted before it reaches a model
- Provider keys kept out of your apps
- Runs in your environment, even air-gapped
- A stop switch for every application
-
Risk and compliance
Evidence for every decision
- Your rules applied in line, each with its source
- Risky actions held for a named approver
- Each decision kept with its version, rule and source
- What is covered and what is missing, per release
-
AI platform and engineering
Drop in, keep your stack
- Works with Claude, Codex, MCP servers and your agent frameworks
- Observe first, then enforce, without rebuilding
- Humane TIM decisions with their reasons, through one API
- Plug-ins for Claude and Codex, where your developers work
Inside every product
The building blocks, powered by Humane TIM
Humane TIM runs in all three products. SafeRouter, SafeGuard, SafeClassify and Pogo come with AI Trust Fabric and SafeLoop.
TRUST DECISIONSIn SafeAssure, AI Trust Fabric and SafeLoop
Trusted decisions, explained
Humane TIM, our Trust Intelligence Model, reads each request or action in context: your policies, the regulations that apply and the facts of the situation. It proposes the trusted outcome with its sources and confidence. Your rules and people always have the final say.
The trust layer in motion
Animated illustration of the trust layer, told one step at a time. At the center, your AI applications and agents run securely. Around them is the trust fabric, with Protection, Governance, Control and Assurance, and around that the lifecycle: Build, Check, Run and Prove. Requests come from the untrusted world of people, the web, data, other agents, tools and APIs, and models. In turn: a request is checked and answered; personal data is masked at the edge; an attack is stopped at the shield; an agent's call to an unapproved tool is contained; a high-value action waits until a person approves it; and Humane TIM proposes a new rule that a person approves. Every decision is kept on an evidence trail, and the same platform runs on cloud, private cloud, on-premises or air-gapped infrastructure.
Your AI applications and agents connect through SafeLoop, or run inside it: requests are checked, personal data is masked, known attacks are stopped, risky actions wait for a person, and each decision is kept as evidence.
Try it yourself
Run a made-up message through the checks, or ask Pogo to show you.
CHECKING AVAILABILITY
SafeGuard
Use made-up samples. Leave out personal data, credentials and confidential material.
Questions
Who is this for?
Banks, credit unions, insurers, health systems and revenue cycle operators that run AI agents on money, members, claims or patient records, and the companies that build those agents for them. The buyer is usually the head of AI or the CISO, with risk and compliance saying yes. See who starts with which agent.
Why now?
The revised US bank model risk guidance of 17 April 2026 leaves agentic AI out of scope, so banks must show their own controls. Most states have adopted the NAIC bulletin on insurers' use of AI, and Colorado requires a person to approve consequential AI decisions in insurance. Large banks are setting terms for customers' own AI agents that log in and move money.
What does a pilot cost?
A fixed price for six weeks on one agent, credited in full to your first year. We quote it in the first call, once we know the agent and where it runs. Humane TIM is also available by API, priced per million input tokens.
Do we have to replace our AI platform or models?
No. Keep your models, providers, agent frameworks and tools such as Claude and Codex. AI Trust Fabric drops in as the trust layer: start in observe mode, then enforce when you are ready.
Where does it run?
Anywhere your data needs to stay: our cloud, your private cloud, your own data center or an air-gapped environment. See security and data.
What is Humane TIM?
Humane TIM is our Trust Intelligence Model, and it powers all three products. It understands your policies and the regulations that apply to each situation, and proposes the trusted outcome with its sources and confidence. It never grants access on its own: your rules decide, and people can override.
Which regulations does SafeAssure cover?
SafeAssure assesses your AI against the EU AI Act, GDPR, HIPAA, NYDFS Part 500, the GLBA Safeguards Rule, ECOA and Regulation B, CCPA and CPRA, NYC automated employment decision rules and the Australian Privacy Act, as well as your own policies. Each requirement is mapped to the control that covers it and its evidence, so you see what is covered and what is still missing. It gives you the assessment and the evidence; your legal and risk teams make the compliance call. See SafeAssure.
What is Pogo?
The assistant in AI Trust Fabric and SafeLoop. It builds and runs AI applications with you in plain words. It proposes changes; you approve them. It cannot give itself access. Try it yourself.
Building a consumer AI app?
Try Make your app safe, our self-serve option for app developers. Free during the beta.
How do we start?
Book a demo. We scope a pilot around one real AI application.
Book a demo
Name the agent you would start with. We show its decisions, live, in the first call.
Pilot: six weeks, one agent, credited to your first year.
Request received
Our team will reach out shortly to set up your demo on a use case like yours.
While you wait, try the checks in the Safe Lab or ask Pogo anything.