How our agents behave.
Six rules that decide what the AI can do on its own and what waits for a person.
Six rules our agents run under.
Each rule is enforced in TruOps AI, and you can check all six on one of your own documents.
- 01Every answer cites its source.
Anything the AI fills in shows its source, how recent that source is, and a confidence score. If it can't find a source, it leaves the answer blank.
- 02The AI never approves its own work.
The AI drafts and people check. Whatever produced a value can never approve it, and an auditor can see in the audit log who did what.
- 03Confidence decides what gets reviewed.
Drafts above a confidence threshold you set are grouped for one-click bulk approval, the middle band gets line-by-line review, and anything ambiguous goes to a person.
- 04Agents get only the access they need.
Agents run with their own limited permissions and follow the same access rules as your users.
- 05When it's unsure, it asks.
When sources disagree, it flags the conflict for a person to resolve. When evidence goes out of date, the control's status drops.
- 06One record of everything.
Every action, by a person or an agent, goes into one audit log. Completed results are saved as of their date, so you can show where you stood on any day.
How you verify this in a demo
In a 30-minute demo on one of your own documents, you should be able to see each of these for yourself:
| Commitment | What to look for |
|---|---|
| Every answer cites its source | Open a pre-filled assessment. Each drafted answer lists the document, control, or past answer it used, how recent that source is, and a confidence score. Where there is no source, the answer is left blank. |
| The AI never approves its own work | The identity that drafted a mapping or finding cannot approve it. Try to complete a run yourself and you will see that a person still has to confirm it. |
| Confidence decides what gets reviewed | A rule you set can accept high-confidence exact matches automatically. The middle band gets one-click review, and ambiguous items wait for a person. Ask us to show all three on a live mapping. |
| Agents get only the access they need | TruPilot runs with limited permissions under the same access rules as your users. |
| When it's unsure, it asks | When two sources contradict each other, both are flagged for review. When evidence goes out of date, the control's status drops. |
| One record of everything | Every draft, acceptance, rejection, and override, whether by a person or an agent, goes into one audit log. Completed assessments are locked as of their completion date. |
What the AI is allowed to do
What matters is which of the AI's changes stay as drafts and which become part of the record.
| The AI may | The AI may not |
|---|---|
| Read documents and tool data inside your tenant | Train foundation models on your documents, telemetry, or corrections |
| Draft answers, mappings, findings, ratings, and fixes, each cited | Approve its own draft, or sign off an assessment |
| Auto-accept only where you set a confidence rule (for example exact-match control maps) | Quietly overwrite a human decision |
| Group the same failure across many assets into one finding | Invent an answer when it has no source |
| Lower a control's status when its evidence gets old | Keep a control green after its evidence is past its check schedule |
How we handle your data
- Your data stays in your program. Documents, telemetry, and corrections stay in your tenant. They are never used to train foundation models or shared with other tenants.
- Least-privilege integrations. Connectors to your cloud, identity, endpoint, vulnerability, and code tools use limited permissions you can revoke.
- Tenants are isolated. Every environment, whether an enterprise, an MSSP client, or a portfolio company, has its own data, evidence, and access controls.
Three questions for any AI GRC vendor
Ask three things: where a person approves, whether the AI's actions are logged well enough to serve as evidence, and whether every answer traces back to a source. TruOps shows all three in a demo. For more, read Agents without a human gate and human-in-the-loop AI in GRC. How we protect the platform itself is on Trust & security.
Questions
Does TruOps use customer data to train AI models?
No. Your documents, telemetry, and corrections improve your own program only. They're never used to train foundation models or shared with other customers.
Can TruOps' AI act without a person approving?
Only within rules you set, such as auto-accepting exact-match control mappings above a confidence threshold. Everything else is reviewed by a person, and the AI can never approve its own work.
What happens when the AI is wrong?
A person rejects or corrects it. The correction stays in the audit log. Later drafts in your tenant can learn from the correction, but it is never used to train a foundation model.
Is the audit log evidence an auditor can use?
Yes. Every action by a person or an agent is recorded. Completed assessment results are saved as of their date, so you can show where you stood on a day an examiner names.
How is this different from a GRC chatbot?
A chatbot answers questions about your data. TruOps does the work itself, drafting assessments, control mappings, and findings with citations, and leaves every decision to a person. See What is AI GRC?.
Related
Why a GRC agent with no human approval step is a liability.
→LearnHuman-in-the-loop AI in GRCWhy AI should draft and people should decide.
→PlatformTruPilotHand TruPilot a task and it does the work, with the records behind every answer.
→CompanyTrust & securitySOC 2 Type II, data isolation, and how a security review works.
→See it run on your own data.
Book 30 minutes with a GRC specialist. Bring a real document, like a SOC 2 report, risk register, or vendor list (redacted is fine), and we'll show TruOps AI working on it.