How our AI behaves.
Six commitments about how TruOps' AI works, published so you can hold us to them — and test them on one of your documents.
- You are writing an RFP for “AI GRC” and need a test that is not a chatbot demo
- Legal, audit, or the CISO asked who is accountable when the model is wrong
- A vendor offered to auto-approve control mappings with no threshold you set
An AI you're responsible for should answer to you.
GRC is the job of proving things. We built TruOps' AI to meet the standard you hold everyone else to. These are commitments about how the product works, not aspirations.
- 01Every answer shows its work.
Anything the AI fills in carries its source, how recent it is, and a confidence score. If it cannot show where an answer came from, it does not give one.
- 02The AI never approves its own work.
The AI drafts; people check. Whatever produced a value cannot be the thing that signs it off, and an auditor can see that.
- 03Confidence decides who looks.
High-confidence drafts are accepted by rules you set, the middle band gets one-click review, and anything ambiguous goes to a person.
- 04Agents get only the access they need.
Agents run with their own limited permissions, under the same fine-grained access rules as your users.
- 05When it is unsure, it asks.
When sources disagree, it flags them for review and never quietly overwrites. When evidence goes stale, status drops instead of staying green.
- 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.
Hold us to these.
Bring one real document. Watch the program get set up from it.
How you verify this in a demo
A principle you cannot test is a slogan. In thirty minutes, on one of your documents, you should be able to see each of these:
| Commitment | What to look for |
|---|---|
| Every answer shows its work | 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. Empty answers are empty — not invented. |
| The AI never approves its own work | The same identity that drafted a mapping or a finding cannot be the approver. Try to complete a run: a person still has to confirm. |
| Confidence decides who looks | High-confidence exact matches can be accepted by a rule you set. The middle band is one-click review. Ambiguous items stay in a human queue. Ask us to show the three bands on a live mapping. |
| Agents get only the access they need | TruPilot runs with scoped rights, under the same fine-grained access as users. Integrations are read-only; proposed changes are drafts. |
| When it is unsure, it asks | Contradictory sources are flagged, not silently merged. Stale evidence lowers control status instead of staying green. |
| One record of everything | Every draft, accept, reject, and override — human or agent — is in one audit log. Completed assessments freeze as of their date. |
What the AI is allowed to do
The useful split is not “AI on / AI off.” It is which writes are drafts and which writes are 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 status when evidence goes stale | Keep a control green after the evidence has aged past its cadence |
How we handle your data
- Your data improves your program, not our models. Documents, telemetry, and corrections stay in your tenant. They are not used to train foundation models or shared across tenants.
- Integrations are read-only. Connectors observe cloud, identity, endpoint, vulnerability, and code tools with least-privilege scopes you can revoke. Proposed changes are drafts for a person.
- Tenants are isolated. Every environment — enterprise, MSSP client, or portfolio company — has its own data, evidence, and access boundary.
Why we publish this
Anyone evaluating AI for governance, risk, and compliance should ask where the human approval sits, whether AI actions are logged to an evidence-grade standard, and whether every answer can be traced to a source. These commitments are our answers. The longer argument is in 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. They are not used to train foundation models or shared across tenants.
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. The next draft in that tenant can use the correction; it does not become training data for 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 drafts the work in the registers and tables — assessments, mappings, findings — with citations, and leaves every decision with a person. See What is AI GRC?.
Related
Why an ungoverned GRC agent is a liability, not a feature.
→LearnHuman-in-the-loop AI in GRCWhy AI should draft and people should decide.
→PlatformTruPilotAI on every screen. Every answer lists the records it used.
→CompanyTrust & securitySOC 2 Type II, isolation, and how a security review runs.
→See it run on your own data.
Thirty minutes with a GRC expert, not an SDR. Bring one real document (a SOC 2 report, a risk register, a vendor list; redacted is fine) and watch TruOps set up a live program from it, with an assessment already pre-filled.