1. Rules decide, not models
Every action an agent proposes should be checked against your rules before it runs, with the decision and the reason recorded. Unknown actions should never run.
2. People for high-risk work
High-risk actions should always need a named person, with maker-checker and four-eyes for the riskiest items. This rule should be locked, not optional.
3. Data protection
- Each business sees only its own data.
- Data hosted in a region you choose.
- Personal data masked for roles that do not need it.
- Every access logged with its purpose.
- Support for erasure requests.
4. Model controls
- Only approved AI models.
- Screening for prompt injection.
- Personal data kept out of prompts.
- Spending caps per business.
- Agents tested on your data before go-live.
5. Control in an incident
An instant kill switch for one agent, one business or everything; an incident process with drills; and live tracing of every run.
6. Proof
A tamper-evident audit trail, versioned rules with full history, an inventory of AI agents with owners and risk levels, and an evidence pack auditors can verify.
7. Regulatory alignment
Map your controls to the frameworks that apply to you, such as India's DPDP Act, sector regulators' guidance, ISO/IEC 42001, the NIST AI RMF or the EU AI Act. A good platform is designed so you can meet your obligations; compliance remains yours to declare.
See AI agents on your own process
DOS runs your operations with AI agents under your rules, with people approving what matters. Try the interactive tour or book a 30-minute demo.
Frequently asked questions
Who owns AI governance?
The business does. A platform should make it easy to define, enforce and prove your rules.
Does governance slow AI down?
Done well, it speeds adoption, because risk and compliance teams can approve automation with confidence.