What to require from Agentic AI software, from day one

• Tenant isolation and data sovereignty: Agents must run in environments where tenant data is isolated and cannot be used to train public models. If a vendor can’t guarantee this, treat it as a non‑starter for regulated data.

• No implicit training: Explicit contractual terms are needed to prohibit vendors from using your inputs for model training or telemetry sharing without consent. Demand Data Processing Agreements (DPAs) and technical assurances.

• Least privilege and identity controls: Agents should authenticate via your enterprise identity provider, use scoped tokens, and obtain elevated privileges only via just‑in‑time mechanisms. Never embed broad static credentials in agents.

• Encryption and secrets management: Encrypt data at rest and in transit , and store keys and secrets in an enterprise vault (Key Vault or equivalent). Rotate secrets on a scheduled cadence and on any personnel change.

• Explainability and full action logs: Every decision and external action must be logged with the extracted evidence and the decision rationale. Logs must be tamper‑resistant, retained for audit periods required by law, and easily exportable for investigations. Think of logs as your legal receipts.

• Human‑in‑the‑loop gating: For any financial, legal or high‑risk customer action, require human approval until the agent meets validated accuracy thresholds and you have a defensible audit trail.

Operational controls and testing

• Treat agents as production services: Include them in change control, CI/CD, monitoring and incident response plans. Apply the same SLAs and runbooks you use for core systems.

• Penetration testing and red teaming: Agents increase attack surface. Require inclusion of agent orchestration layers in annual penetration tests and client‑requested assessments.

• Accuracy targets and validation: Define business‑grade accuracy (for example, 99% for document validation tasks) and require automated validation tests in the deployment pipeline. Measure exception rates and human override frequency.

• Privacy Impact Assessment: Any agent that touches personal data must have a DPIA (or local equivalent) before production deployment. The DPIA should be documented and retained.

Data privacy and explainable decision-logging are the operational backbone of any enterprise agent programme. If you want the productivity and speed gains agents offer, you must also accept the discipline: isolate tenant data, prohibit unauthorised model training, vault credentials, require JIT elevation, and keep immutable, searchable logs for every action. Treat agents as production systems subject to the same change control, testing and incident playbooks as your core services.

Start with a narrowly scoped pilot that requires human‑in‑the‑loop gating for sensitive outcomes, and measurable accuracy and exception KPIs. Use the pilot to validate your technical assurances (tenant isolation, key management, explainability logs) and your contractual protections (DPAs, audit rights, incident timelines). Only scale autonomy when the evidence demonstrates repeatable accuracy, robust auditability and acceptable operational risk.

To onboard governed, private and traceable AI agents, reach out to the CohesionX team today: info@cohesionx.co.za