Best Practices
Ten enterprise-grade practices for governing AI agents in production. Each one closes a class of incident we have actually seen in real enterprise deployments.
Every agent starts with zero tool access. Add allowlist entries explicitly per agent. Most production agents need 5–10 tools, not 50.
Teams that hand agents a full MCP server end up with agents calling tools nobody knew existed.
PII redaction should happen before a prompt leaves your perimeter. Once a payload reaches a vendor model, it is too late.
Redacting in vendor logs is not redaction. Vendors retain raw payloads; your audit log will not.
Modern PII redaction needs to handle UK sort codes, IBANs, MRNs, NHS numbers, and dozens of formats. Use Presidio or equivalent — not a regex grab-bag.
A regex for "SSN" misses 75% of the formats your real customer data contains.
Some policies are fuzzy: consent verification, data-purpose alignment, data minimisation. Use an LLM-as-judge to evaluate them and block on violation.
Trying to encode "is this consent-valid?" as regex always produces false-negatives.
Bulk deletes, financial transfers, mass writes — never autonomous. Require a workflow approval (human or programmatic) at the policy layer.
An agent that "helpfully" cleans up test accounts will eventually delete real ones.
Agents act on their own credential, not the calling user. This makes audit attribution clean and revocation possible without affecting humans.
When agents inherit user credentials, every audit log entry says "the user did it."
Free-text logs are useless under audit. Every entry should be a structured record with tool, policy, redaction events, latency, tokens — ready to filter and ship to your SIEM.
A regulator that asks "show me every agent that touched PII in Q3" cannot wait for you to grep.
Each log entry should map to control IDs in GDPR, HIPAA, SOX, ISO 42001 — whichever applies. Build the mapping once at the platform layer, not per-agent.
Compliance teams that have to rebuild the regulatory mapping for every new agent will revolt — rightly.
A governed agent today is not a governed agent forever. Run scheduled audits across every agent to catch policy drift, new tools, model changes, redaction regressions.
A non-allowlisted preview model swapped in by a developer at 4pm Friday is the kind of thing only an auditor or a regulator catches.
Your governance layer must not be tied to a single LLM provider. Switching models should not require rewriting policy logic.
Vendor-specific governance pins you to a single provider and makes negotiation impossible.
The questions enterprise buyers, risk teams, and AI platform leads ask before deploying agents.
Ready to govern your AI agents? Let us know about your use case and we'll help you get started.