The speed of errors changed
A traditional software bug is static: one defect, one feature affected. An agent's error is dynamic: one wrong judgment triggers a chain of follow-on actions, and by the time a human notices, ten steps have executed. Governance does not aim for a mistake-free agent; it aims for mistakes that cannot amplify and can be rolled back.
A typical scenario (illustrative, not a real client case): a support agent misreads "considering a return" as "confirmed return", raises a refund, emails the customer confirmation, updates inventory, and syncs the record to finance. Each step looks compliant alone; together they are a refund that should never have happened. An agent without approval points does not make one mistake; it makes a whole chain of them.
Where approval points belong: three grades by irreversibility
The yardstick is one question: if this action is wrong, can it be undone at no cost? Three grades follow (the grading table is the figure on this page):
- Grade 1: fully automatic. Reversible, low-impact, internal-only actions — retrieval, drafts, internal summaries. The agent runs them freely; humans sample the logs afterwards.
- Grade 2: human spot-check. Externally visible but retractable, medium-impact actions — sending drafted replies, bulk internal notices, bulk CRM updates. The agent acts; humans spot-check a proportion, and a failed check pauses that class of action.
- Grade 3: approval before action. Irreversible or externally binding actions — sending email, changing production data, moving money, quoting or committing to customers. The agent prepares; a human approves before anything executes.
Two supporting rules: the table is fixed before deployment and anything not on it defaults to Grade 3; and grading never ends — new tools, data sources and channels are graded before switch-on.
Approval is a release, not a brake
The common misunderstanding is that governance slows agents down. Without grading, a company has two options — let everything through (an incident waiting to happen) or have humans watch everything (which throws away the agent's speed advantage). With grading, low-risk actions run faster — nobody approves each small thing — and human attention concentrates where it is genuinely needed.
Audit is not for blame
A complete log — inputs, tool calls, outputs, approvals — exists for review and improvement: which task types fail most, which rules block legitimate work. An agent without logs cannot be iterated, only guessed at.
At minimum, log five things: the input, tools called and their parameters, the output, the approval grade and approver, and timestamps. Fix retention at pilot design time — too short and incidents are uninvestigable, too long and costs grow. Logs also carry compliance value: the AI Act requires high-risk systems to record events automatically (Art. 12), and GDPR requires records of processing (Art. 30).
The compliance dimension: three laws, three jobs
GDPR governs the data. When an agent processes personal data, the controller needs a lawful basis (Art. 6), data protection by design (Art. 25), processing records (Art. 30), appropriate security (Art. 32) and advance assessment of high-risk processing (Art. 35). Note Art. 22: solely automated decisions about individuals, including profiling, are restricted. Fines for breaching the core provisions run up to €20 million or 4% of worldwide annual turnover, whichever is higher (Art. 83(5)).
The AI Act grades systems by risk. High-risk systems require a risk management system (Art. 9), technical documentation (Art. 11), automatic logs (Art. 12), transparency information for deployers (Art. 13), human oversight (Art. 14) and post-market monitoring (Art. 72); deployers must assign oversight to natural persons with the competence, training and authority to do it (Art. 26). Conversational support is usually not high-risk, but users must know they are talking to AI (Art. 50). Fines run up to €35 million or 7% for prohibited practices and €15 million or 3% for other infringements (Art. 99).
NIS2 governs cybersecurity in critical sectors. It applies to essential and important entities in the Annex I and II sectors that meet the size threshold: management bodies must approve the cybersecurity risk-management measures and oversee their implementation (Art. 20), and the measures include supply-chain security (Art. 21(2)(d)) — covering the agent's model and tool suppliers too. Check scope against the annexes and the size threshold; when in doubt, the directive text and national law decide.
All three point at the same evidence: proof that you are in control. The governance framework is that proof, for regulators and customers alike.
What "enough governance" looks like
A workable minimum: the grading table is final and covers every action, approvers are named with cover, the five log fields are captured with fixed retention, and a monthly review covers error rates, overrides and legitimate work blocked.