Human approval for AI agents: three brakes for production
The pilot won the team over. Now the agent needs to issue an invoice, release a refund or change a customer record — and the question that stalls the meeting is a different one: who answers for it if it gets it wrong? Un
The pilot won the team over. Now the agent needs to issue an invoice, release a refund or change a customer record — and the question that stalls the meeting is a different one: who answers for it if it gets it wrong? Until that answer exists, the agent stays in the demo.
For an IT or operations director, "autonomous" isn't a compliment. It's a risk to be sized. This post shows the three brakes that turn an acceptable agent into an approvable one: human approval before sensitive actions, a spending cap per agent, and contextual policies. And, at the end, a distinction that often gets blurred in due diligence: evidence is not certification.
Human approval before sensitive actions
Not every agent action needs a human. Looking up an order, summarizing a history or answering a product question are read operations — cheap to get wrong and easy to fix. The problem lies in actions that change the state of something in the world: payments, cancellations, contract changes, sending communication on the company's behalf.
That's where human approval comes in (HITL, human in the loop). In Zihin, you define the scope of actions that require approval. When the agent reaches one of them, it stops and asks. The request shows up as a card in Zihin Chat, with the context of the action and two buttons: approve or reject. Only after the decision does the process continue — or not.
Two details make a difference in practice:
- One member can approve what another triggered. The governance rule allows the person who triggered the agent to be different from the one who authorizes the action. It's accountability between people, not just between a person and an agent.
- Approval applies outside the console too. Whoever operates the agents through an MCP-compatible client is still subject to the same policies: the action requires approval when the configuration says so.
Tip: Start from the opposite of the obvious. Instead of asking "which actions can the agent take on its own?", ask "which actions would I never accept explaining to an auditor without a human name next to them?". Those go into the approval scope. The rest can run.
A spending cap per agent
The director's second fear is less dramatic and more frequent: the bill. An agent stuck in a loop, a misconfigured integration or simply higher volume than expected — and model consumption gets out of control in the middle of the night.
The brake here is the per-agent budget: a dollar limit set for each agent. When it hits the cap, the agent stops on its own. This applies even when it runs on your own provider key (BYOK) — the limit belongs to the platform, not the provider, so it doesn't depend on alerts that arrive after the damage.
What this changes in the conversation with the board:
- Cost stops being a global estimate and becomes a cap per initiative. The collections agent has one budget; the support agent, another.
- Predictability doesn't require vigilance. Nobody has to watch a dashboard for the brake to work.
- Consumption can be checked in the console, and Zihin Watch sends weekly summaries and monthly reports with executions, AI costs and points of attention to the people responsible.
Contextual policies: what the agent never does
Approval handles specific actions. Budget handles money. The third brake handles everything that should be impossible, no matter who asked: contextual security policies (CSP).
A contextual policy is a rule enforced on every execution, before the tool runs. Five types cover most cases:
- Schedule — outside the defined window (business hours, weekdays, time zone), the agent doesn't run.
- Behavior — tone, language, what it must and must not do, including blocking specific tools and limiting iterations.
- Data — restricting access to sensitive fields and entities.
- Origin — where the agent can be triggered from (allowed IPs and domains).
- Custom — your own rules, in JSON, for what is specific to your operation.
Policies follow scope inheritance: you define them once at the company level and refine them per team, agent or user. The most specific rule wins. An LGPD (Brazil's privacy law) restriction written once applies to every agent, with no need to paste it into each configuration.
These policies add to access control: four-level RBAC (owner, admin, editor, member), enforced in the UI and the API, defines what people can configure. CSPs define what the agent can do at execution time. Details of both layers are on the security page.
What gets recorded — and why it matters
A brake without a record is a promise. The three controls above only serve the director if they can prove, months later, that they worked.
Every agent session is recorded in full: phases, tool calls, tokens, costs and latencies. All of it can be queried in the console and via the telemetry API. Session replay lets you reconstruct what happened in a past execution, call by call. Approval decisions are recorded before the process continues. Integration credentials live in a vault encrypted with AES-256-GCM, never in configuration files.
Note: This is evidence, not certification. Zihin does not hold SOC 2 or ISO 27001 certification. What it provides are verifiable technical controls, an audit trail and a DPA covering the Controller and Processor roles under LGPD and GDPR. Controls and a DPA are not equivalent to a certification — and a vendor that treats the two as synonyms deserves suspicion.
The distinction matters because the auditor's question is rarely "do you have the seal?". It's "show me what the agent did on the 14th, at 2:17 a.m., and who authorized it". Evidence answers that. A certification attests to a process, but it doesn't reconstruct a session.
According to Deloitte, only 21% of respondents in its 2026 survey say their organization has a mature governance model for AI agents. The number isn't about technology — it's about the absence of the brakes described here, with a record. Those who have them are in the minority.
Conclusion
An agent is acceptable to a director when they can answer three questions without hesitating: who approved the sensitive action, how much the agent can spend, and what it could never have done. Human approval, spending caps and contextual policies answer each of them — and leave a record.
If your operation has agents stuck in the demo for lack of those answers, it's worth mapping which actions require approval, which cap fits each initiative, and which rules already exist on paper without being enforced. See the plans, check the FAQ or book a diagnostic.
Sources
- Deloitte, State of AI in the Enterprise 2026 — press release. Survey of 3,235 business and IT leaders in 24 countries; maturity as self-reported by respondents.
- Zihin.ai, Security and FAQ — controls, RBAC, contextual policies and position on certifications.
Originally published in Portuguese on the Zihin blog. I work at Zihin, which connects AI agents to a company's systems and processes, with permissions, human approvals and cost tracking.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.