Controlling AI Agents: Why the Verification Layer Must Be Deterministic
In short
AI agents produce more and more work – but verification must remain deterministic. Zurich-based Klarent's seed round highlights 'deterministic autonomy': AI writes tests, execution stays deterministic. For Swiss SMEs that means fixed rules, confidence thresholds, and human escalation – and this control layer can be outsourced.
The Bottleneck Is Verification, Not Production
As AI agents generate more invoices, contracts, analyses and decision memos, the scarce resource shifts from producing to judging. This is the core shift from execution to evaluation. That principle underpins 'deterministic autonomy' – the approach of Zurich-based startup Klarent, which closed a €7.13M seed round in October 2026 with customers including NZZ, Sixt and JD Sports.
€7.13M
Klarent seed round, October 2026
Why You Must Not Let AI Control AI
The most dangerous mistake is giving verification to a second AI agent that 'corrects' the first. If the verification layer itself hallucinates, you trust a false control – the worst failure mode. Deterministic means: same input, same output, reproducible and auditable. For decision rights, a clear decision matrix helps; insights from executive AI competence discussions can sharpen this, as explored in the KI-Podcast episode on AI agents as decision-makers.
What Deterministic Means in Practice
- Fixed business rules: totals must match, sums must balance.
- Confidence thresholds: below a defined value, the case escalates to a human.
- Human-in-the-loop for exceptions: people review only escalations, not every item.
- Auditability: every decision is reproducible and traceable.
Verification as an Outsourcing Building Block for Swiss SMEs
This combination – agent plus deterministic verification plus human escalation – can be bought as an external service without building your own AI team. Building only the agent and forgetting the verification layer means buying risk, not value. For Swiss SMEs running Abacus, Bexio or M365, the verification rules must align with your master data and processes, and data handling should remain revDSG-compliant. How much human control is really needed is illustrated by the Klarna lesson on hybrid models.
Practical Matrix: Which Verification Layer Fits Your Risk Profile?
- Agent only, no verification layer
- Fast but uncontrolled. Errors surface late; trust suffers.
- AI checks AI
- Scales, but the verification layer can hallucinate. No audit trail.
- Deterministic rules + confidence thresholds
- Reproducible, auditable. Rules require upfront effort.
- Deterministic verification + human-in-the-loop (outsourced)
- Clear escalation, no in-house team needed. Best balance of control and effort.
The Most Expensive Mistake
Never trust a verification layer that is not deterministic. When AI 'corrects' its own control, you no longer have control – only a second opinion with the same weaknesses.
The next step is rarely building your own verification layer from scratch. Usually you start with one clearly scoped agent, define the critical rules, and outsource escalation to an external partner. That keeps you in control without hiring an entire AI team.
Frequently asked questions
- Why should the verification layer for AI agents be deterministic?
- Because only a deterministic layer is reproducible and auditable. Same input leads to same output, so errors cannot be 'corrected away' by a second AI.
- What does 'deterministic autonomy' mean at Klarent?
- AI writes the test code, but execution is deterministic – never by the AI itself. This keeps verification traceable while automating creation.
- How do I implement human-in-the-loop without manually reviewing everything?
- Define fixed rules and confidence thresholds. Only cases below the threshold or with rule violations go to a human; the rest passes automatically.
- Can I outsource the verification layer if I don't have an in-house AI team?
- Yes. The combination of agent, deterministic verification and human escalation can be purchased as an external service. The provider must align rules with your processes and systems.
- What role does revDSG play in AI agent verification?
- Verification often processes personal data. Rules and data handling must be revDSG-compliant, including clear access rights and logging.
Sources
Chris Jon Graf · AI Strategist & CEO
For over 15 years, Chris Jon Graf has built productive Swiss software platforms and advised executive teams on the strategic deployment of artificial intelligence. Would you like to evaluate this topic for your company?