AI decisions Operating model
Human in the loop vs autonomous agents: how much should an AI agent decide on its own?
Keep a human approving each action when an error is costly, hard to reverse or affects people's rights, and let the agent act alone when actions are low-impact, reversible and well measured. Autonomy is not a yes-or-no choice but a setting per action, raised step by step as evidence accumulates. Whatever the level, every action must be logged and a person must be able to stop or override the system.
The options
Human in the loop
The agent prepares or proposes each action, and a person reviews and approves it before it takes effect.
Autonomous agent
The agent executes actions on its own within defined limits, and people supervise through monitoring, sampling and exceptions.
Side by side
| Criterion | Human in the loop | Autonomous agent |
|---|---|---|
| Who decides each action | A person, with the agent's proposal and evidence in front of them. | The agent, within scoped permissions; people decide the rules and handle what is escalated. |
| Throughput and latency | Limited by reviewer capacity; queues build up at peaks and response times stretch. | Scales with volume and responds in seconds, including outside working hours. |
| Cost of an error | Most errors are caught before they reach a customer or a system of record. | Errors reach production and are found afterwards, so reversibility and blast radius matter. |
| Main human risk | Automation bias and fatigue: reviewers start approving by default and the check becomes a rubber stamp. | Loss of situational awareness: nobody looks closely until an incident forces them to. |
| Auditability | Each decision has a named approver, which makes accountability simple to show. | Needs complete logs of inputs, plans, tool calls and outcomes to reconstruct why something happened. |
| Learning signal | Every approval, edit or rejection is labelled feedback to improve prompts, tools and evaluations. | Feedback comes from sampled reviews, exceptions and outcome metrics; weaker unless designed in. |
| Fit with high-risk uses | Close to what the EU AI Act asks for high-risk systems: people able to understand, override and stop the output. | Possible only if effective oversight is still built in: monitoring, intervention and a way to halt the system. |
| Decisions about individuals | A meaningful human review helps meet GDPR Article 22 on decisions with legal or similar effects. | Solely automated decisions with legal or similarly significant effects are restricted under GDPR; check the exceptions and safeguards. |
Choose Human in the loop when…
- The action is hard or impossible to reverse: payments, contract changes, deletions, messages sent to customers.
- The decision affects a person's rights or access to a service, such as credit, insurance, employment or public benefits.
- The use case may be high-risk under the EU AI Act, or falls within GDPR Article 22.
- The agent is new, the evaluation set is small, or the process has many undocumented exceptions.
- You need labelled feedback to improve the agent before widening its scope.
Choose Autonomous agent when…
- Actions are low-impact and reversible, such as drafting, tagging, routing or enriching records.
- Volume or response time makes per-item human review impractical, and the queue would defeat the purpose.
- The agent has a track record on a fixed evaluation set and in shadow mode, with error rates the business accepts.
- Permissions, spending limits and rollback are enforced by the system, not only by the prompt.
When to combine them
Most production agents use both, decided per action rather than per agent. A practical pattern is tiered autonomy: the agent acts alone on low-risk, reversible steps; it asks for approval above a threshold of amount, confidence or impact; and it escalates anything outside its policy. Start in shadow mode, then human approval, then autonomy on the lowest tier, and move each threshold only when evaluation and incident data support it. Keep a stop switch and sampled human review even at the highest level.
Common mistakes
- Treating autonomy as one switch for the whole agent instead of a threshold per action and risk tier.
- Adding a human approval step without time, context or authority to say no, which turns oversight into a formality.
- Letting the agent hold broad credentials and relying on the prompt to keep it within limits.
- Raising autonomy because the demo went well, without an evaluation set, shadow-mode data or a rollback plan.
- Logging only final outputs, so nobody can reconstruct which inputs, plan and tool calls led to an action.
How Thinkia approaches it
We decide autonomy per action, not per agent. With the process owner we list each action the agent can take and rate it by impact, reversibility and who it affects. That gives a small set of tiers: act alone, act and notify, ask for approval, never do. Agents run first in parallel with people, and autonomy grows only as the evidence does: evaluation results, shadow-mode comparisons, rework and incident data.
Technically, we separate reasoning from execution. The model proposes; a governed layer checks each action against policy, executes it with scoped credentials and logs it. Synapse provides that common layer: corporate SSO, a gateway with rate limiting, model routing across providers and RAG on the client's own infrastructure, so approval steps, limits and logs are the same for every agent instead of being rebuilt by each team.
If a use case may fall into a high-risk category under the EU AI Act, we map the role of each party and the human oversight requirements before production: who oversees, with what training and authority, how they override or stop the system, and which logs are kept. Our AI governance guide covers risk tiers, roles and a practical checklist; it is operational guidance, not legal advice.
Thinkia products involved
- SynapseGoverned agentic platform: agents, models, costs and data in one place.
- EU AI Act governance guideRisk tiers, timeline, roles and a 20-point checklist. Not legal advice.
Related AI solutions
- Agentic process automationAutomate the workflows that rule-based tools can't touch.
- AI governance, risk & controlAI your board, legal team and regulators can sign off on.
- AI claims triageFrom first notice to the right desk, on day one.
- AI Act governance for underwritingHigh-risk under Annex III. Documented, logged, defensible.
- AI compliance monitoringStay compliant without making compliance a full-time job.
Frequently asked questions
Does the EU AI Act require a human to approve every decision?
No. Article 14 requires high-risk systems to be designed so that people can effectively oversee them: understand their limits, be aware of automation bias, interpret the output, decide not to use it or override it, and intervene or stop the system. How that is done depends on the risk and context. Check your case against the Regulation text and qualified legal advice.
Is a human in the loop enough to make an agent compliant?
Not on its own. Oversight has to be effective: the reviewer needs the right information, competence, time and authority to reject the proposal. A person who clicks approve on everything provides little protection and little evidence.
How do we decide the autonomy threshold?
Rate each action by impact, reversibility and who it affects, then agree with the process owner the error rate the business accepts for each tier. Measure the agent against a fixed evaluation set and in shadow mode, and allow autonomy only where results stay within that limit.
What should we log for an autonomous agent?
Enough to reconstruct any action: the input, the context retrieved, the plan, each tool call with its parameters and result, the model and prompt versions, and who or what approved it. Define retention and access to those logs in advance.
Can autonomy be reduced again after it has been granted?
It should be. Treat autonomy as configuration, not as a design decision fixed in code, so you can lower a threshold or return an action to approval when monitoring shows drift, a new type of error or a change in the process.
Keep exploring
Related decisions
- EU AI Act provider vs deployer: which role are you, and what does each one owe?
- AI agents vs RPA: which one should automate your process, and when should you combine them?
- AI pilot vs production: what actually changes when you scale?
- LLM evals vs human review: how should you assure the quality of enterprise AI?
- Centralised vs federated AI governance: who should decide what in your organisation?