The Risk of Over-Delegating Decisions to AI Agents
The Risk of Over-Delegating Decisions to AI Agents
It starts small. Let the agent sort your emails by priority. Let it schedule meetings based on your availability. Let it draft responses to routine messages. Each delegation makes sense individually. You are freeing up time for more important work.
But six months in, you notice something. You do not know what is in your inbox anymore. You cannot remember the last time you read a meeting invite. You approved three commitments this week based on one-sentence agent summaries without reading the actual requests.
You did not delegate tasks. You delegated judgment.
What the Research Shows
This pattern has a name in cognitive science: cognitive offloading - delegating cognitive work to external systems. Smartphones already changed how we handle navigation and phone numbers. AI agents are accelerating the same dynamic at higher cognitive levels.
A 2025 meta-analysis found a negative correlation between frequent AI usage for decision-making and critical-thinking scores. The effect was most pronounced for tasks where the AI was handling the judgment layer, not just the execution layer. People who had AI summarize and prioritize their email performed measurably worse on unrelated judgment tasks compared to people who handled their own email.
The Nature paper "Delegation to artificial intelligence can increase dishonest behaviour" (2025) found a related effect: when people delegated to AI agents with vague instructions, they were more likely to endorse outcomes they would have rejected if they had been directly responsible for the decision. The delegation created psychological distance from consequences.
Both findings point at the same mechanism: delegating judgment to an AI agent does not just save time. It degrades your ability to exercise judgment independently.
The Delegation Ratchet
Delegation tends to expand in one direction. You start by having the agent filter spam. Then it prioritizes your real emails. Then it drafts responses. Then it sends responses with your approval. Then you start auto-approving because the drafts are usually fine.
Each step is a small, rational optimization. The cumulative effect is that you no longer have direct contact with the information your decisions are based on. You are making choices based on the agent's interpretation of reality, not reality itself.
The ratchet has no automatic stop. The agent's outputs are usually good enough that there is always a case for going one step further. The question is not whether any individual delegation is reasonable - it is whether the accumulated delegation leaves you with enough direct exposure to exercise independent judgment when it matters.
What Gets Lost
Direct exposure to raw information builds intuition that summaries cannot replicate. Reading customer emails - even the routine ones - gives you a feel for sentiment, urgency, and relationship quality that no summary captures. Scanning your calendar manually maintains a physical sense of time commitment that automated scheduling removes.
The friction of manual processing is not only waste - it is signal processing. When you read the email yourself, you pick up on tone, word choice, what was not said, who was cc'd and who was not. An AI summary captures the explicit content but misses the subtext that experienced judgment relies on.
When you outsource this processing entirely, you lose the ability to notice subtle patterns:
- The slightly unusual phrasing in a client email that signals something is wrong
- The gradual increase in meeting frequency that indicates scope creep
- The one task that keeps getting rescheduled - possibly because it is blocked and nobody has surfaced it
- The pattern of which requests come through formal channels versus informal ones
These signals exist in the raw data. They rarely survive the summarization step.
A Healthier Delegation Model
The useful distinction is between delegating execution and delegating judgment. Execution delegation - having the agent send a calendar invite you approved, draft a response you reviewed, extract action items from a meeting you attended - frees up time without degrading judgment. Judgment delegation - having the agent decide what is important, what requires your attention, what is an acceptable outcome - is where the risk accumulates.
Practical boundaries that preserve judgment while still using agents effectively:
Keep the agent for execution, own the inputs. Let the agent draft responses but read the original email before approving. Let the agent schedule meetings but review the actual requests. The agent saves time on composition and scheduling while you retain direct exposure to the information.
Audit what the agent handled without you. Once a week, look at the actions the agent took autonomously. Not to micromanage - but to check whether you are surprised by what you find. If you regularly discover commitments you did not know about or responses that went out in your name that you would not have approved, the delegation has gone too far.
Protect direct channels for high-stakes relationships. Some communications should never be filtered or summarized before you read them - key clients, your direct reports, your manager, anything that affects important decisions. Being unreachable to an agent for these channels is a feature.
Set explicit autonomy limits by consequence level. An agent can send routine meeting confirmations without review. It cannot agree to project changes or make commitments on your behalf. Document these limits explicitly rather than letting them evolve implicitly.
The goal of AI agents is to amplify your judgment, not replace it. Judgment is a capacity that atrophies without use. The moment you stop exercising it is the moment you start losing access to it.
Fazm is an open source macOS AI agent. Open source on GitHub.