# How do you decide what to delegate to an AI agent?

> An AI agent is not a tool you launch, it is a decision you delegate. Until the delegation doctrine is written, every agent built draws the line in place of the decision-maker, and the sum of those default, scattered choices is the company's real AI policy.

## 01 The turning point

Google Cloud presents Gemini for business as a universal agent to which you can « delegate tasks instead of supervising them ». Microsoft (Copilot Studio) and Anthropic (Agent Skills) also let teams build agents without heavy development. The technical barrier is falling. Feasibility is no longer the only question.

## 02 The question that remains

Building an agent is becoming trivial. Deciding what it is allowed to do is not. An employee who describes an agent in natural language (« analyse customer feedback and email a summary to my team ») makes, without stating it, **four decisions on behalf of the company**:

- which data is used,
- which judgment is delegated,
- which communication is sent in its name,
- which record will remain.
Without a written doctrine, those decisions are made by default, one by one, in silos. The sum of these choices is the company's real AI policy: the one no one signed.

## 03 The delegable / non-delegable matrix

Three criteria are enough to sort what can be given to an agent and what must stay on the decision-maker's desk.

The three criteria for delegating to an AI agent Criterion The question to ask If the answer is no
Reversibility Can the action be undone at no cost? A sent email, a wire transfer, a publication: you don't get them back.
Traceability Can you reconstruct who decided what, six months later? An agent writing to a system with no log cannot be controlled.
Attribution Who answers, by name, if the decision is wrong? If no one can be named, the decision is not delegable.
An action that fails all three is not an agent use case, no matter how convincing the demo. An action that passes all three can be delegated. In between, a named human guardrail is required.

## 04 The doctrine, in five lines

1
### By default, nothing is delegated

Adding an agent is a positive decision, not an experiment. An agent no one asked for does not exist.

2
### One agent = a written goal, a written scope, a named owner

The three fit on one page. Without them, creation is not complete.

3
### Every irreversible action stays under explicit human sign-off

Sending, publishing, transferring, filing, contracting. The agent prepares, a human signs.

4
### Every agent leaves a trail readable without technical skills

Who built it, what it does, what it must not do, when it was last reviewed. An executive must be able to read the inventory.

5
### An agent that loses its owner disappears

No orphan-agent debt. An employee leaving triggers a review of their agents.

Key takeaway Google and its competitors are putting agent building within reach of every employee. That is progress. But every agent built without doctrine does more than automate a task: it moves a decision. Drawing the line of what can be delegated before two hundred agents draw it for you: that is the real AI governance job of 2026.

## FAQ

**What is « agentic work »?**
The term, pushed by Google Cloud, Microsoft and Anthropic, refers to a way of working where AI agents run full workflows (not just tasks) using company data and third-party tools.

**How is it different from a chatbot?**
A chatbot answers a question. An agent pursues a goal: it plans, calls tools, writes into systems, makes micro-decisions. The delegation is deeper and harder to reverse.

**Should agents built by employees be banned?**
No. A ban breeds shadow AI. The right answer is a clear doctrine of what can be delegated and under what conditions, plus a readable inventory of the agents in circulation.

**How do you start a delegation doctrine?**
One page is enough. Three sorting criteria (reversibility, traceability, attribution), five default rules, one named owner for the agent inventory. The rest is refined in motion.

**Isn't the platform's security (Google, Microsoft) enough?**
No. Those platforms handle infrastructure security (encryption, isolation, access control). They do not decide for you which actions an agent is allowed to take on your behalf. That is what the doctrine is for.

**How does this relate to AI Act compliance?**
The AI Act sets legal requirements by risk level (banned, high risk, transparency). A delegation doctrine is an internal governance tool that lets the company actually meet those requirements, agent by agent.

## See also

- [AI agent](https://wyp.agency/en/glossary/ai-agent/)
- [AI agent platform](https://wyp.agency/en/glossary/agent-platform/)
- [Guardrails](https://wyp.agency/en/glossary/guardrails/)
- [Shadow AI](https://wyp.agency/en/glossary/shadow-ai/)
- [Prompt engineering](https://wyp.agency/en/glossary/prompt-engineering/)
- [Human in the loop](https://wyp.agency/en/glossary/human-in-the-loop/)

## Sources

- Google Cloud, « Gemini for business: the universal work agent », page accessed 14 October 2026 (https://cloud.google.com/gemini-enterprise)
- NIST, AI Risk Management Framework (AI RMF 1.0) (https://www.nist.gov/itl/ai-risk-management-framework)
- European Union, Regulation (EU) 2024/1689 on artificial intelligence (AI Act) (https://eur-lex.europa.eu/eli/reg/2024/1689/oj)

---
Entry from the WYP.agency AI glossary · https://wyp.agency/en/glossary/delegating-to-ai-agents/ · Last updated : 2026-10-14