My AI agent refuses to perform certain tasks: why?

You submitted a request to your AI agent, and it replied that it couldn't process it. First reaction: annoyance. You pay for a service, it should obey. Yet this refusal is not a malfunction. It's a sign that the agent encountered a limit it is programmed to respect. Understanding these limits will save you time and help you collaborate better with your assistant.

IWhy does an AI agent refuse certain requests?

An AI agent is not an employee who executes everything you tell it to do. It operates within a framework defined by three types of limits. Ethical limits prevent it from participating in illegal or harmful actions. Technical limits concern what it can or cannot do. Mission limits correspond to what you configured it for. When a request falls outside this framework, the agent stops and notifies you.

IIThe three categories of refusals to know

1. 1
   Ethical refusal

   The agent detects a request that could be illegal, deceptive or harmful. For example, sending prospecting emails to individuals without their consent, or drafting a message that lies about a product.

2. 2
   Technical refusal

   The agent does not have the capabilities to execute the task. You ask it to modify an image when it only processes text, or to read a file in a format it does not recognize.

3. 3
   Mission refusal

   The request falls outside the scope you have defined. An agent configured to follow up on unpaid invoices will refuse to write a newsletter, even if it would be technically capable.

4. 4
   Security refusal

   The agent protects sensitive data. You ask it to send banking details by unsecured email, or to store passwords in a shared document.

IIIHow to distinguish a justified refusal from a real blockage?

A justified refusal usually comes with an explanation. The agent tells you why it cannot act and, often, what you could do instead. A real blockage, on the other hand, looks like a breakdown: the agent doesn't respond, loops endlessly, or produces a technical error message. In the first case, it's a normal limit. In the second, it's a problem to report.

Justified refusalTechnical blockage

Clear message explaining the reason

Incomprehensible error or no response

Suggestion of a possible alternative

No suggestion offered

The agent works for other tasks

The agent seems stuck on everything

The refusal makes sense (legal, ethical, mission)

The task seems simple and authorized

IVExamples of requests often refused

Some requests come up frequently and trigger a refusal. A craftsperson who asks their agent to send quotes to addresses bought online runs into the consent rule. An accounting firm that wants the agent to transmit bank details by regular email receives a refusal related to data securityA family asking its agent to order medication online falls outside what an agent can legally do. These refusals are not whims: they protect you.

VHow to rephrase a refused request?

When the agent refuses, reread its response. It often contains the key. If the refusal concerns the mission, check that your request matches what you defined at the start. If it is a matter of format, convert your file. If it is ethical or legal, ask yourself if your objective can be achieved another way. Sometimes, it is enough to break the task down into simpler steps.

1. 1
   Read the explanation

   The agent almost always indicates the reason for the refusal. Take time to understand it before trying again.

2. 2
   Check the mission

   Does your request fall within the scope you have defined? If not, create another agent or expand the mission.

3. 3
   Simplify the request

   A complex task can be broken down. Ask the agent to do the part it can handle first, then complete it yourself.

4. 4
   Accept the limit

   Some requests should not be automated. A refusal can be an opportunity to rethink your process.

VIWhen to accept the limit rather than work around it

I sometimes see users trying to rephrase their request ten different ways to force a result. That's rarely a good idea. If the agent refuses a task for AI ethicsreasons, insisting won't change anything and you'll waste time. It's better to ask yourself whether this task really needs to be automated. A phone call, an email written by you, a manual check: some actions benefit from staying human.

VIIAdapt your agent's mission to avoid unnecessary refusals

If you frequently receive refusals related to agent missionit may be that you defined a scope that's too narrow. Review the description of what your agent should do. Add the use cases you encounter regularly. Conversely, if the agent accepts tasks it shouldn't handle, tighten its mission. A good balance is found through use: the first few weeks are for adjusting.

• I read the explanation given by the agent
• I verified that my request falls within the defined mission
• I excluded any illegal or deceptive intent
• I considered doing this task myself if it is sensitive

Check what applies to you: the list remembers your answers on this device.

VIIIFrequently asked questions

?Questions we're asked

Can my agent send banking data by email?

No, and that's normal. Transmitting bank details or card numbers by unsecured email exposes you to fraud risks. The agent refuses to protect your data and that of your clients. Use a secure channel for this type of information.

Why is the agent refusing a task it completed yesterday?

The context may have changed. Perhaps your request today contains a new element that the agent identifies as problematic. Compare the two requests to find the difference.

Can I force the agent to execute a refused task?

No, and that's not desirable. The limits exist for good reasons. If you think the refusal is a mistake, reformulate your request or contact support to understand.

Does a refusal consume tokens?

Yes, because the agent analyzed your request before concluding it couldn't execute it. Consumption remains low, but it exists. Avoid repeating the same refused request.

How do I know if it's a refusal or a bug?

A refusal comes with an explanation. A bug produces a technical error or silence. If the agent works normally for other tasks, it's probably a justified refusal.