> ## Documentation Index > Fetch the complete documentation index at: https://docs.conduit.ai/llms.txt > Use this file to discover all available pages before exploring further. # Escalations and AI tags > What it means when the agent flags a conversation, where flagged conversations gather, and how to clear them. Sometimes the agent works out that it should not answer. Instead of guessing, it tags the conversation and leaves it for a person. ## The escalation tags Tags such as **AI needs your help** and **AI is missing information** appear on the conversation in the list and on the message itself. A tagged message is never sent automatically, even on autopilot. This is the difference between a tag and an ordinary proposed reply: a proposed reply is an answer the agent is confident in, whereas a tag is the agent saying it cannot answer safely. Common reasons for an escalation tag: * The knowledge base has nothing covering the question * Two sources disagree, so the agent has conflicting information * The situation needs a judgement call rather than an answer ## The Escalations view Every flagged conversation also appears under the **Escalations** view under Inbox. Rather than hunting through **All inboxes** for tags, you can work the Escalations view directly. These are the conversations that need a human regardless of how automated the rest of your inbox is. That tab does more than gather conversations. It groups recurring escalations into **issues**, each with an AI written explanation of the pattern behind them and specific recommendations for closing it, so you can fix the cause instead of answering the same question repeatedly. If your goal is raising automation rather than clearing today's queue, that is where to spend the time. See [The Escalations tab](/inbox/agent/escalations-tab). ## Your own conversation tags Everything above is about flags Conduit raises on its own. You can also define your own. Custom tags are created under **Settings > Conversation Tags**, where you give the tag a name, describe when the agent should apply it, and set an escalation level. That escalation level is the part that matters here: | Escalation level | What happens | | - | - | | **Not urgent** | The agent keeps replying. The tag is for filtering and prioritizing only | | **Urgent** | The agent stops responding and the conversation escalates to a person, even on autopilot | This is how you guarantee a particular situation always reaches a human. Rather than hoping the agent decides to escalate, you describe the situation once and mark it urgent. A guest reporting that supplies have run out at the property is a good candidate, since the useful answer depends on internal information and a person's judgement rather than anything the guest agent should be reciting. Urgent tagged conversations appear in the **Escalations** view alongside the AI flags above. See [Conversation Tags](/inbox/setup/conversation-tags) for creating and configuring them, and [Writing tag descriptions](/inbox/setup/conversation-tags#writing-tag-descriptions) for making them fire on the right conversations. ## Clearing an escalation Open the conversation and read the **Reasoning** tab inside a proposed reply. It tells you what the agent was missing, which decides what you do next. | What the reasoning shows | What to do | | - | - | | A fact the agent did not have | Add it to the knowledge base | | A situation it handled wrongly | Add or amend a skill | | Something that should always reach a person | Dismiss the escalation and answer it yourself | Dismissing is a legitimate outcome. Some conversations should escalate every time, and there is nothing to fix. Escalations are the cheapest agent feedback you will ever get, because each one is a real guest question your agent could not answer. Working through them in the first weeks is what moves an account from copilot towards autopilot. ## Frequently asked questions No. Flagged messages are held back regardless of mode. Yes. They stay in the To do list of the inbox they belong to. Escalations is a gathered view of the same conversations, not a separate queue. No. Fix the ones caused by a gap you can close. Leave the ones that genuinely need a person, and consider setting a guardrail so the agent keeps escalating them. See [Guardrails](/build-your-agent/what-goes-where/guardrails). Rather than judging escalations one at a time, work from the [Escalations tab](/inbox/agent/escalations-tab). It groups them into recurring issues ranked by how many conversations sit behind each one, which tells you which gaps are actually costing you automation and which are one-offs not worth chasing. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.