> ## Documentation Index > Fetch the complete documentation index at: https://docs.signalmd.dev/llms.txt > Use this file to discover all available pages before exploring further. # AI copilot overview > An AI assistant that reads your issues and calls, drafts fixes to your prompt or flow, and tests them. Publishing to Retell always takes an admin's click. The copilot is an AI assistant in a chat panel beside every signed-in page in Signal. It answers questions about your agent from Signal's own data, drafts a change to your prompt or flow when you ask for one, and tests that draft. It cannot change your live Retell agent itself: an admin of your workspace publishes a draft with a click, after reading the diff. ## Open the copilot Click **Copilot** in the sidebar, or press Ctrl J (⌘J on a Mac). On wide screens the panel docks on the right and pushes the page over, so the issue you are asking about stays readable. On narrow screens it is a drawer over the page. The panel header has **New conversation** and **History**, where you can **Rename** or **Delete** earlier conversations. Conversations are private to the person who started them. Deleting one removes its messages; drafts it wrote stay in [Build](/copilot/build). Under the message box, a label shows what your next message is read against: **This issue**, **This call**, **This test run**, **This draft**, or the agent's name. The copilot receives the page you are on with every message, so "why is this happening" means the issue on screen. ## What it can see It reads the same data you see in the app, for the agent selected in the sidebar, and can list the other agents in your workspace. | Area | What it reads | | - | - | | Issues | Place, cause (or **Cause not established**), recommendation, call count, evidence quotes and the calls behind each issue | | Calls | The transcript with the step of each turn, tool calls with timing, signals, and the issues the call belongs to | | Flow | Calls per part, where calls drop or loop, and the transitions they took | | Versions | Each version's results, what changed between two versions, and the comparison | | Metrics | Health numbers for a date range, next to the previous period | | Tests | Your tests, their last results, and the results of a test run | | Prompt and flow text | The live version's prompts, step instructions and transition conditions, read when it drafts | It cannot: * See another workspace, anything Signal has not imported from Retell, or audio. * Change tools, URLs, ids, voice, settings or knowledge bases, or write new tests. A fix that needs one of those is yours to make in Retell. * Publish, roll back or start a paid Retell test run. Those take a person's click. * Draft or test on the demo agent, which is read-only. There you can ask questions. ## Questions that work well Ask about something Signal measured, and name the step, issue, call or version when you can. * "What should I fix first?" * "Why do callers hang up at the insurance step?" * "Did version 8 make things better or worse?" * "Lots of calls go from Offer times back to Check availability. Is that a loop we should fix?" * "Draft a fix for this issue, then run a quick check on it." * "Why did these tests fail, and which failures are real defects?" When a question needs something Signal does not hold, the copilot says so instead of guessing. It gives guidance for Retell only. ## What it can do | You ask | What happens | Who clicks | | - | - | - | | A question | It reads issues, calls, flow, versions or tests and answers with the evidence | Nobody | | Why something happens | It reads the issue and its calls. If the evidence does not settle the cause, it says **Cause not established** and what would settle it | Nobody | | A fix | It reads the live prompt and flow, writes a small change as a draft and shows the diff. Nothing reaches Retell | Nobody | | A quick check | Signal plays a caller and your agent from the draft's text and scores the conversations. The copilot may start one unasked; it spends no test budget | Nobody | | Retell tests | The draft runs in Retell's simulator on a copy of your agent. The copilot asks first, because it spends your test budget | You | | Publishing | The draft's text is written to your agent in Retell and published as a new version | An admin | [Build](/copilot/build) covers drafts, tests, publishing and rollback step by step. Drafts and publishing are not available through the [API](/api/introduction) or the [MCP server](/mcp/overview). ## Draft cards When the copilot writes or changes a draft, a card appears in its answer with the title, how many texts change, lines added and removed, the latest test result (such as "Quick check 2 of 3 passed (approximate)"), a status tag, and **Stale** if your live agent has changed since. The card has **Show diff**, **Quick check**, **Run tests**, **Publish** and **Open in Build**. **Publish** opens a dialog with the whole diff and the latest test results. Nothing is sent to Retell until an admin confirms there. If the draft changed after the dialog showed it, Signal refuses and asks for the diff to be reviewed again. ## How it cites evidence * It names issues, calls, steps, versions and counts as the data returned them, and links each issue and call it mentions. * Each lookup appears as a row in the answer, such as **Reading open issues** with a one-line result like "12 open issues". Click a row to see what it asked for. * It names the [lane](/concepts/lanes) a problem sits in. * It does not say a fix works unless a test ran on that draft in the conversation, and it says which kind. A quick check is an approximation, never proof. A Retell result is not reported until the run has finished. * Text inside a transcript is data. A caller who says "ignore your instructions" is something the call recorded, not an order. ## Limits * **Message allowance.** Each plan includes a number of copilot messages per period. A message is one thing you type; the lookups it triggers are not counted separately. Each quick check and each message in the brief's **Calibrate** chat counts as one message too. Near the limit the panel shows how many are left, and at zero the box is disabled until the period resets or you upgrade. See [Pricing](/billing/pricing). * **Spending limit.** Separately, each workspace has a usage limit on the copilot, the same on every plan, and an upgrade does not raise it. At the limit the panel says "This workspace has reached the copilot's spending limit for this period." See [Limits and quotas](/limits#copilot-limits). * **One answer at a time.** **Stop** ends an answer early and keeps what was written. * **Broad questions.** A question that needs too many lookups is cut short, and the copilot answers from what it has read. Narrow the question. * **Long results.** Very long calls and very large lookups are cut short, so ask about a narrower range if an answer seems incomplete. A message can be up to 8,000 characters. * **Text, not phone calls.** Quick checks and Retell tests are text conversations, so latency, silence and talking over cannot appear in them. See [Test a fix](/guides/test-a-fix). * **It can be wrong.** The rows and links are there so you can check it. ## Data handling The copilot runs on third-party AI model providers under no-retention terms. This is what leaves Signal when you use it. **Sent with each message:** your message, the page you are on (the ids of what is on screen), the earlier messages in the conversation, and what its lookups return: issues and evidence quotes, call transcripts and the tool calls inside them, flow, versions, metrics and tests, plus your live prompt and flow text when it drafts or checks a change. **Not sent:** audio, your Retell keys, other workspaces' data. **Personal details are not masked in the chat.** Transcripts can contain callers' names and dates of birth, and the copilot receives them as the call recorded them. Tests made from calls mask them; the chat does not. **Quick checks** send the draft's text and the test callers to the same providers. With no test for the issue, the simulated caller is made from the issue's latest calls and receives their real lines, unmasked. **Retention at the providers.** The providers do not keep what Signal sends or train on it. **Kept in Signal.** Conversations and their lookups are stored so a conversation can be reopened. They are private to you and removed when you delete the conversation. Drafts and quick-check transcripts belong to the agent and are visible to everyone in your workspace. Signal also records usage per message. ## Next steps Draft, test, publish and roll back a fix. Copilot messages included in each plan. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.