{"site":{"name":"Koji","description":"AI-native customer research platform that helps teams conduct, analyze, and synthesize customer interviews at scale.","url":"https://www.koji.so","contentTypes":["blog","documentation"],"lastUpdated":"2026-10-12T18:27:46.636Z"},"content":[{"type":"documentation","id":"64c1f5a2-29cf-4da3-897f-5f703f3f7825","slug":"setting-interview-guardrails","title":"Interview Guardrails: Set What Your AI Interviewer Avoids","url":"https://www.koji.so/docs/setting-interview-guardrails","summary":"Guardrails are the rules for what a Koji AI interviewer should avoid. They live in the Studio on the Questions tab, in the Guardrails section below Closing Approach, written one per line. Changes must be saved and the study republished before new interviews follow them, and the same guardrails apply to text, voice and phone interviews. The AI Assistant can add, suggest or remove guardrails on request, keeping existing ones, and saves straight away; the MCP connection accepts up to 10 guardrails of up to 300 characters. With no guardrails, the interviewer follows built-in good-practice rules for the research method. Guardrails are available on every plan.","content":"Guardrails are the list of things your Koji AI interviewer should avoid during an interview: topics it mustn't raise, promises it mustn't make, details it mustn't share. You set them in the **Studio**, on the **Questions** tab, in the **Guardrails** box at the bottom, one guardrail per line. Save your changes, then publish (or republish) the study, and every new interview follows them, whether it runs by text, by voice or as a phone call. Guardrails are available on every plan.\n\nThis guide shows where guardrails live, how to add and change them by hand or with the AI Assistant, and how to write ones the interviewer can actually follow.\n\n## Why guardrails matter\n\nA human moderator walks in with a briefing: don't mention the price change, don't name other clients, don't get drawn into a support ticket. An AI interviewer needs the same briefing, written down.\n\nWithout it, the interviewer will happily follow an interesting thread wherever it goes, which is usually what you want. Following up on what people say is the whole point of an AI-moderated interview, and it's what a static form from SurveyMonkey, Typeform or Qualtrics can't do. But some threads carry risk:\n\n- **Commercial risk.** A participant asks about pricing or a roadmap date, and an unscripted answer becomes a promise.\n- **Confidentiality.** Unreleased features, other customers' names, a deal that isn't public yet.\n- **Sensitive personal data.** Under the GDPR, data about health, religion, political opinions, sexual orientation and a few other areas is \"special category\" data (Article 9) with stricter rules. If your research doesn't need it, tell the interviewer not to ask for it.\n- **Research quality.** An interviewer that offers opinions or suggests answers biases what you hear.\n\nGuardrails let you keep the follow-up depth that makes AI interviews useful while fencing off the areas you never want to enter.\n\n## Where to find guardrails\n\n1. Open your study and go to **Studio**.\n2. Select the **Questions** tab.\n3. Scroll to the last section, **Guardrails**. Hover the help icon and you'll see \"Things the interviewer should avoid\".\n\nThe Questions tab runs in interview order, so guardrails sit below the parts of the conversation they constrain: **Screening Questions** (when your study has them), **Opening Approach**, **Interview Questions**, **Topics to Explore** and **Closing Approach**.\n\n![Koji guardrails settings listing topics the AI interviewer should avoid, like prices and other customers](/email/shots/shot-guardrails.jpg \"Each line is one rule the interviewer follows in every interview\")\n\n## Add or edit guardrails by hand\n\nThe Guardrails box is a bulleted list:\n\n1. Click into the box (it starts empty with the placeholder \"Enter a guardrail...\").\n2. Type your first guardrail.\n3. Press **Enter** to start a new bullet for the next one.\n4. To remove a guardrail, delete its text. Pressing **Backspace** at the start of a line removes that whole line.\n5. Click **Save** in the **Unsaved changes** bar. **Discard** throws away your edits.\n6. Click **Publish** (or **Republish** on a live study) so new interviews pick up the change.\n\nThere's no fixed limit on how many guardrails you can add in the Studio, but a short, specific list works better than a long one. If you publish with unsaved edits, Koji warns you that publishing will use the last saved version and offers **Save & Publish**.\n\nArchived studies are read-only, so their guardrails can't be edited.\n\n![Process for setting interview guardrails in Koji: open the Questions tab, write one rule per line, save, republish, then preview to test.](/visuals/setting-ai-interview-guardrails-process/setting-ai-interview-guardrails-process.png \"Edits only reach participants after you save and republish\")\n\n## Let the AI Assistant draft or change them\n\nYou don't have to write guardrails yourself. With the **AI Assistant** open beside your brief, ask in plain language:\n\n- \"Add a guardrail: don't quote any prices or discounts.\"\n- \"Suggest guardrails for interviews with enterprise customers.\"\n- \"Drop the guardrail about the beta.\"\n\nThe assistant keeps your existing guardrails and changes only what you asked for. The chat shows a step such as \"Updated guardrails\" so you can see what changed. It won't touch your guardrails unless you ask, and it may suggest a few when it first drafts a study.\n\nChanges the assistant makes are saved to the brief straight away. You still need to **Republish** for live interviews to follow them. To clear every guardrail at once, delete them in the box yourself.\n\nA useful habit: if you want the interview to *start* or *end* differently, ask the assistant to rewrite the **Opening Approach** or **Closing Approach** rather than adding a guardrail. Guardrails are for what to avoid; the opening and closing say what to do.\n\nThe **Analyse for Suggestions** button can also propose guardrail changes based on the rest of your brief, which you can review before applying. Studies started from a template often come with a few guardrails already filled in, so check them against your own situation.\n\nIf you build studies through the AI-assistant (MCP) connection, you can set guardrails there too: up to 10, of up to 300 characters each, and sending a list replaces the current one. See the [MCP overview](/docs/mcp-overview).\n\n## How guardrails reach your interviews\n\nWhen you publish, Koji passes your saved guardrails to the interviewer for every way people can respond: text interviews, voice interviews and phone calls all get the same rules. That has two consequences:\n\n- **Edits don't apply to live interviews until you republish.**\n- **Unsaved edits are never published.** Save first, or use **Save & Publish**.\n\nIf you leave Guardrails empty, the interviewer still follows built-in good-practice rules for your study's research method. Your own guardrails add to that with what only you know: your product, your customers and your boundaries.\n\n## Writing guardrails that work\n\nA guardrail is an instruction to an interviewer who has never met your company. The clearer it is, the more reliably it's followed.\n\n| Vague | Clear |\n|---|---|\n| Be careful with pricing | Don't quote prices, discounts or plan limits |\n| Keep it confidential | Don't mention the new mobile app; it isn't announced |\n| No sensitive stuff | Don't ask about health, religion or political views |\n| Stay neutral | Don't give opinions on the product or suggest answers |\n| Avoid support issues | Don't troubleshoot problems; say the team will follow up |\n\n![Comparison of vague and clear AI interviewer guardrails, from pricing and confidentiality to sensitive topics and neutrality.](/visuals/vague-vs-clear-ai-interview-guardrails/vague-vs-clear-ai-interview-guardrails.png \"Specific rules get followed; vague ones get interpreted\")\n\nA few rules of thumb:\n\n- **One rule per line.** Each bullet becomes a separate instruction.\n- **Name the thing.** \"Don't mention Project Atlas\" beats \"don't discuss unreleased work\", because the interviewer can't guess what's unreleased.\n- **Say what to do instead when it helps.** \"Don't troubleshoot; say the team will follow up\" gives the interviewer a graceful way out when a participant pushes.\n- **Don't use guardrails to remove questions.** If a topic shouldn't come up at all, take it out of **Interview Questions** and **Topics to Explore** first.\n- **Keep the list short.** Five sharp guardrails beat twenty overlapping ones.\n\n## Test your guardrails before launch\n\nAfter republishing, open the **Preview** tab on your study's **Interviews** page and take the interview yourself. Try to break your own rules: ask the interviewer what the product costs, or name a competitor, and see how it responds. Previews are free and never count as responses. See [Previewing your interview](/docs/previewing-your-interview).\n\nThen read the first few real transcripts. If the interviewer drifted somewhere you didn't expect, add a guardrail, save and republish. New interviews will follow it.\n\n## Guardrails and the rest of your brief\n\nGuardrails work alongside the other controls on the **Questions** tab:\n\n- **Interview Questions** set what's asked. Koji's structured question types (open-ended, scale, single choice, multiple choice, ranking and yes/no) give you measurable answers alongside the conversation. See the [structured questions guide](/docs/structured-questions-guide).\n- **Follow-ups** on each question set how deep the interviewer probes: **None (skip)** or 1 to 3 follow-ups, with optional custom follow-up guidance.\n- **Topics to Explore** list open areas where the interviewer can follow interesting threads.\n- **Opening Approach** and **Closing Approach** shape how the conversation starts and ends.\n\nThink of it this way: questions and topics point the interviewer toward what you want to learn, and guardrails mark the places it shouldn't go.\n\n## Related Resources\n\n- [Structured Questions Guide](/docs/structured-questions-guide)\n- [Understanding the Research Brief](/docs/understanding-the-research-brief)\n- [Working with the AI Consultant](/docs/working-with-the-ai-consultant)\n- [Editing the Brief Manually](/docs/editing-the-brief-manually)\n- [Previewing Your Interview](/docs/previewing-your-interview)\n- [Publishing Your Study](/docs/publishing-your-study)\n","category":"Reports & Analysis","lastModified":"2026-10-12T04:16:33.128099+00:00","metaTitle":"AI Interviewer Guardrails: Set Topics to Avoid | Koji Docs","metaDescription":"Keep your AI interviewer away from pricing, confidential plans and sensitive topics. Set guardrails in Koji for text, voice and phone interviews on any plan.","keywords":["ai interviewer guardrails","ai interview topics to avoid","control ai moderator","ai interview brand safety","ai interviewer rules"],"aiSummary":"Guardrails are the rules for what a Koji AI interviewer should avoid. They live in the Studio on the Questions tab, in the Guardrails section below Closing Approach, written one per line. Changes must be saved and the study republished before new interviews follow them, and the same guardrails apply to text, voice and phone interviews. The AI Assistant can add, suggest or remove guardrails on request, keeping existing ones, and saves straight away; the MCP connection accepts up to 10 guardrails of up to 300 characters. With no guardrails, the interviewer follows built-in good-practice rules for the research method. Guardrails are available on every plan.","aiPrerequisites":["A Koji account","A study you can edit"],"aiLearningOutcomes":["Find and edit the Guardrails section in the Studio","Apply guardrail changes to live interviews by republishing","Use the AI Assistant to draft or change guardrails","Write specific guardrails an AI interviewer can follow","Test guardrails with a free preview"],"aiDifficulty":"beginner","aiEstimatedTime":"8 minutes"}],"pagination":{"total":1,"returned":1,"offset":0}}