--- title: "How to Answer Security Questionnaires With AI" url: https://www.insulin.dev/blog/answer-security-questionnaires-with-ai/ canonical: https://www.insulin.dev/blog/answer-security-questionnaires-with-ai/ type: Blog description: "Answer security questionnaires with AI in Insulin: an agent drafts each answer from your approved knowledge base and cites its source; a person reviews it." --- # How to Answer Security Questionnaires With AI > Canonical HTML version: https://www.insulin.dev/blog/answer-security-questionnaires-with-ai/ 1. [Home](/) 2. [Blog](/blog/) 3. How to Answer Security Questionnaires With AI # How to Answer Security Questionnaires With AI An Insulin agent can draft questionnaire answers from your policies and approved answers, naming the document behind each one. A person still reviews each draft and copies it in, and nothing is kept for the next questionnaire until someone explicitly saves it. ![Zee Chen](/authors/zee-chen.jpg) Zee Chen Oct 15, 2026 ![How to Answer Security Questionnaires With AI](/images/blog/answer-security-questionnaires-with-ai/hero.png) Explore AI Summary [![](/logos/company/openai.svg)](https://chat.openai.com/?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fanswer-security-questionnaires-with-ai%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Knowledge%20Bases%2C%20Agents. "Summarize with ChatGPT")[![](/logos/company/anthropic.svg) ](https://claude.ai/new?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fanswer-security-questionnaires-with-ai%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Knowledge%20Bases%2C%20Agents. "Summarize with Claude")[![](/logos/company/gemini.svg)](https://www.google.com/search?udm=50&aep=11&q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fanswer-security-questionnaires-with-ai%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Knowledge%20Bases%2C%20Agents. "Summarize with Gemini")[](https://www.perplexity.ai/search/new?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fanswer-security-questionnaires-with-ai%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Knowledge%20Bases%2C%20Agents. "Summarize with Perplexity") Table of Contents - [How do you answer security questionnaires with AI?](#how-do-you-answer-security-questionnaires-with-ai) - [What goes in the answers knowledge base?](#what-goes-in-the-answers-knowledge-base) - [How do you set up the drafting agent?](#how-do-you-set-up-the-drafting-agent) - [How does the agent draft each answer?](#how-does-the-agent-draft-each-answer) - [What should a reviewer check before an answer goes in?](#what-should-a-reviewer-check-before-an-answer-goes-in) - [How do approved answers get reused next time?](#how-do-approved-answers-get-reused-next-time) - [Where won’t AI help with a security questionnaire?](#where-wont-ai-help-with-a-security-questionnaire) - [Frequently asked questions](#frequently-asked-questions) - [Takeaways](#takeaways) _Answering security questionnaires with AI means an agent drafts each answer from your own policies and approved answers, names the document behind it, and a person reviews every draft before it reaches the customer. In Insulin the agent drafts and cites, a person copies each approved answer into the questionnaire, and that answer is kept for next time only when someone explicitly saves it._ * * * “Can an AI agent draft our security questionnaire answers from our own policies, and show where each one came from?” Sales engineers, security leads and deal desks ask it with a long spreadsheet standing between them and a signature. Yes, within limits. The agent reads the questionnaire, searches what your security team has approved and cites the document behind each draft. It can’t decide what’s true about your security program, and Insulin’s documentation describes no way for it to write answers back into the questionnaire. What follows is for answering a questionnaire a customer sent you, not for questioning an AI vendor’s own security. ## How do you answer security questionnaires with AI? **To answer security questionnaires with AI, keep your approved answers and the policies behind them in a knowledge base, attach the questionnaire to an agent that drafts each answer from those documents and cites its source, and have a person review every draft before copying it in.** In Insulin that’s five steps, each with an edge worth knowing before you start: Step Who does it What happens The limit **1\. Build an answers knowledge base** Security, once; approvers after that Approved answers and the policies behind them are indexed for search PDF, DOCX and text formats, up to 25 MB a file; no Excel, so save past questionnaires as CSV **2\. Attach the questionnaire** Whoever is answering it The drafting agent reads the Excel, Word or PDF file in the chat For a PDF, what it reads depends on the model, and some models can’t read a scanned one **3\. Draft each answer** The drafting agent It searches up to three knowledge bases and cites the source document for each result A question no document covers has nothing to cite **4\. Review every answer and copy it in** A person Each draft is checked against its source, then pasted into the questionnaire No documented write-back to the questionnaire file or portal **5\. Keep it current** An approver After approval, explicitly ask an Edit-mode curator to create, update, or deprecate a KB file It never deletes; a deprecated file can be restored The two agents are the design: the one that reads a customer’s file can’t change your approved answers, and the one that can is used only by approvers, one explicit request at a time. ## What goes in the answers knowledge base? **An answers knowledge base is the knowledge base your drafting agent searches: the answers your security team has already approved, plus the policies those answers rest on.** Build it as an Organization knowledge base, which takes org admin access, so it can be shared with everyone who answers questionnaires. If you haven’t built one before, [how to ground an AI agent in your documents](/blog/how-to-ground-an-ai-agent-in-your-documents/) covers the basics. An agent can search up to three knowledge bases at query time, so split the material by who maintains it, into three at most: Knowledge base What goes in it How it stays current Mode on the drafting agent **Approved answers** One file per approved answer, with its approver and approval date at the top; past questionnaires saved as CSV Approvers, through the curator agent Read **Security policies** The policies and procedures your answers draw on Policy owners upload revisions, or the knowledge base syncs a Google Drive folder Read **Public security pages** (optional) Your published security pages, crawled from their URL A manual **Re-sync**, because websites don’t re-crawl on their own Read [Files](https://doc.fours.com/insulin/knowledge-base/#files) can be PDF, DOCX or text formats such as `.csv` and `.md`, up to 25 MB each, but not Excel: save a past questionnaire as CSV, keeping only the answers you’d give again. Keep approved answers as files, one per answer: Deprecate works only on files, not on website or connector documents, and one answer per file lets you retire a superseded answer on its own, while a citation names the exact answer a draft used. Before attaching it to an agent, run **Test search** with questions from a recent questionnaire. It previews what an agent would retrieve, with each result’s source document and relevance score, so you can see which questions nothing answers yet. Then share. Access to an organization knowledge base [comes only from a per-knowledge-base share or org-wide sharing](https://doc.fours.com/insulin/knowledge-base/#sharing-and-roles), never from an org role by itself. Share each one as User, which searches and reads, with everyone who answers questionnaires. Give Editor, the role that adds content, only to whoever maintains each one: approvers on the approved answers, policy owners on the policies. ## How do you set up the drafting agent? **A drafting agent is a custom agent whose only job is turning a questionnaire into cited drafts: an Organization agent with the knowledge bases above in Read mode, no integrations, and a system prompt that asks for a source or a flag on every question.** Only org admins can create an [Organization agent](https://doc.fours.com/insulin/agents/#user-level-vs-org-level-agents), the only kind that can be shared. 1. **Attach the knowledge bases in Read mode,** which [retrieves and searches](https://doc.fours.com/insulin/agents/#integrations-and-knowledge-bases) a knowledge base “for grounded context only.” The agent reading a customer’s file has no use for Edit. 2. **Add no integrations.** It shouldn’t hold a connection that text inside someone else’s file could talk it into using. 3. **If questionnaires arrive as PDFs, pick a default model that reads them.** An Organization agent can use the providers your organization has connected, and Anthropic, OpenAI and Gemini models “read a PDF itself, layout, images and charts included.” 4. **Share it by name, as USER,** with the people who answer questionnaires; USER uses the agent without changing it. Sharing the agent doesn’t share its knowledge bases, so they need User on those too. Then give it a system prompt that allows two outcomes per question, a cited draft or a flag: ``` You draft answers to security questionnaires for a person to review. You never send or submit anything. For each question in the attached questionnaire: 1. Search the approved answers first, then the policies. 2. Draft a short answer that says only what those documents say, and name the source document. 3. If no document covers the question, write NO SOURCE instead of an answer. 4. If two documents disagree, quote both with their names. Don't choose. Never state a certification, date, number or commitment that a source document doesn't state. ``` ## How does the agent draft each answer? **You attach the questionnaire to a chat with the drafting agent and ask for one section at a time, a batch one review pass can check; the agent reads the file, searches the knowledge bases for each question, and returns a draft naming its source document, or a flag.** Attach it with the paperclip, by dragging it onto the composer or by pasting it: [up to 10 files a message](https://doc.fours.com/insulin/agents/#file-attachments), each up to 25 MB. The agent reads PDFs, Word (`.docx`, `.doc`), Excel (`.xlsx`, `.xls`), PowerPoint (`.pptx`) and text files; a file in another format still uploads, but the agent is told it can’t read it. Given a choice, take Excel or Word, because what the agent gets from a PDF depends on the model the turn runs on. Ask for each section in a fixed shape: the questionnaire’s own question number, the draft, and its source document, or NO SOURCE. At query time an agent “can search up to three knowledge bases and cites the source document for each result,” so each draft arrives with the document a reviewer needs to open. To see exactly what the agent retrieved, expand its search in the conversation: tool call details show their inputs and outputs. ## What should a reviewer check before an answer goes in? **That the cited document exists, is current, and says what the draft says, and no more.** A citation turns review into a source check: instead of judging whether a paragraph sounds right, the reviewer opens the document it names and compares. [How to read an AI citation properly](/blog/why-ai-answers-need-citations/) covers what a source does and doesn’t prove. For each draft: 1. **A source is named, and you’d still stand behind it.** A NO SOURCE flag, or a document you’d retire, sends the question to its owner, not into the questionnaire. 2. **The draft says no more than the source.** Watch for “all,” “always” and “never,” and for any certification, date or number the document doesn’t state. 3. **It answers every part of the question.** A two-part question with a one-part answer reads as evasive. 4. **It’s fit to send.** Policies are written for staff, and an accurate sentence isn’t always one to share with a customer. 5. **You copy it in.** Paste each approved answer into the questionnaire yourself. **There’s no documented write-back.** The agent reads the attached questionnaire, and Insulin’s documentation describes no way for it to write answers back into that file or into the portal a customer sent it from. The person who copies each answer across is also the last check before it leaves. ## How do approved answers get reused next time? **Only when someone saves them: after approval, explicitly ask an Edit-mode curator to create, update, or deprecate a KB file.** The drafting agent can search the knowledge bases but can’t change them, so nothing a reviewer approves or edits reaches the approved answers by itself. **A curator agent is a second custom agent that maintains the approved answers and does nothing else.** Build it as an Organization agent with the approved-answers knowledge base in Edit mode, which can “create, update, and deprecate files in the knowledge base (never delete).” Leave the policies and integrations off it, share it by name only with the approvers who hold Editor on that knowledge base, and give it a one-line system prompt: save exactly the text an approver gives you, under the file name they give, and nothing else. Each change is one explicit request: When Ask the curator What happens A new answer is approved ”Create `data-retention.md` with this answer, approved by \[approver\] on \[date\]: …” A new file, searchable once it’s indexed An approved answer changes ”Update `encryption-at-rest.md` to this text: …” Updated; searches see the new text once it’s indexed An answer is superseded ”Deprecate `vpn-access.md`.” Out of search but kept, and restorable A deprecated file is [retired from search but not deleted](https://doc.fours.com/insulin/knowledge-base/#retiring-a-document): its chunks are kept, and **Restore** brings it back, re-indexing it on the way. Deprecate is offered only once a file has finished indexing, and deleting a file stays with people. [When to deprecate, replace or delete a knowledge-base document](/blog/deprecate-replace-or-delete-a-knowledge-base-document/) compares the three. Ask by file name, so a request can’t land on the wrong answer, and expand the tool call afterwards to see exactly what was saved: a draft saved by mistake becomes a source the drafting agent will cite. Edit mode covers files in the knowledge base, not the questionnaire, so keep customer files out of the curator’s chat entirely. ## Where won’t AI help with a security questionnaire? **Where there’s nothing to read, nothing to cite or nowhere to write.** Each is a point where a person takes over. Situation What happens What to do **A PDF, on a Fours-hosted or open-source model** Those models get only its text, up to 50,000 characters: a scanned PDF has none, so the model is told it couldn’t read the file, and a long one can lose its later questions Set an Anthropic, OpenAI or Gemini default model, or ask the customer for Excel or Word **A reply that failed over** The reply ends with an [italic note naming both models](https://doc.fours.com/insulin/agents/#when-a-model-fails-over) If the second is a Fours-hosted or open-source model and the questionnaire is a PDF, check that every question was drafted **A question no document covers** There’s nothing to cite, so the system prompt has the agent write NO SOURCE Route it to its owner, then save the approved answer through the curator **Filling in the questionnaire or portal** No write-back is documented A person copies each approved answer in **A request for evidence**, such as an audit report The agent can name a document, but deciding what to share isn’t drafting A person decides, under your own sharing rules ## Frequently asked questions ### Can AI fill out a security questionnaire for you? It can draft the answers, and a person fills in the file. An Insulin agent reads the attached questionnaire and drafts each answer from your knowledge bases, naming its source. Insulin’s documentation describes no write-back to the questionnaire file or portal. ### Which questionnaire formats can an Insulin agent read? Excel (.xlsx, .xls), Word (.docx, .doc), PDF, PowerPoint (.pptx) and text files, up to 25 MB each. For a PDF, the model matters: Fours-hosted and open-source models get only its text, so they can’t read a scanned PDF. ### Does the AI show where each answer came from? Yes, when the answer comes from a knowledge base. An agent can search up to three knowledge bases at query time and cites the source document for each result. A question no document covers has nothing to cite, so have the agent flag it rather than guess. ### Are approved answers saved for the next questionnaire? Only when someone saves them. After approval, ask a curator agent with Edit mode to create or update a knowledge-base file. The drafting agent has the knowledge bases in Read mode, so it can search them but can’t change them. ### Why use two agents instead of one? So the agent that reads a customer’s file can’t change your approved answers. The drafting agent has the knowledge bases in Read mode and no integrations. The curator has Edit mode on the approved answers, and approvers ask it for each change. ## Takeaways - **The agent drafts; a person answers.** Every draft names its source document, so review is a source check, and nothing reaches the customer until a reviewer copies it in. - **There’s no documented write-back.** The agent reads the questionnaire; filling in the file or the portal is a person’s job. - **Two agents, two modes.** The drafting agent only reads. After approval, explicitly ask the Edit-mode curator to create, update or deprecate a knowledge-base file: nothing is kept for next time by itself, and nothing is deleted. - **Prefer Excel or Word.** For a PDF, what the agent reads depends on the model, and a scanned one needs an Anthropic, OpenAI or Gemini model. Every step here rests on one thing: a knowledge base of answers your team has approved, searched and cited on every draft. To see how that works with your own documents, start with [knowledge bases in Insulin](/knowledge-bases/). ## Sources Primary sources for the platform rules cited above. Last verified October 15, 2026. Cloud providers change fees, eligibility, and program terms without notice — check the source before relying on a figure. - [Insulin Knowledge Bases — Fours Doc](https://doc.fours.com/insulin/knowledge-base/) — Files: PDF, DOCX and text formats such as .csv and .md, up to 25 MB per file; creating an organization knowledge base requiring org admin access; only documents that finish indexing becoming searchable; Connectors: organization knowledge bases supporting the Google Drive connector, and websites re-crawled only on a manual Re-sync; Retiring a document: Deprecate dropping an indexed file from search while keeping it, offered only once a file has finished indexing, and Restore bringing it back and re-indexing it, files only; Testing Search: the matching content, its source document and a relevance score; an agent searching up to three knowledge bases at query time and citing the source document for each result; Sharing and Roles: access to an organization knowledge base only from a per-knowledge-base share or org-wide sharing, never from an org role by itself, Editor adding content and User limited to search and read - [Insulin Agents — Fours Doc](https://doc.fours.com/insulin/agents/) — Only org admins creating org-level agents, and only Organization agents shareable; an Organization agent drawing on the organization's connected providers; a knowledge base attached in Read mode for grounded context only, or in Edit mode to create, update and deprecate files, never delete; integrations as an optional allowlist; USER using the agent in conversations; File Attachments: up to 10 files per message and 25 MB per file, the agent reading PDFs, Word (.docx, .doc), Excel (.xlsx, .xls), PowerPoint (.pptx) and text files, a file in another format uploading but unreadable, and Anthropic, OpenAI and Gemini models reading a PDF itself while Fours-hosted and open-source models get only its text, up to 50,000 characters, a scanned PDF having none; tool call inputs and outputs expandable; failover surfaced as an italic note naming both models ## Keep reading - [AgentsEnterprise AI Agent Memory: What Should Agents Remember—and Forget?Aug 20, 2026](/blog/enterprise-ai-agent-memory/) - [AgentsHow to Build a Grounded AI Agent for Marketplace OpsAug 18, 2026](/blog/how-to-build-a-grounded-agent-marketplace-ops/) - [Knowledge BasesHow to Ground an AI Agent in Your DocumentsAug 2, 2026](/blog/how-to-ground-an-ai-agent-in-your-documents/) - [IntegrationsAI Agent for Stripe: Payouts, Invoices, TaxOct 15, 2026](/blog/ai-agent-for-stripe/) [Browse every post on the Insulin Blog](/blog/) ### Stay Updated New posts, product updates and marketplace strategy are shared on LinkedIn as they publish. [Follow Fours on LinkedIn](https://www.linkedin.com/company/suger-inc)