Agents Lookup

Recipes / Security & compliance

Security Questionnaire First-Draft Agent with Cited Sources

An agent drafts security questionnaire answers from your own trust documents and approved answer library, cites every source, and marks gaps NEEDS OWNER for a responsible person to sign.

CategorySecurity & compliance
Steps9
Human sign-offs4

The problem

Small software and service companies get security questionnaires from prospects: short SIG-lite style forms, custom spreadsheets with a hundred rows, or portal forms. Each one takes hours, pulls an engineer or founder away from other work, and usually gets answered by copying from the last questionnaire, which spreads old or inconsistent answers. The real risk isn't only the time. It's answering wrong, such as claiming a control you don't have or a certification you haven't earned, in a document a customer may rely on in a contract. This recipe sets up an agent that drafts a first pass using only your own approved material: your public trust page, security policies you have published, and an answer library your security owner has signed off. Every answer cites its source. Anything not covered is marked NEEDS OWNER instead of guessed, and a responsible owner reviews and signs the final response.

Who it's for

Not a fit when

Tools

Google Sheets or ExcelWorking copy of the questionnaire with columns for draft answer, source, confidence, owner, and review status
Approved answer libraryA sheet or doc of reviewed question-and-answer pairs, each with owner, last-reviewed date, and source document
Google Drive or a shared folderStores the current versions of your public trust documents and approved policies the agent may cite
Claude or a similar LLMMatches each question to library entries and source documents, drafts a cited answer, or marks it NEEDS OWNER
Slack or emailSends NEEDS OWNER items to the right control owner and notifies the reviewer when a draft is ready
Document signing or approval logRecords who reviewed and approved the final response and when

Inputs you need first

Step-by-step build

  1. Set the sourcing rules

    Write these rules into the agent instructions and your team process. The agent may only use the approved answer library and the documents in the approved sources folder. Every answer must cite at least one source by document name and section, or by library entry ID. If no source covers the question, the answer is NEEDS OWNER with a short note on what is missing. The agent must never state or imply a certification, audit, or control that isn't in the sources. Answers are drafts until the responsible owner signs them.

  2. Build the approved answer library

    Create a sheet with columns: entry_id, topic, question_variants, approved_answer, source_document, owner, last_reviewed, scope_notes. Seed it from past questionnaires, but have each control owner review entries in their area before they count as approved. Remove anything uncertain. Use short factual answers and note limits, for example 'applies to production systems only'. Set a review interval, such as every six months, and mark entries older than that as stale so the agent treats them with lower confidence and flags them.

  3. Organize the approved sources folder

    Put current versions of your trust documents in one folder, each with a clear name and date, such as Access Control Policy v3 2026-05. Remove drafts and old versions, since the agent will cite whatever it is given. Only include material you are comfortable sharing with prospects or that is already public. If a policy is internal-only, cite it in the library entry rather than giving the agent the full text, and confirm with the owner that its summary can be shared.

  4. Prepare the questionnaire sheet

    Copy the incoming questionnaire into a working sheet and add columns: question_id, question_text, draft_answer, yes_no_na if the form uses it, source_citation, confidence (High, Medium, Low), status (Draft, NEEDS OWNER, Reviewed, Approved), owner, reviewer_notes. For portal-based questionnaires, export or copy the questions first. Keep the original file untouched so the final answers can be pasted back in the customer's exact format.

  5. Run the first draft

    Send questions to the agent in batches of 20 to 40 along with the library and sources. For each question, it looks for a matching library entry first, then a source document passage. It writes a draft answer that stays within what the source says, cites the source, and sets confidence. High means a library entry directly answers it. Medium means a source document covers it but the wording needed adapting. Low or no match means NEEDS OWNER. The agent never fills a yes or no field without a source.

  6. Route NEEDS OWNER items

    Filter the sheet for NEEDS OWNER and Low confidence rows. Group them by control area and send each group to its owner through Slack or email with the question text, any partial sources found, and the deadline. Owners answer in the sheet, attach or name the evidence, and decide whether the answer should be added to the library. If the honest answer is no or not yet, write that plainly, with any compensating control or planned date the owner approves.

  7. Test on a past questionnaire

    Before using it on a live deal, run the workflow on a questionnaire your team already answered and approved. Compare draft answers with the final approved ones. Look for three problems: answers that go beyond the source, missing citations, and questions marked answered that should have been NEEDS OWNER. Any answer that claims more than the source supports is a serious failure; tighten the prompt and library until it doesn't happen in the test set.

  8. Review, sign, and submit

    The responsible owner reads every answer, not just the flagged ones, checks citations against the sources, and edits where needed. They pay close attention to yes or no answers, certifications, encryption and data location statements, incident notification timelines, and anything that could become a contract commitment. When done, they set each row to Approved, record their name and date in the approval log, and only then is the response pasted into the customer's format and sent by the sales or security contact.

  9. Feed the library and maintain it

    After each questionnaire, add new owner-approved answers to the library with source and review date. Every quarter, check for stale entries, policy changes, new subprocessors, or controls that changed, and update the sources folder and library together. When a policy is retired, remove it from the folder the same day so it can't be cited. Keep a log of which library version was used for each customer response in case a question comes up later.

Example agent instructions

Adapt this to your business. Placeholders in braces are filled by your automation tool.

You draft first-pass answers to a security questionnaire for {company_name}. Your drafts are reviewed and signed by {responsible_owner} before anything is shared.

You may use only these sources:
- Approved answer library: {answer_library}
- Approved source documents: {source_documents}

Rules:
1. Every answer must cite at least one source as [Library: entry_id] or [Doc: document name, section].
2. Say only what the source supports. Do not broaden scope, add detail, or upgrade 'planned' to 'in place'.
3. Never state or imply a certification, audit report, attestation, or control unless a source explicitly confirms it with a date.
4. If no source answers the question, set the answer to NEEDS OWNER and say in one sentence what information is missing.
5. For yes/no fields, answer only if a source directly supports the answer; otherwise NEEDS OWNER.
6. If a library entry is marked stale, use it but set confidence to Low and add a note.
7. Keep answers short, factual, and in plain language. No marketing language.

Questions:
{questions}

Return JSON for each question: question_id, draft_answer, yes_no_na, source_citation, confidence (High, Medium, Low), notes.

Human sign-off points

Guardrails

Expected time saved

Estimate: if a typical questionnaire has about 120 questions and your team currently spends about 3 minutes per question finding and writing answers, a cited first draft that covers most routine questions could reduce that to a review pass plus focused work on NEEDS OWNER items, possibly saving 3 to 5 hours per questionnaire. This assumes a well-maintained answer library. Your results depend on volume and data quality.

How to tell it's working

FAQ

Can the agent help us answer questions about controls we don't have yet?

It can mark them clearly as NEEDS OWNER and tell you what is missing, which helps you see gaps. It shouldn't write an answer that suggests a control exists when it doesn't. The honest answer, such as 'not currently in place' with any compensating control or planned date your owner approves, is better than an overstated one that could become a contract problem later.

What if the customer asks for evidence like policies or audit reports?

Evidence requests are handled by people. The agent can note that a source exists and which document it is, but the security owner decides what to share, under what NDA or terms, and sends it through your normal channel. Don't place confidential reports in the agent's sources folder unless your owner has approved that handling and your tool settings meet your data policies.

Is it safe to put our security documents into an AI tool?

Check the data handling terms of the AI tool you use, including whether inputs are retained or used for training, and choose a business plan or settings that fit your policy. Start with documents that are already public, like your trust page and privacy policy. Add internal summaries only after your security owner reviews the vendor terms and approves it.

Want this set up for you?

Build it yourself with the steps above, or tell us what you need. We scope the work with you and send a written quote before anything starts.

Request a quote

Sources

Written by Agents Lookup. Official pages for the tools this recipe uses: