The stack
The problem
Security questionnaires and RFPs are deal-blockers that quietly eat days right when momentum matters most. A 200-question security review lands late in a hot cycle, the AE pings engineering and security, answers trickle in over a week between everyone's real jobs, and the deal stalls precisely when it should be accelerating. Worse, the same answers get re-typed from scratch for every single deal because nobody maintains a clean, current answer library, so the team pays the full cost every time.
Here is the thing almost nobody internalizes: nearly every answer already exists somewhere in your own documents. Your SOC 2 report, your security policy, your DPA, your architecture docs, and your previously completed and approved questionnaires already contain the truth. The work is not figuring out the answer; it is finding the right approved language and rephrasing it to fit each new questionnaire's specific wording and format. That is retrieval and rephrasing, which is exactly what AI is good at, as long as it draws only from your real documents and never invents a control you do not actually have.
Pointing an AI at your own approved source documents lets it first-draft each answer in the new questionnaire's format, cite which document the answer came from, and flag anything the docs cannot answer for a human to handle. Security still reviews everything, but they review and approve drafts in hours instead of writing from a blank page over days. The flagged-and-low-confidence rows tell expert reviewers exactly where to spend their limited time.
The opinion that must govern this entire workflow, because the downside is severe: this is built around source-only answering and mandatory human sign-off, full stop. A hallucinated security claim, an AI confidently asserting a certification or a control you do not have, is not a minor error. It is a compliance violation, a breach of contract waiting to happen, and a trust catastrophe with a customer's security team. The prompt is engineered to refuse rather than guess, and a human is accountable for every answer that gets submitted. There is no version of this that auto-submits.
How it works
- Assemble your approved, current security and compliance docs into one source-of-truth library
- Load the source docs into a Claude Project (persistent context) or ChatGPT file upload
- Get the incoming questionnaire into a clean, structured list with the expected answer format per question
- Have AI draft each answer using ONLY your source docs, citing the source and a confidence level per answer
- Route every 'NEEDS HUMAN INPUT' and low-confidence answer to the right expert; the rest just need sign-off
- Security or the accountable owner reviews and approves every answer before it goes back, then it feeds the library
See it run
The playbook
Build the approved source-of-truth library
Gather your real, approved compliance and security materials into one folder: the SOC 2 or ISO report or summary, your security and data-handling policies, the DPA, the sub-processor list, architecture and data-flow descriptions, the business continuity and incident response plans, and any previously completed and approved questionnaires. Those prior approved questionnaires are gold, because they are pre-vetted language that security has already signed off on, which means re-using them is low-risk and fast.
Include only approved, current documents. Anything outdated or unapproved sitting in the library becomes a confidently wrong answer later, and a wrong security answer is the exact failure mode you cannot afford. Treat the library as a maintained single source of truth, not a junk drawer; assign someone to keep it current.
Note the sensitivity tier of each document. Some answers (a SOC 2 report itself) only go out under NDA, and the workflow should know to reference availability under NDA rather than pasting confidential contents into a questionnaire.
TipYour single best source is last quarter's completed and approved questionnaire. Those answers are reusable, already-signed-off language, so the AI rephrasing them carries far less risk than the AI interpreting a raw policy document for the first time.
Load the source docs into the AI
Upload the source documents into a Claude Project. In Claude, create a new Project from the left sidebar, give it a name like 'Security Questionnaire Drafting,' and add your approved documents as Project knowledge. The advantage of a Project is that every drafting session automatically has the full approved library in context without re-uploading, which is ideal for a recurring task. ChatGPT's file upload or a custom GPT works similarly.
If the documents are large and you are hitting context limits, prioritize the highest-yield sources: the security policy, the SOC 2 summary, the DPA, and the most recent completed questionnaire. Those four answer the large majority of standard questions.
Put the source-only-answering instruction into the Project's custom instructions so it applies to every session automatically, rather than relying on remembering to paste it each time. The guardrail should be impossible to forget.
Structure the incoming questionnaire
Most questionnaires arrive as a spreadsheet or a portal. Get the questions into a clean list, with a Google Sheet of one question per row working well. For each question, note the expected answer format: yes/no, free text, or pick-from-a-list. Clean structure going in means clean, mappable answers coming out that you can paste straight back into the customer's grid without reformatting.
If the questionnaire is a PDF or a locked portal, copy the questions into your sheet so you have a workable list to feed the AI in batches. The few minutes of structuring pays for itself many times over in clean output.
Add empty columns to your sheet for the draft answer, the source citation, and the confidence level. This is the grid the AI will fill and the human will review, and having it laid out before you start keeps the whole process organized.
Draft answers grounded strictly in your docs
Run the drafting prompt in your Claude Project. The non-negotiable instructions are that the AI answers ONLY from the provided source documents, cites which document each answer came from, and explicitly outputs a flag for any question the docs do not cover instead of guessing. A confident wrong answer to 'Do you encrypt data at rest?' is a legal and trust catastrophe, so the prompt is deliberately engineered to make the model refuse and escalate rather than invent.
Process in batches of fifteen to twenty questions. Smaller batches let you review as you go, keep the source documents in clear focus for the model, and make it obvious if the model starts drifting toward guessing on a run. Larger batches tend to degrade and are harder to audit.
Demand a confidence level per answer. The high-confidence, well-sourced answers usually just need a fast sign-off; the low-confidence and flagged ones are precisely where expert time should go. That sorting is most of the time savings.
You are drafting answers to a security/RFP questionnaire using ONLY the approved company documents available to you in this Project. This is compliance-sensitive and accuracy is critical.
ABSOLUTE RULES:
- Answer ONLY from the provided documents. NEVER invent, assume, or infer a control, certification, or practice that is not explicitly supported.
- For each answer, cite the source document and section, e.g. [Source: Security Policy, sec 4.2].
- If the documents do NOT clearly and explicitly answer a question, output exactly: "NEEDS HUMAN INPUT - not covered in source docs" and DO NOT guess or hedge into an answer.
- For confidential sources (e.g. SOC 2 report contents), reference availability under NDA rather than pasting confidential detail.
- Match the requested answer format (yes/no vs free text vs pick-from-list).
QUESTIONS (batch):
{{PASTE_15-20_QUESTIONS_WITH_EXPECTED_FORMAT}}
Output a table, one row per question:
Question # | Draft answer | Source | Confidence (high / med / low)TipSort the output by the confidence column. The 'NEEDS HUMAN INPUT' and low-confidence rows are exactly where security should spend their time; the high-confidence sourced rows are a fast approval pass, not a writing task.
Route flagged questions to the right humans
Pull every 'NEEDS HUMAN INPUT' and low-confidence answer into a short list and send each to the right owner: security, legal, or engineering depending on the question. This is dramatically less work than answering the whole questionnaire and focuses scarce expert time only where it is genuinely needed. On a typical standard questionnaire, the flagged set is a small fraction of the total, which is the entire point of the approach.
When an expert provides a genuinely new approved answer, add it back into the source library (often as an addition to your maintained answer document or your next 'master questionnaire'). Next time, that question drafts automatically at high confidence, so the library compounds and each questionnaire makes the next one faster.
Resist the temptation to let the requesting rep guess at a flagged answer to save time. The flag exists precisely because the answer is not established; a rep's guess on a security control is the same hallucination risk you engineered the AI to avoid, just from a different source.
Mandatory review and final assembly
Have security or the accountable owner review and approve every answer, not only the flagged ones. The AI produces drafts; a human being is accountable for what is actually submitted to a customer. The high-confidence sourced answers should be a quick read-and-approve, but they still get read, because an approved-but-wrong answer is still wrong and still your liability.
Once approved, paste the answers back into the customer's grid or portal in the exact format requested. Keep the completed, approved version in your library to compound the speed advantage on the next deal, and update any document the process revealed to be out of date.
Treat the human sign-off as the load-bearing control of the whole system, not a formality. The AI made the work fast; the human review is what makes it safe. Skipping it to save an hour is how a deal-accelerating tool becomes a compliance incident.
What you get
Drafted, sourced questionnaire answers with confidence flags and explicit human-input markers, ready for security review.
Q# | Draft answer | Source | Confidence
Q12 | Do you encrypt data at rest? | Yes. All customer data is encrypted at rest using AES-256. | [Source: Security Policy, sec 3.1] | high
Q13 | Do you hold SOC 2 Type II? | Yes. A SOC 2 Type II report is available under NDA; the most recent audit period covers the trailing 12 months. | [Source: SOC 2 Summary] | high
Q27 | Where is customer data hosted? | Customer data is hosted in AWS, primary region us-east-1, with documented data residency controls. | [Source: Architecture Overview, sec 2] | high
Q41 | Describe your incident-response SLA for customer notification. | NEEDS HUMAN INPUT - not covered in source docs | (route to Security) | low
Q55 | Do you support customer-managed encryption keys (CMEK)? | NEEDS HUMAN INPUT - not covered in source docs | (route to Engineering) | low
Summary: 51 of 60 answered at high confidence from source docs; 9 routed to humans. Estimated review time: ~1 hour vs. a multi-day manual effort.Pitfalls to avoid
Hallucinated controls or certificationsAn AI claiming a control or certification you do not have is a compliance violation and a trust catastrophe with the customer's security team. Enforce strict source-only answering and refuse-to-guess behavior, and never relax it to save time.
Stale source documentsOutdated or unapproved documents in the library produce confidently wrong answers. Keep the library current and approved-only, and assign an owner to maintain it rather than letting it drift.
Skipping human sign-offThe AI drafts; a person is accountable for what is submitted. Security or legal must review and approve every answer, including the high-confidence ones, because an approved-but-wrong answer is still your liability.
Reps guessing at flagged answersWhen a question is flagged as not covered, a rep guessing to keep the deal moving reintroduces the exact hallucination risk you designed the AI to avoid. Route flagged questions to the accountable expert, not back to the rep.
Not growing the libraryIf you do not feed newly approved answers back into the source library, you redo the same expert work next quarter. Each completed questionnaire should measurably reduce the human effort on the next one.