Enterprise deals get stuck in AI security review when the buyer can't confirm what happens to their data inside your AI features. The usual causes are vague answers, no evidence, unclear model provider terms, doubt about whether customer data trains a model, no admin controls, no named owner, and AI features shipped without evals. Each one creates a follow-up round, and each round adds time. The fix is to answer from your real system, with proof, and to close the engineering gaps that make honest answers hard.
This guide is for SaaS founders, CTOs and sales leads in India and the US who have watched a signed-off deal sit in review. It explains why AI triggers the extra scrutiny, who is reading your answers, and what to do about it.
Why AI features trigger an extra review
A buyer's security team already knows how to review a normal SaaS product. They check access control, encryption, backups and incident response. AI features add risks their standard checklist doesn't cover:
- A new data flow. Customer data now leaves your app and goes to a model provider, often in another country.
- A new subprocessor. The model provider is a third party the buyer's privacy team has not approved.
- The training fear. Buyers worry their data could improve a model that also serves their competitors.
- New copies of data. Prompts, outputs, embeddings and logs are all copies, each with its own retention.
- Unpredictable output. A model can be steered by a crafted input. Prompt injection is first on the OWASP Top 10 for LLM Applications 2025.
- Regulation. EU buyers now work under the EU AI Act, whose bans on prohibited practices applied from 2 February 2025 and whose rules for general-purpose AI models applied from 2 August 2025. Indian buyers are preparing for the DPDP Act, most of whose duties start on 13 May 2027.
Standard questionnaires have caught up. Shared Assessments added an Artificial Intelligence domain to the SIG in its 2024 edition, and the Cloud Security Alliance released its AI Controls Matrix with the AI-CAIQ questionnaire in July 2025. So procurement asks about your AI features because their own templates now tell them to.
Who is reading your answers

Your champion bought the product. They are not the people who approve it. In a larger company, four groups may read the AI section, and any one can hold the deal:
| Reviewer | What they want to know | What stalls them |
|---|---|---|
| Security | What they want to know: Can the AI leak data across tenants or be manipulated? Is it logged? | What stalls them: No architecture detail, no test results |
| Privacy | What they want to know: Where does personal data go, for how long, and to which subprocessors? | What stalls them: A model provider not on your subprocessor list, no retention numbers |
| Legal | What they want to know: Do the contract and DPA cover AI use and training? | What stalls them: No clause on training, provider terms that conflict with yours |
| AI risk committee | What they want to know: Is the use case acceptable? Is there human oversight and an owner? | What stalls them: No AI policy, no named owner, no way to turn features off |
Not every buyer has an AI risk committee. Regulated buyers such as banks, insurers and healthcare companies are the most likely to. Where one exists, it may meet on a fixed cycle, so missing a meeting can add weeks.
The seven reasons deals stall
1. Vague answers
"We use industry-standard safeguards" doesn't answer anything. The reviewer can't mark it as passed, so they send it back. Each vague answer turns into a follow-up question, and follow-ups come back in batches, days or weeks apart.
2. Claims with no evidence
A yes is a claim. Reviewers increasingly ask for the proof: the provider DPA, the retention setting, the isolation test. If the evidence doesn't exist yet, someone on your team has to create it while the deal waits.
3. Model provider terms nobody has read
Buyers ask what your model provider does with their data. Many teams don't know which terms they signed up under. API terms and consumer terms differ. OpenAI's API documentation says API data isn't used for training unless you opt in, and that abuse monitoring logs are kept for up to 30 days by default. (OpenAI) That 30-day copy surprises buyers if you said "we don't store anything".
4. Doubt about training on customer data
This is the question that most often escalates to legal. If your terms of service give you a broad right to use customer data to "improve our services", the buyer's lawyer will read that as permission to train. Even if you never do it, expect to be asked to change the clause or carve out AI.
5. No opt-out and no tenant controls
Many enterprise buyers want to switch AI features off for some teams, or for everyone, until their own review is done. If AI is always on and there is no admin toggle, the reviewer can't approve a partial rollout. The same applies to tenant isolation: if retrieval runs over a shared index and you can't show the tenant filter, security won't sign off.
6. No owner
When the questionnaire arrives, it goes to whoever is free. Sales forwards it to the CTO, the CTO forwards it to an engineer, the engineer guesses at the privacy answers. Nobody owns the full set, so it moves slowly and answers contradict each other. Buyers notice contradictions, and they cost trust.
7. AI features shipped without evals
Reviewers may ask how you know the AI feature gives correct, safe answers, and how you know it still does after you change the model or the prompt. If the honest answer is "we tried it before launch", the reviewer has nothing to mark as passed. An eval is a fixed set of test cases, run on every change, with the results kept. Three kinds answer most of these questions: quality on realistic tasks, safety cases such as prompt injection and leaks across tenants, and regression runs that show a model or prompt change did not make things worse.
This is governance debt: the feature works, but nothing on file shows it. Frameworks buyers cite expect the evidence. The Measure function of NIST AI RMF 1.0 covers assessing, benchmarking and monitoring AI risks, and ISO/IEC 42001:2023 has controls for the AI system life cycle. Start small: a set of real cases from your own customers' workflows, scored the same way every release, with the results dated and stored where you can attach them to a questionnaire answer.
What to do this week
If a deal is stuck now, focus on the open items, not on a governance programme:
- Get the exact open questions. Ask your champion for the reviewer's follow-ups in writing. Don't guess what's blocking.
- Name one owner. One person collects every answer and checks it with the engineer who built the feature.
- Pull the provider terms. Find the commercial terms and DPA you are actually under, and note the clauses on training and retention.
- Draw the data flow. One diagram: what data goes where, to which provider, in which region, and where it is stored.
- Answer the hard ones honestly. Where you have a gap, say so, give any compensating control, and give a dated plan. Our guide on how to answer the AI section of a security questionnaire has sample answers for each group.
- Offer a call. A 30-minute call between your engineer and their security reviewer often clears more than another written round.
What to fix this quarter

The gaps that keep coming back are usually engineering, not paperwork. These are the changes that turn hard questions into easy ones:
- Tenant data isolation. Filter every retrieval query by tenant on the server, and add tests that try to cross tenants.
- Retention. Set a limit for prompts, outputs, embeddings and logs, and build the job that deletes them.
- Logging. Log AI requests with user, tenant, feature and model, but keep prompt text out of application logs.
- Model provider terms. Make sure you are on commercial API terms, data sharing is off, and the provider is on your subprocessor list.
- Admin controls. Let customer admins turn AI features on or off per organisation, and show usage in the audit log.
- Internal AI use. Find out which AI tools your own team uses with customer data. A shadow AI audit maps this.
How to stop it recurring
The first AI review is slow because you are building answers from scratch. The next one shouldn't be. Four assets make the difference:
An answer library
Keep one approved answer for each common AI question, with the evidence file linked and the date it was last checked against the code. Reuse it across the SIG, the AI-CAIQ and buyers' own forms. Review it whenever you ship an AI feature or change a provider.
A trust page
Publish what you can: which AI features exist, which providers they use, whether you train on customer data, retention periods and admin controls. Many reviewers read it before they send the questionnaire, and some questions never get asked.
An AI use policy
A short policy with a named owner: approved models and providers, what data may go to them, how new AI features are reviewed before launch, and how employees may use AI tools. It doesn't need to be long. It needs to be true. You can reference NIST AI RMF 1.0 or ISO/IEC 42001:2023 as frameworks, but don't imply a certification you don't hold.
Admin controls in the product
Per-organisation toggles, usage visibility and clear AI labels answer a whole block of questions with a screenshot. They also let a cautious buyer start with AI off and switch it on later, which keeps the deal moving.
Run a tech audit before the deal, not during it
Most of the stalls above are found by the buyer's reviewers, one follow-up at a time, while the deal waits. A pre-deal tech audit finds them first, by reviewing your system the way a security and architecture reviewer will. These are the areas they look at:
| Area | What reviewers look for |
|---|---|
| Architecture and data flow | What reviewers look for: A current diagram of where customer data goes, including model providers, regions and every store of prompts, embeddings and logs |
| Tenant isolation | What reviewers look for: Server-side tenant filters on every query and retrieval, with tests that try to cross tenants |
| Identity and access | What reviewers look for: SSO for customer users, roles with least privilege, and controlled access for your own staff to production data |
| Logging and audit trail | What reviewers look for: Who did what, when, including AI requests and admin changes, with prompt text kept out of application logs |
| Data handling | What reviewers look for: Encryption, retention limits that are actually enforced, deletion at the end of the contract, and backups |
| Secrets and dependencies | What reviewers look for: No keys in code, a process for patching vulnerable dependencies, and a record of when it last ran |
| AI controls | What reviewers look for: Provider terms and training settings, admin toggles per organisation, and evals with dated results |
| Incident response | What reviewers look for: A written plan with a named owner, and breach notice times that fit the contract |
Run it before your first big enterprise deal, and again after a major AI release or a change of model provider. Our guide to enterprise readiness for your first big customer covers SSO, audit logs and AI controls in more depth, and our tech audit is the version we run for you.
India and US: what changes
The questions are mostly the same. The emphasis differs:
- US enterprise buyers tend to send long standard questionnaires, often the SIG, and frame AI questions around frameworks such as NIST AI RMF and its Generative AI Profile, NIST AI 600-1. Legal review of training rights is usually thorough.
- Indian enterprise buyers, especially in banking, insurance and other regulated sectors, focus on where data is processed and on DPDP Act readiness. They will ask whether you act as their Data Processor and what you can show for security safeguards and deletion. Our vendor evidence pack covers that side.
- Indian SaaS companies selling to the US face both. Data flows from US customers to Indian engineering teams and back to US-hosted model providers need to be drawn clearly, or reviewers will ask.
Where Mono helps
We help teams answer AI security questionnaires when a deal is waiting. We review the AI section of your live questionnaire, find the gaps between what you can claim and what your systems show, and fix the engineering: logging, tenant data isolation, retention, model provider terms and admin controls. Then we assemble the evidence so each answer is backed by proof. It is part of our AI governance work for SaaS. If you want to reduce dependence on a single model provider, see own your AI.
We are engineers. This is not legal advice and not a certification, and your lawyers still own the contract.
FAQ
Why do enterprise buyers ask about our AI features in security review?
AI features send customer data to a model provider, create new copies of data in prompts, embeddings and logs, and can be manipulated through crafted inputs. Standard questionnaires now include AI sections, such as the Artificial Intelligence domain added to the SIG in 2024 and the CSA AI-CAIQ released in 2025. So procurement teams are following their own templates when they ask.
What are the most common reasons a deal stalls in AI security review?
Vague answers, claims with no evidence, unclear model provider terms, doubt about whether customer data trains a model, no admin controls or tenant isolation proof, no single owner for the answers, and AI features shipped without evals. Each one creates a follow-up round. Fixing them usually takes a mix of honest answers and small engineering changes.
Who reviews the AI section of a vendor questionnaire?
Usually security, privacy and legal, and in larger or regulated companies an AI risk committee as well. Security looks at isolation, logging and prompt injection; privacy at data flows, retention and subprocessors; legal at training rights in the contract; the committee at the use case, oversight and ownership. Any one of them can hold the deal.
How can a SaaS company speed up AI security reviews?
Keep an answer library with evidence linked, publish a trust page about your AI features, write a short AI use policy with a named owner, and build admin controls so customers can switch AI on or off. These let you answer most questions from existing material and give buyers a way to start cautiously.
Do Indian and US enterprise buyers ask different AI security questions?
The core questions are the same: training, retention, subprocessors, isolation and ownership. US buyers often send the SIG and reference NIST AI RMF, with close legal review of training rights. Indian buyers tend to focus on data location and DPDP Act readiness, including your role as a Data Processor.
