Wondering how to choose AI software for business without getting burned? Start by sorting each tool by risk, then ask every vendor about data use, testing, contracts and monitoring. A short, repeatable checklist beats a gut feeling.
This checklist is one part of the broader subject of ai tools for business, which covers the main product types and where each one fits.
| Item | Value |
|---|---|
| EU AI Act maximum fine, prohibited practices (Article 5) | Up to €35 million or 7% of global annual turnover |
| EU AI Act maximum fine, high-risk AI breaches (Article 99) | Up to €15 million or 3% of turnover |
| Vendor risk tiers in a tiered review | 4 tiers (critical, high, medium, low) |
| AI-specific risk categories standard reviews miss | 6 categories |
| Documentation categories auditors commonly ask for | 8 categories |
| SOC 2 Trust Services Criteria | 5 criteria, with Security required in every report |
These figures come from the Prediction Guard vendor due diligence checklist, dated June 2026. Rules and penalties can change, so check the source before you rely on them.
Why AI software needs more than a standard software review
Most buyers begin with the paperwork they already know: a SOC 2 report, a privacy policy, a security questionnaire, a data processing agreement and a subprocessor list. One compliance publisher says these are important but not enough for AI products, because AI vendors raise questions that traditional vendor review often misses (Captain Compliance).
The Prediction Guard checklist names six AI-specific risk categories that standard SaaS reviews tend to skip:
- Model opacity: you often cannot tell why a system made a decision, which matters in hiring or lending.
- Training data lineage: you need to know where the data came from and whether it fits privacy rules.
- Algorithmic bias liability: you bear responsibility for the disparate impact of selection tools you use, including those built by third-party vendors.
- Prompt injection: attackers can manipulate inputs to gain unauthorized access or extract hidden system instructions.
- Sub-processor chains: third-party AI dependencies can fail without warning.
- Model drift: you must keep monitoring for deviations from expected data distributions or model behavior.
The stakes grow with the market. The Responsible AI Institute noted in August 2023 that the global AI market was valued at $136 billion in 2022 and projected to grow at a compound annual growth rate of 37.3% from 2023 to 2030 (Responsible AI Institute).
Sort every tool by risk before you ask questions
The first decision gate is risk tiering. The depth of your review should match the risk: critical vendors handling sensitive data warrant full due diligence, while low-risk vendors providing commodity services need only basic business verification and sanctions screening (Prediction Guard).
That source describes four tiers you can adapt:
- Critical: health data, personal data, financial data or proprietary algorithms. Run the full review and ask for a SOC 2 Type II report; ISO 27001 is encouraged.
- High: internal confidential data or non-sensitive customer data. Add a security and legal review with a detailed questionnaire.
- Medium: lower-sensitivity data. Use a focused questionnaire plus a certification check.
- Low: minimal-sensitivity data. Verify the business and screen for sanctions.
Another guide says high-risk systems, such as those involved in hiring, credit decisions, healthcare or public-facing interactions, deserve more rigorous review than a low-risk productivity assistant (Optro). A governance library review of one assessment guide makes the same point: not every system deserves the same scrutiny (AI Governance Library).
If your team is small, read about ai tools for small business before deciding how much of this checklist to apply. The same review notes that smaller vendors may struggle to answer every question at depth, and that not every buyer has the bandwidth to interpret complex answers.
What should you ask every AI vendor?
Start with data, because that is where most surprises hide. Captain Compliance suggests asking whether the vendor uses customer data to train or improve models, whether prompts and outputs are stored, and whether training can be disabled (source).
Then work through the system itself:
- Does the vendor build its own model or rely on a third-party foundation model?
- Do subprocessors also touch your AI data?
- Has the system been tested for accuracy, bias, security and reliability?
- Does it create audit logs, and can humans override the AI output?
- Will the vendor notify you of model changes?
The Responsible AI Institute condenses its guidance into three questions that work well in RFPs and procurement documents. They cover responsibility at the organizational level (policies and processes), at the system level (development and management) and at vendor handoff (documentation and support after go-live) (Responsible AI Institute).
Ask for specifics, not assurances. The AI Governance Library highlights questions that elicit operational practice, such as whether a vendor supports BYOK (Bring Your Own Key), client-configurable retention periods or model scorecards (source).
If the product is a customer-facing assistant, the same questions apply, and the category has its own wrinkles covered in ai chatbots for business.
Contract and compliance checks before you sign
Contracts turn answers into obligations. One procurement guide warns that organizations routinely sign contracts that allow vendors to use enterprise data for model training (Optro), so check that your contract prohibits unauthorized model training (Captain Compliance).
Certifications give you independent evidence. Ask for copies of SOC 2 Type II reports or ISO 27001 certificates and record the date you verified them. The AICPA SOC 2 criteria include CC9.2, which requires identification and evaluation of vulnerabilities arising from vendor and business partner relationships, so your own vendor records can be examined too (Prediction Guard).
Map your questions to recognized frameworks rather than inventing your own. The AI Governance Library says its reviewed guide aligns with ISO/IEC 42001, the NIST AI RMF and the EU AI Act (source). The NIST AI RMF GOVERN 1.1 function requires that legal and regulatory requirements involving AI are understood, managed and documented, and that obligation extends to every vendor supplying AI capabilities to you (Prediction Guard).
Penalties give the board a concrete reason to care. As of the Prediction Guard checklist updated in June 2026, EU AI Act fines reach up to €35 million or 7% of global annual turnover for prohibited practices under Article 5, and up to €15 million or 3% of turnover for high-risk AI breaches under Article 99 (source).
A step-by-step way to run the evaluation
Run the steps in order, because tiering first sets the scope of everything after it.
- Describe the use case. Write down what the tool does, whose data it touches and whether it influences decisions about people.
- Assign a risk tier. Decide how deep the review needs to be before you send any questionnaire.
- Send the questionnaire. Cover data use, model sourcing, testing, audit logs and human oversight.
- Verify the evidence. Check certifications and test reports instead of accepting a yes or no.
- Negotiate the contract. Address training on your data, retention, model-change notices and audit rights.
- File the paperwork. Keep dated answers and sign-off records.
- Monitor after launch. Review the vendor again when the model, the data or the use case changes.
Captain Compliance stresses that vendor due diligence should not be a one-time questionnaire buried in procurement (source). Treat it as a lifecycle process tied to your wider governance and vendor management.
Evidence to keep once you have chosen
Auditors commonly ask for evidence across eight documentation categories, and many organizations struggle to produce all of them on short notice (Prediction Guard). A tidy evidence file for each vendor should hold:
- Questionnaires with dated responses and sign-off evidence
- Risk rating methodology and tiering records with approvals
- Copies of SOC 2 Type II reports or ISO 27001 certificates, with verification dates
- Data processing agreements with sub-processor addenda
- AI incident reports and breach notification records
- Change management logs covering model updates and API modifications
- Records of audit rights exercised and follow-up actions
- Offboarding evidence, including signed data deletion certificates
Keep watching for model drift too. The same source says you must continuously monitor for signals that flag deviations from expected data distributions or model behavior, which is an ongoing responsibility beyond the first assessment.
Frequently asked questions
Do I need the full checklist for every AI tool?
No. Scale the review to the risk. According to the Prediction Guard checklist, low-risk vendors providing commodity services need only basic business verification and sanctions screening, while critical vendors handling sensitive data need full due diligence (source).
Is a SOC 2 report enough to trust an AI vendor?
Not on its own. A SOC 2 report, privacy policy and security questionnaire are important, but Captain Compliance says they are not enough for AI, because they miss questions about training on your data, stored prompts and model changes (source).
What are the EU AI Act fines?
As of the Prediction Guard checklist updated in June 2026, prohibited practices under Article 5 carry fines up to €35 million or 7% of global annual turnover, and high-risk AI breaches under Article 99 carry fines up to €15 million or 3% of turnover (source). Check the legal text for your own situation.
Which frameworks can I use to structure my questions?
The sources point to the NIST AI Risk Management Framework, the EU AI Act and ISO/IEC 42001, plus SOC 2 and ISO 27001 for security controls (AI Governance Library).
