In this guide · 7 sections
Short answer
Usually not in full. AI Verify is voluntary and built for organisations that develop or deploy AI models, with about 100 process checks plus technical tests. An SME can borrow its ideas: document what each AI tool does, test it against wrong and harmful answers, keep a human in the loop, and record the results.
What AI Verify is, in plain English
AI Verify has two parts. The testing framework is a structured list of principles, desired outcomes, processes and the evidence that shows you followed them. The testing toolkit is software that runs technical tests on a model and produces a report. Both are maintained by the AI Verify Foundation, a not-for-profit set up with IMDA.
| Date | Milestone |
|---|---|
| 25 May 2022 | AI Verify framework and toolkit launched by IMDA and PDPC as a minimum viable product for pilot |
| 7 June 2023 | Framework and toolkit open-sourced on GitHub |
| 29 May 2025 | Updated framework released to cover generative AI risks |
| 18 May 2026 | AI Tester Accreditation Programme (AI TAP) announced for third-party AI testing firms |
None of this is a legal requirement. It sits alongside IMDA's voluntary Model AI Governance Framework, which our AI governance framework guide explains for small businesses.
What the AI Verify framework checks
The framework assesses an AI system against 11 internationally recognised governance principles and is mapped to the US NIST AI Risk Management Framework, the Hiroshima Process Code of Conduct and ISO/IEC 42001. Each principle has a set of process checks, each backed by documentary evidence.
| Principle | Process checks | What it means for a small business |
|---|---|---|
| Transparency | 15 | Tell users when they are dealing with AI and what it does |
| Explainability | 1 | Be able to explain how an answer or decision was reached |
| Reproducibility | 10 | Record versions, prompts and settings so results can be repeated |
| Safety | 17 | Stop harmful or dangerous outputs |
| Security | 14 | Protect the system and data from misuse and attacks |
| Robustness | 10 | Behave sensibly with odd or unexpected inputs |
| Fairness | 6 | Avoid unfair treatment of groups of people |
| Data governance | 4 | Know what data goes in and where it comes from |
| Accountability | 14 | Someone is responsible, with clear processes |
| Human agency and oversight | 9 | A person can review, override or stop the AI |
| Inclusive growth and well-being | 6 | Consider wider impact on people and society |
The checks come as a PDF, an Excel tracker, and a software tool for generative AI. A large organisation might spend weeks gathering evidence for all of them. A small business can read the list in an hour and pick the handful that apply.
The toolkit and Project Moonshot: what they test
The AI Verify Testing Toolkit runs technical tests for fairness, explainability and robustness on traditional machine learning models, such as a model that scores loan applications or predicts customer churn. Its user guide covers preparing tabular and image datasets and models, and installing it is a developer task.
Project Moonshot is the generative AI side. It is an open-source toolkit that tests large language model (LLM) applications through benchmark tests and red teaming, implementing the benchmarks recommended in IMDA's Starter Kit for Safety Testing of LLM-based Applications. For agent-based systems, see our guide to IMDA's agentic AI testing starter kit.
AI Verify Testing Toolkit
- For traditional ML models you trained or deployed
- Fairness, explainability, robustness tests
- Needs your model and datasets
- Developer set-up required
Best for: Businesses with their own scoring or prediction models.
Project Moonshot
- For LLM apps such as chatbots
- Benchmarks and red teaming
- Connects to the model or app endpoint
- Developer set-up required
Best for: Teams building customer-facing AI assistants.
Light SME version
Most SMEs- A register of AI tools in use
- A short written test script
- Human review of risky outputs
- No software install
Best for: Businesses using ChatGPT, Copilot or a vendor chatbot.
When a customer or regulator might ask about AI testing
AI Verify is voluntary, so nobody can demand an AI Verify report by law. In practice, the question about AI testing usually comes from somewhere else:
- Enterprise or government customers sending a vendor questionnaire that asks how your AI features are tested and supervised.
- Regulated sectors such as finance, insurance and healthcare, where sector guidance expects risk management for AI decisions.
- Investors or partners doing due diligence on an AI product you sell.
- A complaint or incident, for example a chatbot giving a customer wrong prices or advice, where you need to show what checks you had.
- Personal data questions from the PDPC if AI tools process customer data, which brings in your PDPA obligations.
If you sell an AI product to larger organisations, mapping your own testing to the AI Verify principles, or using an accredited tester under AI TAP, gives you a recognised answer. If you only use AI tools internally, a short documented process is usually enough.
A lighter AI check an SME can borrow
You do not need the toolkit to apply its thinking. For an AI chatbot, a ChatGPT workflow or AI-drafted customer replies, this half-day exercise covers the principles that matter most for a small business:
List every AI tool and what it does
Name, vendor, what data goes in, who uses it, and whether customers see its output. This is your transparency and data governance check.
Name an owner for each tool
One person decides how it is used and reviews problems. That covers accountability.
Write 20 test questions
Include normal questions, tricky ones (prices, refunds, medical or legal topics) and attempts to make it misbehave. This is a mini red-team.
Run the tests and record the answers
Save the prompts, answers, date and model version in a spreadsheet. Mark each as good, wrong or harmful.
Fix and add guard-rails
Limit topics, add your real prices and policies as source material, and route risky questions to a human.
Repeat after every change
Rerun the same 20 questions when you change the model, prompt or knowledge base. Keep the results.
Do it yourself or get help?
If you only use AI tools internally, the light check above is something you can do yourself in an afternoon. If you are putting an AI assistant in front of customers, it is worth having it set up with tight limits from the start, and tested before launch.
AI check for a small business
Our AI Website Assistant (S$790) is trained on your own website, FAQs and documents, with guard-rails so it only answers about your business, hand-over to WhatsApp or email, and a usage report after the first 30 days.
Frequently asked questions
No. AI Verify is a voluntary testing framework and toolkit run by the AI Verify Foundation with IMDA. There is no general law requiring businesses to use it. Some customers, partners or regulated sectors may ask how your AI is tested, and AI Verify gives a recognised structure for answering.
Yes. The AI Verify Testing Framework and Toolkit were open-sourced on GitHub in June 2023, and Project Moonshot is also open source. The cost is the time and technical skill to install the tools, prepare your models and data, and gather evidence. Paid third-party testers are a separate option.
Project Moonshot is an open-source toolkit from the AI Verify Foundation that tests the safety and reliability of large language models and LLM applications through benchmarks and red teaming. It implements the benchmarks in IMDA's Starter Kit for Safety Testing of LLM-based Applications and is aimed at developers and deployers of AI apps.
Not directly in a useful way. The toolkit's technical tests target models you control, such as your own machine learning models, and Moonshot needs a developer to connect to an LLM app. A small business using ChatGPT is better served by a register of AI uses, a written test script and human review of important outputs.
AI TAP, announced on 18 May 2026, is run by the AI Verify Foundation. It accredits third-party firms that can technically test AI systems for safety, reliability and governance risks. Businesses that need independent testing of an AI product can use an accredited firm for added confidence.