In short
An AI risk assessment checks each AI tool your people use against your data and security requirements: what data goes in, where it is stored, whether it trains on your data, who can access it and what it can do. You score the risk, decide to approve, restrict or block, then record the decision and review it.
Want help with AI risk assessment?
We can plan it with you or do it for you. Tell us where you are up to.
What an AI risk assessment is (and is not)
An AI risk assessment, in the sense most UK businesses need, is a structured check of the AI tools staff use or want to use. It asks whether each tool is acceptable for the data and tasks involved, and what controls are needed if it is.
It is not the same as:
- A DPIA. A data protection impact assessment is a legal requirement under UK GDPR for processing likely to result in high risk to individuals. An AI risk assessment often feeds a DPIA, but it also covers confidential business data, credentials and source code that are not personal data. See AI DPIA.
- Model risk management. If you build or train AI models, you also need testing for accuracy, bias and robustness. That is a different discipline; Check Point's AI red teaming and AI security products cover the security side of AI you build.
Done well, an AI risk assessment gives you a defensible approved list for your AI acceptable use policy, and a record for clients and auditors of why each decision was made.
Five steps to assess AI risk
- Inventory: discover what is in use. List the AI tools, AI features in SaaS apps, browser extensions, desktop apps, coding assistants, agents and MCP connectors in use, approved or not. Surveys depend on people remembering and recognising AI features; discovery tooling records actual use. See shadow AI discovery.
- Classify the data. For each tool, record what data people put in or connect to it: public, internal, confidential, personal, special category, credentials, source code. Use your existing classification scheme.
- Assess each tool. Check the provider's terms and settings: training use, retention, storage location, access controls, admin features, sub-processors, and what the tool can do (read files, send email, run code). The table below lists the questions.
- Score and decide. Combine the sensitivity of the data with what the tool's terms and features allow. Decide to approve, approve with restrictions, or block, and name the controls.
- Record and review. Write the decision down with its reasons, owner and review date. Review when the provider changes its terms or features, when use changes, and on a set cycle.
The first step matters most. You cannot assess a tool you do not know about, and the tools nobody approved are often the riskiest.
What to assess for each AI tool
These questions cover most of the risk. Answer them from the provider's own published terms, privacy policy and admin documentation, not from marketing pages or third-party summaries.
| Question | Why it matters | Where to look |
|---|---|---|
| Does the provider train models on our inputs? | Training use can mean your data influences outputs for others and is retained longer | Privacy policy, business terms, admin settings |
| How long are prompts and files retained? | Long retention increases exposure if the provider is breached | Data retention documentation, admin settings |
| Where is data stored and processed? | Transfers outside the UK bring UK GDPR transfer rules into play | Privacy policy, data residency documentation |
| Can we use work accounts with single sign-on? | Personal accounts leave data outside your control when people leave | Plan comparison, admin documentation |
| What admin controls exist? | You need to control apps, connectors, sharing and retention centrally | Admin console documentation |
| What can the tool access or do? | Agents and connectors act with real permissions on files, email and code | Connector and agent documentation |
| Which sub-processors are involved? | Your data may pass to other providers and model hosts | Sub-processor list |
| What security assurance is published? | Independent assurance reduces supplier risk | Trust or security pages, certifications |
Answers differ sharply between plans from the same provider. For example, OpenAI states that it does not train its models on business data by default for ChatGPT Business and Enterprise, while personal accounts are governed by consumer terms; our page on ChatGPT security risks explains the difference. DeepSeek's own privacy policy says it collects, processes and stores personal data in the People's Republic of China, and lists user input among the data it uses to train its models, with a right to opt out; see DeepSeek security risks. Microsoft says Microsoft Copilot only surfaces organisational data that a user has at least view permission to, which makes your file permissions part of the assessment; see Microsoft Copilot security.
Not sure which option applies to you? Ask us
Tell us about your organisation and we will point you to the right option.
Scoring and deciding
Keep scoring simple enough that people use it consistently. A three-level decision works for most organisations:
| Decision | Typical profile | Controls |
|---|---|---|
| Approve | Business plan with work accounts, no training on your data by default, acceptable storage location, admin controls | Listed in the acceptable use policy with permitted data types; DLP still applied for the most sensitive data |
| Approve with restrictions | Useful tool with gaps: limited admin control, uncertain retention, or only suitable for some data | Allowed for low-sensitivity tasks; sensitive data redacted or blocked; warnings shown to users |
| Block | Consumer terms only, training on inputs with no business control, unacceptable storage location, or excessive permissions | Access blocked; staff pointed to an approved alternative |
Write down the reasoning, not just the outcome. When a provider changes its terms or launches a business plan, you can revisit the decision quickly. Our guide on how to block AI tools explains why restricting is often better than blocking outright.
AI risk assessment template
Use these fields for each tool. A spreadsheet is fine; what matters is that every AI tool in use has an entry.
- Tool name, provider and plan (for example, consumer, business or enterprise)
- Business owner and the teams using it
- Purpose and approved use cases
- Data types used, and the highest data classification involved
- Training use, retention period and storage location, with the source you checked
- Account type, single sign-on and admin controls available
- Integrations, agents, extensions and connectors, with their permissions
- Sub-processors and security assurance
- Risk rating and decision: approve, approve with restrictions, or block
- Controls required (DLP rules, redaction, copy and paste limits, blocking, user guidance)
- Whether a DPIA is needed, and its reference
- Date assessed, assessor, and next review date
This template is general information, not legal advice. For the security policy that sits around these assessments, see AI security policy, and for the wider structure, our AI governance framework guide.
Prefer a hand with this?
We can help you plan it and set it up, alongside your own team.
Using Workforce AI Security for the Assess stage
Check Point builds Workforce AI Security around four stages: Discover, Assess, Govern and Protect. Several of them feed a risk assessment directly. Based on Check Point's material:
- Discover gives you the inventory: sanctioned and shadow AI apps, tools and MCPs, broken down by app, session and user, plus AI use inside connected SaaS apps and adoption trends.
- Assess helps you "understand AI risks relative to your security and compliance requirements". It uses AI to classify the content of prompts into tens of use-case categories (Check Point's examples include Marketing, Debugging and Legal), so you can see what each tool is actually used for, not just that it is used.
- The AI application catalogue lets you search for any app, including ones not yet used in your organisation, so you can assess a tool before someone requests it.
- Govern and Protect then apply your decisions: per-app policies, blocking of unauthorised AI apps, different policies for managed and unmanaged apps, AI-powered DLP with redaction, copy and paste restrictions, and runtime controls for risky agent actions.
The result is a risk assessment based on evidence from your own organisation rather than assumptions. AI used through web applications is covered by Essentials; desktop AI apps, coding assistants and agents need Enterprise. See how Workforce AI Security works.
How Forever Systems helps
We can help you get a first, evidence-based AI risk assessment done and keep it current:
- We can run a discovery pilot with Workforce AI Security, so the assessment starts from what is really in use.
- We help you work through the questions for the tools that matter most, and turn the decisions into policies: allow, restrict, redact or block.
- We help you choose between Essentials and Enterprise and confirm entitlements at quote.
For a wider look at your cloud security posture, our sister site offers a cloud security assessment. Ask us for a quote for Workforce AI Security. To talk about an AI risk assessment, contact us.
How Forever Systems helps you use AI safely
We are an independent UK IT and cyber security provider. We supply Check Point Workforce AI Security and help organisations put sensible, workable rules around AI at work.
- See what is really in use: we can start with a discovery pilot, so you know which AI tools, browser extensions, coding assistants and agents your people use before anyone writes rules
- Policy you can enforce: we help you decide what to approve, restrict or block, and turn your AI acceptable use policy into Workforce AI Security policies
- The right tier: Essentials for browser-based AI, Enterprise when desktop apps, developer code assistants and agents are in the picture, confirmed at quote
- Staged roll-out: from monitoring to warnings, redaction and blocking at a pace your people can absorb, with reviews along the way
- Around the tool: Microsoft 365 permissions ahead of Copilot, device management for the extension and agent, and safe use of AI training for staff
We work with businesses, schools, charities and public sector organisations. Our guides are general information, not legal advice.
Frequently asked questions
What is an AI risk assessment?
It is a structured check of the AI tools your organisation uses or plans to use: what data goes into each, how the provider handles it, what the tool can access or do, and whether the risk is acceptable. The outcome is a decision to approve, restrict or block, with the controls needed.
Is an AI risk assessment the same as a DPIA?
No. A DPIA is a UK GDPR requirement for processing of personal data likely to be high risk. An AI risk assessment is broader, covering confidential business data, credentials and source code too, and it often provides the information a DPIA needs.
What should an AI risk assessment template include?
Tool, provider and plan; owner and users; purpose; data types and classification; training use, retention and storage location; account and admin controls; integrations and permissions; sub-processors; risk rating and decision; required controls; DPIA reference; and review dates.
How often should we review AI risk assessments?
On a set cycle, and whenever a provider changes its terms or features, a new plan becomes available, or the way your people use the tool changes. AI tools change quickly, so annual reviews alone are rarely enough.
Is there an AI risk assessment tool?
Discovery and governance tools make the work much easier. Check Point Workforce AI Security discovers AI apps, tools and MCPs in use, classifies what prompts are used for, assesses risk against your requirements and applies the resulting policies.
Still have a question about AI risk assessment?
Ask us and a member of our team will reply by email.