AI System Impact Assessment - Template
An editable AI system impact assessment procedure, with the assessment record to complete for each AI system. Replace the bracketed placeholders with your own details.
[Organization] develops, provides or uses AI systems that can affect people who never chose to interact with them. This procedure sets out how we assess the potential impacts of each AI system on individuals, groups of individuals and society, before it is deployed and throughout its life, how we decide whether and how to proceed, and how we record the result. The assessment record in section 15 is completed for each AI system and updated at every review.
1. Purpose
This procedure defines how [Organization] identifies and assesses the potential consequences an AI system can have for individuals, groups of individuals and society as a result of its development, provision or use. It exists so that those consequences are understood and acted on before harm occurs, and so that every decision to deploy, change or retire an AI system can be traced to a documented assessment.
This assessment is separate from, and feeds into, the AI risk assessment. The risk assessment asks what an AI system could mean for [Organization]; this assessment asks what it could mean for the people and communities it touches.
2. Scope
This procedure applies to every AI system [Organization] develops, provides to others or uses in its operations within the scope of [our AI management system], including AI features built into third-party products where [Organization] decides how they are used. It applies from the point a new AI system is proposed until it is retired.
Any exclusion is recorded here with its reason: [exclusions and reasons, or "None"].
3. When an assessment is required
No AI system is deployed, and no significant change is put into use, until the impact assessment for it has been approved under section 10. An assessment is completed, or the existing one reviewed and updated, in each of the following situations:
- A new AI system is proposed, whether it is built in-house, bought, or enabled as a feature of a product we already use.
- A significant change is proposed to an existing AI system, for example a new model or model version, new training or input data, a change in the degree of automation, or the removal of a human review step.
- An AI system is to be used for a new purpose, in a new context or location, or by or on a new group of users or affected people.
- An incident, complaint or near miss suggests the system has caused, or could cause, an impact the current assessment did not identify.
- The periodic review falls due, at least [annually] for each AI system in scope.
4. Roles and responsibilities
- Assessment owner - [role, for example the product or system owner] leads the assessment for each AI system, gathers the information in section 5, completes the assessment record and tracks mitigation actions to completion.
- Contributors - the assessment owner involves the people who understand the system and the people it affects, including [roles, for example the data owner, a technical lead, the privacy lead, legal or compliance, and customer-facing staff]. Where practical, the views of affected people or their representatives are sought [through (method)].
- Approver - [role, for example the AI governance committee or a named executive] reviews each completed assessment and makes the decision in section 10. The approver is not the assessment owner for the same system.
- AI governance lead - [role] owns this procedure, keeps the register of AI systems and their assessments, and reports on assessment status to [management forum] at least [quarterly].
5. Describing the AI system
Each assessment starts with a description of the AI system, recorded with the assessment record. It covers:
- Intended purpose - what the system is for, the outputs or decisions it produces, and what it is not intended to be used for.
- Users - who operates the system or acts on its outputs, inside and outside [Organization].
- Affected people - who is subject to its outputs, including people who never interact with it directly.
- Data - the data used to train, test and operate it, where that data comes from, and whether it includes personal or sensitive information.
- Deployment context - where and how it is used, in which locations and jurisdictions, and what else in the process depends on it.
- Degree of automation and human oversight - whether outputs are advisory, reviewed by a person before they take effect, or acted on automatically, and who can intervene or override.
- Supplier and version - for a bought or third-party system, the supplier, the model or product version assessed, and the information the supplier has provided about it.
6. Identifying who is affected
The assessment identifies everyone the AI system could affect. Affected people include those who do not use the system and may not know it is involved in a decision about them. Where a group could be affected differently from others, the assessment names that group rather than describing affected people only in general terms.
Affected people are identified at three levels:
- Individuals - the people a single output is about or acts on, for example an applicant, a customer or an employee.
- Groups of individuals - groups who could be affected differently from others, for example by age, disability, language, location or another characteristic relevant to the context.
- Society - effects beyond the people directly involved, for example on access to an essential service, on public trust or on the environment.
7. Assessing potential impacts
For each individual, group and societal effect identified in section 6, the assessment considers the potential impacts of the AI system, beneficial and harmful, both when it works as intended and when it fails or produces errors. Each impact is rated for severity and likelihood on the scales in the table below, before and after mitigation, and recorded in the assessment record. Where information is missing, the assessment records the gap and the action to close it rather than assuming the impact is low. [Use the same scales as your AI risk assessment methodology where you have one, and replace these if they differ.]
At a minimum the assessment considers:
- Fairness and non-discrimination - whether outputs could be less accurate, or less favorable, for some groups than for others.
- Privacy - whether the system collects, infers or exposes personal information beyond what people would reasonably expect.
- Safety - whether an output, or acting on one, could cause physical or psychological harm.
- Access to services - whether the system could restrict or delay someone's access to work, credit, housing, education, health care or another service.
- Transparency and the ability to contest a decision - whether affected people know AI is involved, can get an explanation they understand, and can challenge an outcome and reach a person.
- Wellbeing - whether the system could affect people's autonomy, dignity or mental health, for example through persuasive design or continuous monitoring.
- Environmental impact - the energy and resources the system uses to train and run, weighed against its benefit.
- [Any other impact relevant to this system or sector.]
| Rating | Severity - how serious the impact would be | Likelihood - how likely it is to occur |
|---|---|---|
| Low | Minor, short-lived inconvenience that is easily reversed. | Not expected to occur in normal use. |
| Medium | Noticeable harm to individuals that can be put right with some effort. | Could occur in some circumstances. |
| High | Significant harm to individuals or a group, or harm that is hard to reverse. | Expected to occur at some point in normal use. |
| Very high | Serious or lasting harm to individuals, a group or society, or harm that cannot be reversed. | Expected to occur often, or already occurring. |
8. Reasonably foreseeable misuse
The assessment also considers how the AI system could be used in ways it was not intended for but that can reasonably be foreseen, and the impacts that would follow. Each foreseeable misuse with a material impact is recorded as its own line in the assessment record. Consider at least:
- Using outputs for a decision the system was not designed or tested for.
- Relying on outputs without the human review this assessment assumes.
- Using the system on people, data or locations outside those it was tested on.
- Deliberately manipulating inputs to obtain a particular output.
9. Mitigation measures and human oversight
For each impact that is not acceptable as it stands, the assessment records one or more mitigation measures, the owner of each and the date by which it will be in place, then re-rates the impact to give the residual concern. Mitigations can include changes to the system or its data, limits on where and how it is used, information given to affected people, a route to contest a decision, monitoring, and human oversight.
Where human oversight is relied on as a mitigation, the assessment names the role that provides it, what that person reviews, the authority and time they have to override or stop the system, and how their interventions are recorded. A human review step that cannot change the outcome is not recorded as a mitigation.
10. Decision and approval
On the basis of the assessment, the approver makes one of three decisions and records it, with the reasons, in the assessment record. Any impact with a residual severity of [High] or above is escalated to [role or forum] for the decision. The approver records their name, title and the date of the decision, and the assessment owner records when each condition has been met and what shows it.
The three decisions are:
- Proceed - the residual impacts are acceptable and no further conditions apply.
- Proceed with conditions - the system may be deployed or changed only once the listed mitigations are in place, or only within stated limits, each condition with an owner and a date.
- Do not proceed - the residual impacts are not acceptable. The system is not deployed, or the change is not made, until it is redesigned and reassessed.
11. Link to the AI risk assessment
The results of each impact assessment are passed to [role responsible for the AI risk assessment] and considered in the AI risk assessment. Each impact with a residual severity of [High] or above, and each open condition, is reflected in the AI risk register as a risk to [Organization], cross-referenced to the assessment record by its reference number. Where the risk assessment selects a treatment that changes the system, the impact assessment is updated to match.
12. Documenting and retaining results
Each completed assessment, including the record, the evidence it relied on and the approval, is stored in [location, for example the AI governance register] under a unique reference and a version number. Earlier versions are kept, not overwritten, so that the assessment in force on any past date can be shown. Assessments are retained for [period, for example the life of the AI system plus (number) years], or longer where a legal or contractual obligation requires it.
13. Communicating results
Where appropriate, [Organization] makes the results of an impact assessment, or a summary of them, available to relevant interested parties such as customers, users, affected people or their representatives, and regulators where required. [Role] decides what is shared, with whom and in what form, taking account of confidentiality, security and legal obligations, and records that decision. Affected people are told how to raise a concern or contest a decision about them [through (channel)].
14. Review
Each assessment is reviewed at least [annually], and whenever one of the triggers in section 3 occurs. The review checks whether the system, its use, the people it affects or its impacts have changed, whether conditions and mitigations are in place and working, and whether any complaint, incident or monitoring result points to an impact the assessment missed. A review that finds no change is still recorded, with its date.
This procedure is owned by [role]. Approved by [name], [title], on [date]. Version [x.x]. Next review [date].
15. Assessment record
Complete one record for each AI system, and keep it with the system description from section 5.
- Reference: [AIIA-001]. AI system and version: [name, version].
- Assessment owner: [role]. Contributors: [roles]. Date: [date].
- Trigger: [new system / significant change / new use or users / incident / periodic review].
- Decision: [proceed / proceed with conditions / do not proceed]. Reasons: [summary]. Approved by: [name, title, date].
- AI risk register reference: [risk IDs]. Next review: [date].
| Impact area | Who is affected | Description | Severity | Likelihood | Mitigation and oversight | Residual concern | Owner |
|---|---|---|---|---|---|---|---|
| Fairness and non-discrimination | [Group, for example applicants whose first language is not English] | [How outputs could be less accurate or less favorable for this group] | [High] | [Medium] | [For example, test results by group before release, and a person reviews every adverse outcome] | [Medium, monitored (monthly)] | [Role] |
| Transparency and the ability to contest | [Individuals the output is about] | [Whether they know AI is involved and can challenge the outcome] | [Medium] | [High] | [For example, a notice at the point of use and a named contact who can review the decision] | [Low] | [Role] |
| Privacy | [Individuals whose data is used] | [What personal information is used, inferred or exposed] | [Severity] | [Likelihood] | [Mitigation] | [Residual concern] | [Role] |
| Foreseeable misuse: [describe] | [Who is affected] | [What would happen] | [Severity] | [Likelihood] | [Mitigation] | [Residual concern] | [Role] |
| [Impact area] | [Who is affected] | [Description] | [Severity] | [Likelihood] | [Mitigation] | [Residual concern] | [Role] |
This template gives you a working procedure and a record to complete for each AI system, but a generic procedure on its own does not make an ISO/IEC 42001 AI management system. That needs the whole set - an AI policy, an AI risk assessment, a Statement of Applicability and the other documents your controls require - each tailored to your AI systems and consistent with the others. Some laws also set their own assessment requirements for certain AI systems; check which apply to you, because this template is not written to satisfy any particular law. PolicyMint generates the AI System Impact Assessment Procedure and the rest of the set, tailored to your business and checked before you see it.
