AI Impact Assessment Template

An AI impact assessment asks how a proposed system may change outcomes for the people and groups affected by it. It goes beyond whether the model performs its task: who gains convenience, who bears mistakes, who may be excluded, what alternatives remain and how someone can question an outcome.

A useful template connects those questions to design decisions, evidence and named owners. It should not end as a list of ethical principles beside an unchanged implementation plan. If the assessment identifies that some customers cannot use the proposed channel, the team needs an operating response and a way to test whether that response works.

This guide provides an original assessment structure for Saudi organisations building AI governance setup. It is designed for business sponsors, service owners and reviewers, with a fictional recreation-class booking assistant illustrating how impact findings change the service design.

Set the assessment boundary around people and decisions

Describe the service problem, intended benefit, proposed AI function and the decision or action it influences. Identify the current process and who experiences it. “Improve booking efficiency” is insufficient if the actual change routes every customer through a conversational interface before they can see available classes.

Record the system boundary and the broader service boundary. The AI may only recommend a class, but its output can influence which options a clerk presents or which applications receive attention. Assess the actual path from input to consequence, including human decisions, queues, communications and fallback channels.

Identify the stage of development. An early assessment can influence whether AI is appropriate and which alternatives to test. Before operation, update it to reflect the actual configuration and evidence. A statement written when the system was only a concept should not be treated as validation of the final service.

Keep impact assessment distinct from other reviews

A technical evaluation asks whether a capability behaves as intended under specified tests. A privacy assessment examines relevant personal-data processing. A risk register tracks events, treatments and ownership. An impact assessment brings the affected person's experience and wider consequences into the decision, drawing on those other reviews without replacing them.

ISO/IEC 42005:2025 provides guidance on AI system impact assessment. Canada's Algorithmic Impact Assessment is a separate instrument supporting its federal automated-decision policy. Neither makes the original worksheet below an official Saudi compliance form or an ISO certification assessment.

Map applicable requirements separately with appropriate specialists. Internal impact findings can justify stronger design choices than a minimum requirement, while a favourable assessment cannot remove an obligation that applies to the use. Keep the source and status of each requirement visible.

Use this original impact assessment record

AI impact assessment fields to adapt for a defined use

Section	Questions to answer	Useful output
Purpose and baseline	What problem is being addressed, and how is it handled now?	Specific service objective and current process
Proposed intervention	What will AI do, recommend or execute?	Bounded function and downstream decision map
Affected people	Who uses it, who is affected without using it and who may be excluded?	Stakeholder and access map
Potential benefits	Who benefits, through which mechanism and with what evidence?	Testable benefit hypotheses
Potential harms	What adverse outcomes are plausible and how could they arise?	Scenario and consequence analysis
Alternatives	Can a different process achieve the purpose with different effects?	Comparison of viable options
Participation	Whose views were sought and whose are missing?	Consultation findings and limitations
Design commitments	What changes will address the identified effects?	Owned controls and service changes
Evidence	How will the commitments be tested?	Evaluation plan and results
Challenge and remedy	How can a person question or correct an outcome?	Accessible operational route
Decision	Proceed, narrow, repair, defer or avoid, and why?	Authorised conclusion with conditions
Reassessment	Which changes or observations reopen the assessment?	Triggers, owner and review point

Use references to existing records where they answer a question, but check that their scope matches. A generic accessibility policy does not establish that this booking journey has an assisted route. A supplier's aggregate performance claim does not show outcomes for the service's actual users.

For every unanswered field, identify the uncertainty and what decision it affects. Do not mark an unknown as not applicable solely to complete the form. The assessment can support a narrow discovery activity while withholding a broader operating recommendation.

Identify people who are absent from the user list

Start with direct users, then consider people affected by decisions or workload shifts. A booking assistant may be used by reception staff while customers bear the consequences of incorrect recommendations. Staff who receive escalations may experience additional work even if the pilot's headline metric reports shorter conversations.

Consider access barriers relevant to the service: language, digital confidence, disability-related needs, device availability and reliance on an assisted channel. These are questions to investigate, not demographic assumptions about every member of a group. Avoid collecting unnecessary personal details merely to make the assessment appear comprehensive.

Ask who cannot reach the point at which feedback is collected. A satisfaction survey completed only by successful online users misses people who abandoned the channel or could not enter it. Design participation and evidence collection to expose those missing experiences appropriately.

Trace potential harm through a causal chain

Write a scenario from system behaviour to consequence. For example: the assistant assumes a mobile number is required; a customer using an assisted channel cannot complete the flow; the clerk offers no alternative; the customer loses access to a suitable class. This chain identifies several possible interventions beyond improving the model's wording.

Separate severity, likelihood, duration and reversibility. An incorrect class suggestion can be easy to correct before reservation but harder after limited places are allocated. A misleading confirmation can cause someone to travel to a session they were never booked into. The timing and operational context matter.

Identify who benefits and who bears the burden of each control. Requiring a customer to telephone repeatedly may protect the automation from errors while transferring cost and effort to that person. A fairer design may require staff capacity or a different reservation process. Make that tradeoff explicit for the sponsor.

Worked example: recreation-class booking assistance

Consider a fictional local recreation service that wants an assistant to explain available classes and prepare a reservation for a clerk to confirm. The intended benefit is clearer information and less repeated data entry. The service includes Arabic and English interactions and an assisted route for people who do not use the online interface.

The assessment identifies a design assumption: the proposed workflow requires a mobile-number verification step before the clerk can complete every reservation. That assumption blocks some legitimate assisted scenarios. The team must decide whether the requirement is truly necessary for this service and how an authorised alternative can operate, rather than treating every blocked case as user error.

A functional exercise uses twenty-four deliberately constructed scenarios: ten Arabic online paths, six English online paths and eight assisted paths. Twenty complete successfully. The four failures all occur in the assisted group. The overall completion result is 20/24, about 83.3%, while the assisted-path result is 4/8, or 50%. These are fictional test-case results, not estimates of population outcomes or actual service performance.

Illustrative functional scenarios reveal a hidden access problem

Scenario family	Completed	Total	Interpretation
Arabic online	10	10	Selected online paths complete under test assumptions
English online	6	6	Selected online paths complete under test assumptions
Assisted	4	8	Four valid assisted paths are blocked by the workflow
All constructed paths	20	24	The aggregate obscures the concentrated failure

The assessment changes the design commitment: provide an authorised assisted reservation route with clear identity and booking checks appropriate to the service, train clerks to use it and make the alternative discoverable. The team then retests the failed cases and additional variations. Passing the repaired examples alone is not evidence that every access barrier has been resolved.

The decision can permit a bounded service trial only after the agreed route works and the operating owner accepts its staffing and monitoring. It should retain questions about real-world demand and outcomes for later observation. The assessment has produced a practical service change, not merely a statement that inclusion is important.

Compare alternatives before committing to the design

Consider a clearer conventional form, improved search, additional staff guidance or a limited assistant that explains information without preparing reservations. Compare them against the same service objective. AI may be useful for some parts of the journey while another mechanism is better for a consequential step.

Assess alternatives realistically. A manual channel is not automatically accessible if it is understaffed or available only at inconvenient times. A digital form is not automatically clear because it avoids a language model. Identify the resources and operating assumptions each option requires.

Record why the selected design is preferred and what would change that conclusion. This gives future reviewers a reason to reconsider the decision if benefits fail to appear or the alternative channel becomes overloaded. The risk acceptance record can capture any residual exposure requiring a distinct authorised judgement.

Turn consultation into decisions and evidence

Seek input from people who understand the service and, where appropriate, people likely to experience its barriers. Use accessible participation methods and protect personal information. A small consultation can reveal issues without proving that every affected group is represented; document who participated and who remains missing.

Summarise the concern, the team's response and the resulting design choice. If a suggestion is not adopted, explain why and what alternative addresses the underlying issue. Consultation should influence the assessment rather than serve as a count of meetings that validates a predetermined plan.

Use a separate evidence route for confidential or sensitive experiences. The public-facing summary can explain that a barrier was identified and addressed without exposing the participant's personal story. Internal reviewers should have enough controlled evidence to evaluate the conclusion.

Design a challenge route that can change an outcome

Explain how a person can ask for clarification, correct inaccurate information or obtain human reconsideration where relevant. Identify the receiving team, response process and authority to correct the result. A feedback button that only sends model-training data is not necessarily a remedy for a wrongly handled reservation.

Preserve the information needed to reconstruct the relevant interaction within appropriate retention and access rules. The reviewer may need to know the source version, assistant output, clerk action and final booking state. Avoid requiring the customer to prove a system failure using evidence only the organisation possesses.

Test the challenge route in Arabic and English and through the assisted channel. The route itself should not reproduce the barrier that caused the original problem. Link significant failures to the incident response process while keeping ordinary service correction manageable.

Turn each affected-person concern into a design commitment

Use a concern-to-commitment record rather than a general list of stakeholder comments. Capture who may be affected, the barrier or consequence, the evidence supporting it, the proposed response, the owner and the test. This preserves the link between listening and implementation. It also makes unresolved disagreements visible when different groups prefer different service arrangements.

For the recreation service, a concern about mobile verification should lead to a concrete decision about the assisted route. The team needs to establish which identity and booking checks are necessary, who can perform them and how the result is recorded. Do not collect additional personal information by default merely because the digital route used a particular field.

Illustrative concern-to-commitment record

Concern	Design commitment	Acceptance evidence
Valid assisted cases cannot complete	Provide an authorised route with appropriate service checks	Clerk completes realistic assisted scenarios
Customer does not know an alternative exists	Make the route discoverable at the point of difficulty	User can find and understand the next step
Correction request returns to the blocked flow	Give the receiving team authority to correct the booking	Challenge exercise reaches a changed outcome
Staff burden is hidden	Resource and monitor the assisted workload	Operating owner reviews demand and unresolved waiting

Assign a decision for concerns the team cannot fully address. The outcome might be a narrower service trial, additional research, a different design or deferral. Avoid recording “noted” as if it were a treatment. The sponsor should understand which affected-person issue remains, why it matters and what would support reconsideration.

Keep the consultation evidence proportionate. A summary can preserve the concern and response without identifying an individual or copying a sensitive account into every governance document. Where detailed evidence is necessary, provide controlled access and distinguish the participant's account from the team's interpretation of it.

Test the whole journey, including scarce places and corrections

A booking journey can appear successful because the interface returns a friendly confirmation, while the reservation is not actually recorded or the selected class is no longer available. Define completion using the service outcome: the appropriate reservation state is established, the person understands it and the next step is clear. This may require evidence from both the assistant and the booking process.

Test a place becoming unavailable during the interaction. The assistant should not imply that a seat is held if the underlying process has not established that state. Test a correction to the requested class, an interrupted assisted session and a person returning through another channel. These cases examine whether the service preserves intent and avoids duplicate or misleading outcomes.

Consider the timing of harm. A wrong recommendation corrected before reservation may have little lasting effect. A misleading confirmed-booking message can cause travel, missed participation or loss of a scarce place. The assessment should identify when intervention must occur to prevent those consequences rather than assuming every error is reversible through a later apology.

Use the findings to set trial boundaries and monitoring. The operating owner may need to review incomplete journeys, corrections, channel transfers and staff workload, not only successful conversational responses. Record how evidence will be collected without unnecessary personal detail and how a significant pattern reaches the responsible decision-maker.

Revisit the assessment after the bounded trial with actual operating evidence. The fictional twenty-four-case exercise is useful for exposing a design defect, but it cannot establish real-world demand, every access need or long-term outcomes. Compare observed use with the assumptions, ask whether the alternative route works at the actual workload and decide whether expansion is supported. This closes the loop between the assessment's promises and the service people experience.

Make the conclusion conditional on observed commitments

For each design commitment, record the owner, evidence and acceptance condition. “Offer human support” becomes a named team, accessible contact route, hours or capacity assumption, test and escalation arrangement. The responsibility map should make these obligations operable.

State the assessment outcome and unresolved limits: proceed within a boundary, repair first, narrow the use, defer pending evidence or choose an alternative. Do not average a serious unaddressed effect into a favourable overall score. Connect the conclusion to the classification and approval process.

Discuss an AI impact assessment with Open for AI using the service journey, affected groups and a real decision. The useful output is an evidence-backed set of design and operating commitments, with a clear explanation of who may benefit, who may be disadvantaged and what the organisation will do about it.