Social impact — PACT Standard assessment axis

What this axis protects.

Social impact assessment protects people and communities from harms that disappear when evaluation stops at an individual prediction. A system can treat each case consistently yet still reduce access to a public service, intensify work, weaken professional judgement, exclude non-digital users, concentrate supplier power or shift resources away from the people who need them most.

This axis assesses intended benefits and adverse effects across the full lifecycle. It includes direct, indirect and cumulative impacts; the scale, scope, duration and reversibility of change; and the distribution of benefits and burdens. Benefits do not cancel violations of rights. They must be tested alongside, not traded against, the minimum protections covered by the other axes.

The right question is not only “does the system work?” It is “what kind of institution and community does its use create?”

Assess six fields of change.

Access and inclusion. Effects on availability, eligibility, affordability, language, disability, digital exclusion, waiting time and the ability to use a non-digital route.
Work and skills. Job quality, autonomy, workload, surveillance, pace, deskilling, reskilling, task allocation, worker voice and conditions in the wider AI value chain.
Wellbeing and relationships. Stress, dignity, trust, isolation, dependency, exposure to harmful content and changes in relationships between staff, institutions and communities.
Institutional power. Accountability, professional discretion, contestability, procurement dependency, concentration of expertise and the ability to withdraw or change supplier.
Democracy and information. Participation, public debate, pluralism, access to reliable information, civic space and unequal ability to influence how a system is governed.
Environment and resources. Energy, water, hardware, raw materials, emissions, waste, local infrastructure demand and whether a smaller or non-AI method can achieve the same public value.

Where social impact assessment fails.

Technology as the starting answer. The organisation compares models but never compares AI with process reform, additional staff, a simpler rule or no intervention.
Benefits counted, burdens described. Efficiency is quantified while time, exclusion, stress, deskilling and environmental use remain anecdotal.
Consultation after commitment. Affected people are invited to react after procurement, budget and success measures are fixed.
Only direct users considered. Families, frontline workers, non-users, communities, suppliers and people represented in data remain outside the stakeholder map.
A short pilot hides cumulative effects. Testing cannot reveal dependency, feedback loops, labour displacement or changes in institutional behaviour that appear at scale.
Efficiency captured by the institution. Time saved by automation increases case volume or monitoring rather than improving service quality, access or worker conditions.
No withdrawal scenario. The organisation cannot estimate what happens to people, work and records if the tool, supplier or funding disappears.

Assessment questions.

What public or organisational problem is being addressed, for whom, and what evidence shows that an AI system is a proportionate response?
What are the realistic alternatives, including process change, additional human capacity, a simpler technology or not proceeding?
Who receives the intended benefit, who carries cost or risk, who is absent from the process and whose behaviour or data makes the system possible?
How may access, work, autonomy, skills, wellbeing, trust, democratic participation and environmental demand change at pilot and full scale?
Which impacts are severe, widespread, long-lasting or difficult to reverse, even if their probability is uncertain?
How have affected people, workers and representative organisations shaped the purpose, success measures, controls and conditions for withdrawal?
Which indicators can reveal benefits, burdens and unintended effects over time, and who has authority to respond?
How will the organisation provide remedy, support adaptation and restore services or work if the system is restricted or removed?

Evidence to collect.

Problem and alternatives: A documented need, baseline, theory of change and comparison of AI, non-AI and no-action options across benefit, cost, rights, uncertainty and reversibility.
Stakeholder map: Direct and indirect users, affected non-users, workers, communities, public-interest groups, suppliers and future generations, including influence and vulnerability.
Engagement record: Who participated, how barriers were removed, what was heard, which decisions changed, what did not change and why, with feedback returned to participants.
Impact register: Positive and adverse impacts assessed for scale, scope, duration, reversibility, likelihood and distribution, with cumulative and system-level effects included.
Measurement plan: Baseline and indicators for access, service quality, work, wellbeing, trust, environmental demand and distribution of outcomes, combining quantitative and qualitative evidence.
Mitigation and transition: Owners, resources, timelines, remedy, worker and community support, supplier exit, data return, service continuity and criteria for restriction or withdrawal.

Decision thresholds.

Proceed: The need and proportionality are evidenced; affected stakeholders shaped the plan; benefits and burdens are measured; severe and cumulative impacts have governance, mitigation and remedy.
Proceed with conditions: Uncertain effects are tested through a reversible, time-limited and independently reviewed deployment with protected alternatives and no exposure to irreversible harm.
Stop or redesign: A simpler alternative provides the benefit with less harm; essential access or worker rights are weakened without remedy; burdens fall disproportionately on vulnerable groups; stakeholder evidence is absent; or major impacts are irreversible and ungovernable.

Participation must be able to change the decision. Record which people can influence the purpose, design, deployment conditions and stop criteria. Listening sessions without decision rights are communications activity, not impact governance.

Lifecycle checks.

Before first use

Establish a baseline, compare alternatives, engage affected people and workers, assess rights and cumulative impacts, and set measurable conditions for scale, continuation and withdrawal.

During operation

Measure distribution of benefits and burdens alongside technical performance. Maintain ongoing routes for worker, community and affected-person evidence, and publish proportionate findings where public interest requires it.

When conditions change

Repeat the assessment when deployment scales, tasks or staffing change, the supplier or model changes, a new population is included, environmental demand grows or evidence shows institutional dependency and feedback loops.

Sources and further reading.

Primary and institutional sources reviewed for this working document. Legal applicability must be assessed for the specific system and context.