Putting AI at the Service of Inclusion

Why inclusion is a right, a design discipline and a condition of service performance.

CX for AI · Season 1 · S01A09

An AI-enabled service can perform well on average and still exclude some of the people it is intended to serve.

Overall accuracy may conceal that some voices are recognised less reliably. A high completion rate may hide users who cannot enter the journey, understand the instructions or recover when something goes wrong. Successful automation may coexist with additional human intervention elsewhere.

The dashboard remains positive because these experiences are diluted within an average—or absent from the measurement altogether.

Inclusion is a right before it is a business case. People should not have to demonstrate commercial value to deserve access, autonomy and fair treatment.

But exclusion also has measurable consequences. It restricts participation, increases abandonment, creates avoidable support demand and prevents services from delivering their intended outcomes.

Inclusive design therefore belongs at the centre of both responsible AI and service performance.

A right—and a performance condition

Approximately 100 million people in the EU live with a disability, and the European Accessibility Act has applied to covered products and services since 28 June 2025. European Commission

Compliance establishes an essential minimum. Inclusion asks a broader question: can people genuinely participate in the service and achieve the outcome it was created to provide?

A technically accessible interface may still use confusing language. An alternative channel may exist but require disproportionate effort. A human hand-off may be available but lose all the information the customer has already provided.

Access to the door is not the same as successful participation.

The distinction matters for AI because automated services can scale both access and exclusion. A barrier built into a conventional process may affect one interaction at a time. A barrier built into an AI-enabled service can be repeated rapidly, consistently and with little visibility.

When AI expands participation

AI can help people interact through speech, text, images or other inputs. It can describe visual information, adapt interfaces and translate between different modes of communication.

Jennifer Wexton’s experience offers a powerful example. After progressive supranuclear palsy affected her ability to speak, AI recreated her voice from previous recordings, allowing her to continue participating publicly in a voice recognisably her own. C-SPAN

The significance was not the novelty of synthetic speech. It was the restoration of agency.

This is AI at the service of inclusion: technology expanding the ways in which someone can participate rather than requiring them to conform to a standard interface.

When average performance conceals exclusion

The same technology can create barriers when it is designed around a narrow understanding of its users.

A speech-recognition system may achieve strong overall accuracy while performing less reliably for particular accents or speech patterns. An identity-verification journey may work smoothly for most customers while repeatedly rejecting people whose circumstances were underrepresented during design and testing.

Aggregate performance does not invalidate these failures. It can conceal them.

The relevant questions are therefore not reduced to accuracy of the system. They must also include: for whom does it work less well? What happens to those people? Can they recover?

Recovery is particularly important. No system will handle every circumstance successfully. An inclusive service must allow people to question an outcome, correct inaccurate information, change channel or reach someone able to help—without having to restart the entire journey.

Designing from lived experience

I encountered the human importance of adaptive design during my time as a trustee at Lifelites, a UK charity providing assistive and inclusive technology for children using palliative-care services.

That experience reinforced a principle: people should not have to adapt themselves to the assumptions built into a standard interface.

Technology creates meaningful access when the interaction can adapt to different abilities, circumstances and forms of expression. Empathy may initiate that work, but effective inclusion requires the discipline to understand real contexts, test with the people concerned and redesign around what is learned.

I saw the same distinction from an operational perspective while supporting adoption of public digital services used by farmers for administrative processes, funding applications and visibility of project progress.

These services were not AI-enabled, but they demonstrated that making a journey available online does not automatically make it effectively accessible.

Adoption depended on whether people understood the process, could relate it to their circumstances, received practical support and knew how to proceed when the standard journey did not work for them.

Co-design and direct engagement helped teams treat adoption as a service outcome—not simply as the consequence of releasing digital functionality.

AI does not change that principle; it raises the stakes because automated interactions can make exclusion faster, less visible and more difficult to challenge.

Making exclusion visible

Inclusion changes what organisations need to measure.

These measures require care. Examining performance by cohort or access need must be lawful, appropriate and proportionate. Quantitative evidence should also be combined with accessibility testing, operational evidence and direct research with affected people.

Metrics alone cannot represent those who were unable to enter the service. Nor can they explain why a difference exists.

The objective is not to produce a single inclusion score. It is to identify where apparently successful services produce unequal experiences or outcomes—and then act on that evidence.

Four disciplines for inclusive AI

Four disciplines help translate inclusion from principle into practice.

1. Make AI behaviour understandable

People should know when AI is involved, what it can do, which information influences its output and how they can question or correct the result.

Explanation should help someone act. Disclosure alone is not enough.

2. Provide equivalent pathways

Different modes of interaction should not lead to an inferior, slower or disconnected service.

A human route may be necessary, but it should preserve context and enable the person to continue rather than begin again.

3. Design with the people concerned

Lived experience reveals barriers that design teams, datasets and standard testing scenarios frequently miss.

Co-design should not be limited to a final accessibility review. It should influence the problem definition, journey, operating model, testing and measurement.

4. Measure differences, not only averages

Overall performance should be examined alongside variations in access, completion, errors, intervention and recovery.

A service cannot be understood fully if its strongest results conceal who is being left behind.

Inclusion as part of AI governance

This performance discipline is increasingly reflected in regulation.

The EU AI Act classifies specified AI systems used in areas such as education, employment and access to essential public and private services as high-risk because they may pose serious risks to health, safety or fundamental rights. The relevant high-risk requirements are due to apply from December 2027 and include risk management, data quality, traceability, human oversight and post-market monitoring. European Commission

Regulation does not replace inclusive design. It reinforces the need to understand how systems perform for the people affected by them.

That understanding cannot be created entirely before launch. Organisations must continue examining who uses the service, who encounters difficulty, where human intervention becomes necessary and whether people can recover from an unsuccessful outcome.

Performance beyond the average user

Inclusion changes the unit by which performance is judged.

It asks organisations to look beyond the average user and examine whether the population a service is intended to support can enter, participate, achieve an outcome and recover from failure.

AI can remove barriers that conventional services have left in place. It can also reproduce those barriers at greater speed and scale.

The difference lies in whose experience shapes the design, which variations the service can accommodate and whether unequal outcomes remain visible after launch.

A high-performing service is not one that works efficiently for the expected user.

It is one that does not require people to become the expected user before they can participate.

Publication note: Part of CX for AI, a series exploring the infrastructure of trust, demand and growth in AI-mediated markets. It revisits and substantially updates “Putting AI at the Service of Inclusion”, first published on Medium in October 2025.