What Matters
An AI agent is not inclusive simply because anyone can sign up. Inclusion means people can reach it, afford it, understand it, use it in their language, recover when it fails, and appeal a consequential decision to a human. DN’s Agentic Inclusion Standard turns those expectations into six deployment gates that can be tested before an agent is trusted at scale.
The six inclusion gates
| Gate | Minimum standard | What failure looks like | What to measure |
|---|---|---|---|
| 1. Access | Works on realistic devices, bandwidth and assistive technologies. | Only high-end users can use it reliably. | Device, bandwidth and accessibility success rates. |
| 2. Affordability | Cost is proportionate to user value and local purchasing power. | Usage becomes exclusionary after a few interactions. | Cost per successful task and regional affordability. |
| 3. Language | Core workflows work in the user’s language, not just chat. | Fluent conversation but broken tool use or structured output. | Task success by language. |
| 4. Understandability | Users can understand what the agent did, why, and what happens next. | Opaque decisions or hidden automation. | Explanation comprehension and correction rate. |
| 5. Human appeal | Users can challenge, correct or escalate consequential decisions. | No path to a person when the agent is wrong. | Escalation availability, time and resolution. |
| 6. Safe fallback | Failures degrade safely without trapping the user. | Blocked access, duplicate actions or unrecoverable state. | Recovery rate and exception resolution. |
DN Inclusion Readiness Score
Access is more than account creation
An agent can be technically available and still be inaccessible. Real access depends on device support, bandwidth, keyboard navigation, screen-reader compatibility, readable contrast, captioning, input alternatives and whether the experience works outside ideal desktop conditions.
Affordability is a systems metric
A cheap subscription can still produce an expensive outcome if the agent requires many retries, consumes large token volumes or fails often enough that a human must redo the work. Inclusion therefore needs cost per successful outcome, not just monthly price.
Language support must include execution
DN’s Multilingual Agent Gap Index showed why chat fluency is not enough. Inclusion requires the same tool accuracy, structured output, interface navigation and error recovery in the target language that the system achieves in English.
Related research: The Multilingual Agent Gap Index 2027.
Users need to understand what the agent did
Explanations should be useful, not decorative. A user should be able to answer: what did the agent do, what evidence or rule mattered, what changed, what remains uncertain, and how can I correct it?
Human appeal is a hard gate for consequential systems
Where an agent can affect money, access, employment, education, health, identity or another important interest, a user should have a meaningful route to challenge the result. “Contact support” is not enough if support cannot inspect the agent’s decision trail.
Safe fallback matters most when automation fails
An inclusive system should degrade safely. A failed agent should not lock a user out, duplicate a payment, erase state or create a dead end. Recovery procedures must work for people with limited technical knowledge, not only developers.
Pass / conditional / fail
| Band | Score | Deployment meaning |
|---|---|---|
| Pass | 85 to 100 | Strong inclusion controls, with continued monitoring required. |
| Conditional | 65 to 84 | Usable only with explicit remediation plan and human fallback. |
| Fail | Below 65 | Do not scale to high-impact use until critical gaps are fixed. |
Methodology and limitations
Version 1.0, 28 October 2026. DN synthesizes W3C accessibility principles, NIST AI risk-management guidance, World Bank AI-readiness constraints and UNESCO human-centred AI principles into an agent-specific deployment framework.
Limitations: inclusion is context-dependent. A public-benefit service should face a higher bar than an internal low-risk tool. Legal obligations also differ by country and sector.
Commercial disclosure: no vendor paid for inclusion or weighting in the standard.
Primary sources
- W3C, Web Content Accessibility Guidelines 2.2
- NIST, AI Risk Management Framework
- World Bank, Strengthening AI Foundations
- UNESCO, Recommendation on the Ethics of Artificial Intelligence
FAQ
What is inclusive AI?
AI designed so people can access, understand, afford, challenge and recover from the system across different languages, abilities and circumstances.
Is accessibility the same as inclusion?
No. Accessibility is one part of inclusion. Affordability, language, transparency, human appeal and safe recovery also matter.
When is human appeal essential?
It is especially important when an agent can affect money, access, employment, education, identity or other consequential interests.
Can an agent pass if it is only English-language?
Only in a narrowly scoped English-only deployment. A globally deployed system should validate core workflows in the languages it claims to support.
