Blog
“From Information to Action”: Shakul Sonker Presents I-Stem's Field Lessons on Assistive AI at Empower 2026, IIT Delhi
Shakul Sonker, co-founder and COO of I-Stem, today presented the paper “From Information to Action: Field Lessons from Multimodal Assistive AI in India” at Empower 2026, the 9th Annual Assistive Technology Conference at IIT Delhi. The paper, co-authored with Kartik Sawhney, co-founder and CEO of I-Stem, draws on I-Stem’s experience of running Nclude between September 2024 and November 2025.
The question behind the paper
Assistive technology has made a great deal of information reachable. Acting on it is a different matter. A screen reader can read a job description while the application form behind it stays inaccessible. A benefits scheme can be explained clearly and still need another person to finish the transaction.
Generative AI can now read interfaces and attempt multi-step tasks. The paper asks what changes when assistive AI moves from giving information to executing tasks, and whether user independence should be measured separately from system autonomy.
Nclude in the field
Nclude is one service available on web, WhatsApp and IVR, so it can be used over an ordinary phone call. Its features include career discovery, government scheme information and task automation, which covers job applications, forms, ride booking and food ordering. Registered users grew from 20,404 in March 2025 to 64,549 in November 2025, more than three times.
What the evaluation found
In November 2025, all 64,549 registered users were invited to a voluntary survey. 2,462 responded, and 85 percent of them were blind or low vision. All outcomes are self-reported and are shares of all respondents, not success rates for the feature.
1,313 people (53.3 percent) completed a previously inaccessible task through AI alone or AI with escalation to Nclude staff.
543 people (22.1 percent) completed one through AI alone.
1,791 people (72.7 percent) reached government scheme information.
The gap of 31.3 points is 770 people. They did something previously closed to them through a supported pathway inside the service, with the decisions staying with them, rather than relying on a relative or neighbour. If success were scored only on what AI finishes unaided, these 770 people would disappear from the record. This is the paper’s central point: user independence and system autonomy need separate measurement.
What was hard
Once the system acted across outside websites, forms and logins, small errors added up. Human support became part of the design, as long as it arrived predictably and the user stayed in control. Some users trusted the AI to do more than it could, while others would not let it act. Translation alone did not achieve localisation, so dialects had to be tested. IVR and WhatsApp lowered the cost of getting started, and disability organisations carried onboarding and escalation.
Implications for the field
Shakul closed with three recommendations for anyone running AI-enabled assistive technology:
Measure user independence separately from system autonomy.
Budget for failure, with an escalation path that is itself accessible.
Test channels and localisation with users, and treat community partners as part of the infrastructure.
He also called for future evaluations to record uptake, AI-only success, escalation-assisted success and failed attempts separately.
For more on the paper, write to shakul@inclusivestem.org.



