AI trust · Personalisation · India

Personal, But On My Terms

What Indian users need before they let an AI assistant use their personal data

A global technology company · 2026 · Anonymised

Company, product and feature names have been anonymised for confidentiality.

At a glance

  • Client: A global technology company (anonymised)
  • Year: 2026
  • My role: Staff UX Researcher; principal researcher
  • Method: 12 moderated remote usability sessions, about 75 minutes each
  • Participants: 12 AI-fluent users across nine Indian cities, aged 19 to 39
  • Focus: Comprehension, trust, control and cultural fit in a personalisation feature

Context

A major AI assistant was introducing a personalisation feature. With the user's permission, it could draw on past conversations and connected apps, such as email, calendar, cloud files, photos and video history, to tailor its answers.

The promise is obvious: an assistant that knows you. So is the risk. Personalisation asks people to hand over some of the most intimate data they have, and it only works if they trust what happens next. The team needed to understand how Indian users would experience that trade-off from end to end.

The question

Would people opt in, and if they did, would they understand what they had agreed to? Where did trust build, and where did it break? And did the experience fit the way Indian users actually live with their data?

Under those sat the question I cared about most: when personalisation fails, is it because people resist it, or because the product hasn't earned it?

My role

I was the principal researcher on the study. I owned it end to end: research design, recruitment and operations, moderating all 12 sessions, analysis and synthesis, and the final recommendations.

Approach

  1. 01

    End-to-end usability sessions

    12 moderated remote sessions over screen share, about 75 minutes each, covering the full journey: opt-in, settings, the home screen, and chat.

  2. 02

    Live opt-in where possible

    four participants enabled the feature live in the session. Others walked through the flow from screens, and several already had the feature on.

  3. 03

    Participant-authored prompts

    each person wrote three prompts from their own life, so we tested personalisation on real needs, not scripted tasks. That produced 35 prompts across eight kinds of use, from recommendations to self-reflection.

  4. 04

    Trust and control probes

    questions on transparency, privacy, correction and what users would want switched off.

  5. 05

    India-context probes

    a dedicated lens on cultural fit, local defaults and how Indian users store and think about their data.

Participants came from nine cities across four regions, including Tier 2 and Tier 3 cities such as Dhanbad, Guwahati, Kottayam and Amaravati, alongside the metros. All were heavy users of several AI tools.

What we found

  • 10 of 12Hesitated or refused photo access
  • 9 of 12Opened the app list before agreeing
  • 6 of 12Credited chat history alone
  • 12 of 12Had at least one strong win
  1. 1

    People weren't resisting personalisation. They were resisting the unknown.

    10 of 12 hesitated or refused when photo access came up, mostly because the disclosure mentioned using data to train AI models. But 9 of 12 opened the app list before agreeing, and most then connected nearly everything. They wanted to look before they agreed.

  2. 2

    In India, cloud storage is a document vault.

    For several participants, their cloud drive held copies of Aadhaar, PAN and voter ID. Bundling it with email and calendar turned a reasonable request into an alarming one, and one participant refused the whole bundle because of it. Consent controls designed around "work files" didn't match how people actually use the service.

  3. 3

    Invisible value is no value.

    6 of 12 credited the personalisation to chat history alone, even with every app connected. One said: "The apps have not added anything in this response." If users can't see a source contributing, the case for granting access disappears.

  4. 4

    Trust broke on weighting, not accuracy.

    The failures followed a pattern. The assistant trusted an IP address over a city the user had stated many times, long-term history over the last two weeks, one passing question over hundreds of consistent ones, and global defaults over Indian context. As one participant put it: "Just because I asked one time about scripting does not mean that I'm a scripter."

  5. 5

    Accurate isn't the same as useful.

    A profile summary could describe someone correctly and still feel, in one participant's words, "like something an astrologist would say." Recognition alone didn't create value.

  6. 6

    When it worked, it was magic.

    All 12 participants had at least one strong win in their own prompts. The best moments drew on sources people could trace, like an exact figure pulled from hundreds of spreadsheets, or a recommendation tied to a video watched the night before.

The signal hierarchy users expected
Should winOften won instead
  1. StatedInferred
  2. RecentStale
  3. DurableOne-off
  4. LocalDefault

What we recommended

Make the value concrete at opt-in. Name the actual apps, and show one example of how a connection changes an answer, so people agree to an outcome rather than a mechanism.
Split bundled consent and add granular control. Let people exclude their cloud drive while keeping email and calendar, address the photo-training concern directly, and explore a temporary pause, data-type exclusions and a view of what the assistant already knows.
Show your sources. Add light inline cues when a connected app contributes to an answer, so users can see the value they are paying for with their data.
Fix the signal hierarchy. Prioritise stated, recent, durable and local signals over inferred, stale, one-off and default ones.
Make Indian context the default when signals support it. That means India-anchored suggestions and content defaults, and labels tested for local meaning.

Impact

This study was one in an ongoing series of research rounds that shaped the feature as it developed, so its influence is cumulative and doesn't trace to a single launch. What it gave the team was a prioritised set of quick fixes and longer-term changes across opt-in, controls, transparency and ranking. It reframed the problem from "users are worried about privacy" to "users will trade data for value they can see and control," which is a far more actionable brief. And it brought India-specific realities, such as cloud storage as an identity vault, into the product conversation.

The hard parts

  • Testing a feature that was still moving. Only four of twelve participants could opt in live, and the screens were still changing. I designed the sessions to work across live, partly live and walkthrough conditions without losing comparability.
  • Separating real failures from thin data. When the assistant felt shallow, was it broken, or did it simply know too little yet? Untangling weighting failures from missing data needed careful coding of every response against what the participant had actually shared.
  • Being honest about the sample. The group skewed young, male and Android, reflecting the testing pool and the market. I flagged where findings were elicited rather than spontaneous, and where they were directional, so the team knew how far to push each one.

Reflection

This study confirmed something I keep finding in trust research: people are more willing than we assume, and less forgiving than we hope. They will hand over remarkably personal data, but only when they can see what it buys them and take it back when they want.

It also showed why cultural context isn't a localisation layer added at the end. Whether a cloud drive holds work files or your Aadhaar card changes what consent means. Getting personalisation right in India starts with understanding how people actually live with their data.