Can AI replace user research and design thinking for Indian product teams?

Every product leader we speak to in Pune, Bengaluru and Mumbai is being asked the same question by a founder or a CFO: if ChatGPT can interview a hundred “users” in an afternoon, why are we still paying for discovery sprints? It is a fair question, and 2026 has finally produced enough evidence to answer it properly. This post sets out what AI does well inside a design thinking process, where the research says it fails, and why the failure modes matter more in India than almost anywhere else.

What can AI actually do inside a design thinking process?

Design thinking is usually taught as five stages: empathise, define, ideate, prototype and test. AI helps in every stage, but unevenly. Its strongest contribution is to the work that happens between conversations with users: transcribing, clustering, drafting and rendering. Its weakest contribution is the conversation itself.

Stage	Where AI helps today	What it cannot do
Empathise	Transcribe and translate interviews; draft discussion guides; find patterns across 40 transcripts in minutes	Replace the interview; notice what a user does not say; observe a shop floor or a classroom
Define	Cluster pain points; draft problem statements and personas from real data	Decide which problem is worth solving for your market
Ideate	Generate hundreds of variations; pull analogies from other industries	Judge which idea an Indian SME owner or a ZP teacher will actually adopt
Prototype	Produce clickable prototypes and copy in hours instead of days	Know whether the prototype makes sense on a ₹8,000 Android phone on a 2G patch
Test	Analyse usability recordings; summarise feedback; score task completion	Stand in for the test participant

The pattern is clear: AI compresses the time between user contacts, which means a team can afford more of them. The mistake is to use that time saving to have fewer.

How accurate are AI “synthetic users” compared with real people?

Synthetic users are LLM personas prompted to answer surveys or interviews as if they were a customer segment. In April 2026, Jim Lewis and Jeff Sauro of MeasuringU reviewed 12 peer-reviewed studies from 2023 onwards that compared LLM-based synthetic users with real humans across psychology experiments, surveys, social research and UX interviews. Their tally: 9 encouraging findings and 14 discouraging ones. Their summary was that results were “not universally bad, but not great”.

The discouraging findings share a shape. High-level averages often look plausible, but the variance is too narrow, subgroup results and regression weights diverge, and in one replication attempt only 3 of 14 classic social-science studies reproduced. In the single UX-interview study, 19 researchers recreated their own projects with GPT-4-Turbo and were initially surprised by how similar the narratives looked, then found the responses lacked contextual depth over several turns. Not one of the 12 studies found synthetic users matched human variance or qualitative depth.

Practitioners have reached the same conclusion. A May 2026 survey of 150 research professionals by User Interviews, as reported, found that 97% use AI somewhere in their workflow, yet only 8% regularly use synthetic-participant tools, 64% are sceptical or opposed, and just one respondent saw them as a standalone substitute for real people. The most common worries were accuracy (89%), stakeholders over-trusting AI findings (80%) and bias amplification (78%).

Why does this matter more for Indian product teams?

Because the accuracy problem is not evenly distributed. A World Values Survey benchmark cited in the same literature found that LLM answers tracked real populations best in wealthy, Western, English-speaking countries and worse elsewhere. The models learn from what is written on the internet, and the internet under-represents a kirana owner in Nashik, a Class 7 teacher in a Marathi-medium Zilla Parishad school, or a first-time UPI user in Tier-3 Bihar. Ask a synthetic persona how that teacher would react to a VR lesson and you get a fluent, confident answer that is grounded in nothing.

We have lived this. Abhigyaan, our VR learning platform, is deployed in 1,350 Maharashtra Zilla Parishad schools reaching 1.5M+ students. The decisions that made it work, from session length to how a teacher launches a lesson without a trained lab assistant, came from sitting in classrooms, not from prompting a model. We described the method in should an Indian startup run design thinking before building its MVP, and the economics in does design thinking deliver ROI for Indian product teams.

There is a second India-specific reason. Twenty-two scheduled languages, code-mixed speech and low-literacy interfaces are exactly the conditions under which the “flattening” the studies describe does the most damage. A model that averages away variance will average away the users you most need to design for.

Where should Indian teams use AI in discovery?

Generously, but on the processing side of the line. The uses below are the ones we now build into every engagement:

1. Multilingual transcription and translation: run interviews in Hindi, Marathi or Tamil and get English notes the same day, which removes the biggest excuse for interviewing only English speakers.
2. Synthesis across transcripts: cluster quotes, count recurring pain points and surface contradictions before the team reads a line.
3. Discussion-guide drafting and pilot testing: use a model as a rehearsal partner to tighten questions before the first real session.
4. Rapid prototyping: turn a sketch into a clickable flow in hours so that real users test something concrete on day three, not day ten.
5. Desk research: summarise regulation, competitor flows and public data as inputs to the define stage.
6. Hypothesis generation, clearly labelled: ask a model what a segment might say, write it down as a hypothesis, then go and check with five real people.

What does a hybrid AI-plus-design-thinking sprint look like?

Our five-day discovery sprint has not got shorter since 2024; it has got denser. The same five days now carry roughly twice the user contact because the admin around each contact has collapsed.

1. Day 1, Empathise: six to eight real interviews, in the user’s language; AI transcribes and translates overnight.
2. Day 2, Define: AI clusters the transcripts; the team argues over the clusters and writes the problem statement by hand.
3. Day 3, Ideate and prototype: AI generates variants and builds two clickable prototypes before lunch.
4. Day 4, Test: five real users on each prototype; AI scores task completion and pulls verbatims.
5. Day 5, Decide: the team presents evidence, not opinions, and commits to a build or a second loop.

The rule we apply: AI may touch anything a user said, but never say anything on a user’s behalf.

How do you stop AI from quietly replacing the user?

Three guardrails. First, label provenance: every quote in a research deck is marked as a real participant or a model output, because the 80% of researchers who fear stakeholders will over-trust AI findings are describing a real risk. Second, set a floor: no define-stage decision without a minimum number of real conversations, whatever the model says. Third, govern the data. Interview transcripts contain personal data, so under the Digital Personal Data Protection Act, 2023 you need a lawful basis to collect them and, under section 8, remain responsible for any AI vendor that processes them as a data processor. Tell participants which tools will touch their words, keep transcripts only as long as the purpose needs, and prefer tools that do not train on your data.

Bring evidence, not opinions, to your next product decision

Qodequay’s design thinking consulting runs discovery sprints for Indian product teams that combine real, multilingual user research with an AI-assisted synthesis and prototyping workflow. We are ISO 9001:2015 and ISO 27001:2022 certified, and we have shipped products into Indian classrooms, compliance teams and industrial training programmes. Book a 30-minute call to scope a sprint for your next release.

This article is for general information and is not legal advice. Confirm obligations against the official text of the Digital Personal Data Protection Act, 2023 and the DPDP Rules, 2025, and consult qualified counsel for your situation.