Does AI actually work in Hindi? A field report from twelve deployments.

"Does AI work in Hindi?" is the question every owner asks within the first ten minutes of our first call. The honest answer is: it depends what you mean by "work." We have shipped twelve Hindi-first or Hindi-bilingual deployments — across coaching institutes, hospitality, healthcare clinics, and a CA practice with a Marwari-speaking client base. Some have been outstanding. Two had to be partly rolled back. Here is what we learned, with as little hedging as possible.

What Hindi AI does well in 2026

The frontier models — Anthropic Claude (Sonnet, Opus), GPT-4 class, Gemini 2 — handle Hindi well enough today that it is no longer a meaningful constraint for most business workflows. Specifically:

• Triage and routing.Reading a Hindi WhatsApp message, classifying intent, extracting key entities (name, batch, fee question), and routing to the right human or automated reply — excellent.
• Grounded retrieval.Answering a Hindi question from a known Hindi (or English) source document, with citations — excellent. This is what powers our concierge bots.
• Form & document extraction.Pulling fields out of a scanned Hindi invoice or handwritten form — very good with the right vision model.
• Summarisation.Compressing a long Hindi conversation or document into bullet points — good, with caveats around register.
• Conversational drafting.Writing a polite Hindi reply that sounds professional — good, but tone needs tuning per business.

Where it fails, and how it fails

Three patterns we now treat as red flags during scoping.

Code-switching is the rule, not the exception

Indian users do not speak "Hindi" or "English." They speak whatever serves the moment. Within a single WhatsApp message, you'll see Devanagari, Roman-script Hindi, English, and the occasional regional word. A typical message: "Sir maine 10th standard pass kiya hai, NEET ke liye consult karna hai, aap WhatsApp pe call kar sakte ho?"

Most generic chatbots break on this. The model handles it fine if it's prompted to expect it — but the failure mode when it doesn't is silent: the bot picks one language and replies, and the user feels misunderstood without quite knowing why.

Regional dialects are a hard ceiling

Marwari, Bhojpuri, Haryanvi, Awadhi, Magahi — these are not "Hindi with an accent." They are linguistically distinct, and frontier models handle them weakly. We have shipped exactly one Marwari-aware deployment (the Alwar resort), and even there we explicitly bound the bot's promise to "tries Marwari, escalates fast if it loses the thread."

If a vendor promises you "fluent Marwari" or "perfect Bhojpuri" — they are either lying or have not deployed at scale. The honest position is: dialects work as a polite gesture, not as a primary medium.

High-stakes judgment in Hindi is not safe yet

We do not deploy Hindi-AI for medical triage beyond appointment-booking. We do not deploy it for legal advice. We do not deploy it for emotionally-loaded customer-service situations (bereavement, complaint escalation). The model can produce a fluent Hindi response — but the calibration of "how confident am I in this answer?" is shakier in Hindi than in English. The wrong polite-sounding answer is worse than no answer.

How we draw the boundary

For every Hindi-AI deployment, we run a five-step boundary exercise during scoping:

1. List every conversation type.From three days of real production data, not what the client thinks happens.
2. Classify each as bot-safe, human-required, or grey-zone.Bot-safe = retrievable answer with low ambiguity. Human-required = anything emotional, financial-advisory, or non-recoverable.
3. Set escalation triggers.Specific phrases, sentiment shifts, repeated misunderstandings — all explicit.
4. Native-speaker review of 200+ shadow conversations before public cutover. Always.
5. Quarterly accuracy audit with the client's senior staff for the first year.

What this means for your business

If you are evaluating a Hindi WhatsApp bot, three questions to ask any vendor:

• Show me the escalation triggers.Specific words and phrases. If they have none, walk away.
• How do you handle code-switching mid-message?If the answer is "the model handles it" without specifics — they have not stress-tested it.
• Who reviewed the templates?If the answer is "we did" without naming a native speaker, the bot will sound foreign to your customers.

Hindi AI is at the point where it is genuinely useful for the right tasks, and genuinely unsafe for the wrong ones. The line between the two is what a good practice gets paid to draw.