What the research says.
How Telybot is built rests on published research about language, speed and trustworthy AI. Here is what each study found, what we built because of it, and where to read it.
- CSA Research · 2020Can’t Read, Won’t Buy
- KPMG in India and Google · 2017Indian Languages – Defining India’s Internet
- ACL Anthology · 2014English–Hindi code-mixing on Facebook
- NeurIPS 2020 · EMNLP 2021Answers grounded in retrieved text
- Harvard Business Review · 2011The Short Life of Online Sales Leads
People buy in their own language
In a survey of 8,709 consumers in 29 countries, 76% of online shoppers said they prefer to buy products with information in their native language, and 40% said they will never buy from websites in other languages.
What we built Telybot replies in the language each customer writes in — 30 languages, Hindi and Hinglish included.
CSA Research, “Can’t Read, Won’t Buy – B2C”, July 2020. Read the source
India’s internet speaks Indian languages
Indian-language internet users overtook English users in 2016: 234 million against 175 million. The study projected 536 million Indian-language users by 2021, against 199 million English users.
What we built Hindi, Bengali, Tamil, Telugu, Kannada, Malayalam, Marathi, Gujarati, Punjabi and Urdu are all among the bot’s 30 languages.
KPMG in India and Google, “Indian Languages – Defining India’s Internet”, April 2017. Read the source
People mix languages as they type
A study of English–Hindi posts on Facebook measured how much people mix the two languages. It found that most mixed words sit on a continuum between borrowing and mixing, with non-standard spellings and transliteration that automatic processing has to handle.
What we built A question typed in Hinglish is answered in Hinglish. Telybot’s language work began in the Keyda keyboard, used by millions of Indians.
Bali, Sharma, Choudhury and Vyas, “‘I am borrowing ya mixing?’ An Analysis of English-Hindi Code Mixing in Facebook”, First Workshop on Computational Approaches to Code Switching, 2014. Read the source
Answers grounded in documents are more factual
Models that write answers from retrieved documents produce more specific, diverse and factual language than models answering from memory alone. In chatbots, the same approach substantially reduces knowledge hallucination — confident statements that are not true — as verified by human evaluations.
What we built Telybot answers only from your content. When nothing in it matches, it says so and hands over instead of guessing.
Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”, NeurIPS 2020. Read the source
Shuster et al., “Retrieval Augmentation Reduces Hallucination in Conversation”, Findings of EMNLP 2021. Read the source
Online enquiries go cold fast
Research into how companies handle potential customers’ online enquiries concluded that most companies are not responding nearly fast enough. Online leads, as the title says, have a short life.
What we built Telybot answers the moment a question arrives, at 9:40pm just as at 11am, and a welcome flow can take the customer’s name and number while they are still interested.
Oldroyd, McElheran and Elkington, “The Short Life of Online Sales Leads”, Harvard Business Review, March 2011. Read the source
How we check our own work
Every release runs more than 1,700 automated checks against a real server. Among them:
- A question its content does not cover is logged as unanswered, not answered.
- Once you answer it, the same question is answered from your words.
- A reply from a saved answer is never counted against your plan.
- One business can never read another business’s knowledge or chats.
- The bot is told never to invent, shorten or rewrite a link.
Found an answer it got wrong? Report it
Read them yourself
Figures are quoted as each source publishes them. We did not carry out these studies.
- CSA Research, “Can’t Read, Won’t Buy – B2C”, July 2020
- KPMG in India and Google, “Indian Languages – Defining India’s Internet”, April 2017
- Bali, Sharma, Choudhury and Vyas, “‘I am borrowing ya mixing?’ An Analysis of English-Hindi Code Mixing in Facebook”, 2014
- Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”, NeurIPS 2020
- Shuster et al., “Retrieval Augmentation Reduces Hallucination in Conversation”, Findings of EMNLP 2021
- Oldroyd, McElheran and Elkington, “The Short Life of Online Sales Leads”, Harvard Business Review, March 2011
Bring your website.
Leave with an assistant.
Free during early access. About fifteen minutes from sign-in to your first tested answer.
Create your assistant