Launch HN: ScopeAI (YC W17) – Extract insights from customer conversations

77 points by iloveluce ↗ HN
I'm Luciano, co-founder and CTO of ScopeAI (https://www.getscopeai.com/). We’ve built a product that automates the process of extracting and communicating user insights to product and operation teams. To extract product insights, we integrate with support channels such as Zendesk, Intercom and desk.com and use NLP to automatically tag, categorize and cluster support tickets.

Customer support teams currently spend hours manually tagging customer support tickets to track trends in user feedback. The process is inefficient, lacks consistency and is reported retrospectively. This process typically fails to capture the granular insights requested by product and operation teams.

Natalie, our CEO is a former UX researcher. In that world, the process for extracting trends from user interviews was completely manual. It involved codifying the conversations and counting how frequently certain feedback was mentioned. It was definitely difficult to scale. We recognized that there needed to be a better way of extracting trends from unstructured data and started working on ScopeAI!

Some things we’ve learned/be happy to discuss further:

-Our process for extracting key phrases from tickets - currently done through a custom pipeline built using spaCy

-How we connect similar phrases - currently using a word2vec model trained on both GloVe vectors and text from tickets in our system

-How we assign broader categories and sentiment analysis using Tensor Flow

Here's an example of an insight we'd extract:

There were 67 requests for subscription cancellations for company x during the month of July:

• 24 requests “slow service”

• 19 requests “I have another account with y company”

• 8 requests “login issues”

Knowing this is really valuable for this company because they can make better decisions - in this case, making the software faster became a much higher priority.

Happy to answer questions and looking forward to hearing any feedback!

24 comments

[ 3.0 ms ] story [ 61.3 ms ] thread
Yelp, Google and other review aggregators do an interesting technique where they isolate significant phrases and intents from reviews and collect them into one report, and present it to the end user to help them make an educated decision. How does your technique compare? Are there significant differences in the objectives when catering to an internal audience (ticket management) instead of an external audience (public reviews)?
Great question! It’s similar in that we extract significant phrases and measure their frequency. Our differences:

• We combine similar significant phrases together. I.e. cancel subscription and end services would be clustered together.

• We track and display their distribution over time not just their total frequency. i.e. is it trending? Is the current downward trend significant?

• We ‘learn’ from each company the topics that interest them by both manually allowing them to subscribe to topics or mark them irrelevant as well as by automatically tracking click through rates and time spent.

As for the internal audience vs. external audience, we focus on the questions that are important to a product or operations team: how to reduce churn, feature requests, sentiment (what improves customer satisfaction) which is different from information valuable for consumers.

As an operator of a subscription service, this sounds valuable. I would be interested in an app where I could connect my Drift account and get some data on this.

I would also be interested in a SaaS that simply managed cancellations and collecting feedback there.

Also, interested in a SaaS that does the same except for tracking true chain of referral (tracks down the customer and makes them answer 'Where did you initially find us', 'What ultimately triggered the purchase'

Is there any integration for review channels on iOS/Android app stores? Google play console currently has some tools around review analysis but I don't believe the Apple app store does. This could be really helpful.
We currently don’t have iOS/Android app store integrations but they’re definitely on our roadmap! Many of our current customers already have their customer support channels all integrated, for example they integrate Zendesk with their Google app reviews so when we integrate with their main support channel i.e. Zendesk, we end up analyzing all of their customer conversations!
Awesome. Have recently started work on a well established product with a ton of data but it's painful to go through all these channels manually and tag it all. Ideally we'll get better at that over time but I know that we'll miss out on some of those new tags, or adding verbatims to tags easily.

Requested access but 404 on your Calendly link to book a call...

What is your revenue model? Are there costs to use the platform? Etc.
We have a monthly subscription revenue model based on volume of tickets analyzed (different tiers for different ranges of ticket volume).
Whats the ideal number tickets coming through to get valuable feedback?
The higher the volume of tickets the more precise our insights can be but as a general rule 100 tickets are required to gauge trends in feedback. For some companies this is an hour of customer requests and for other companies this could take a week to accumulate.
Sorry for the harsh words, but this looks like a solution looking for a problem. Any CEO of a company that's lucky enough to have customers contacting them would spend a not trivial amount of their time reading and replying to the contacts. For example, Jan Koum, Whatsapp founder said that they only started hiring dedicated customer service staff when they reach 150M users [0].

[0] https://youtu.be/8-pJa11YvCs?t=849

That’s a great point!

We completely agree that anyone making product decisions should spend a considerable amount of time listening and seeking customer feedback but we see ScopeAI as a tool that can make that process more efficient and accurate. Using algorithms for mining trends can prevent product decisions from being made with biases like Recency Effect[0] for example.

We also think ScopeAI can customer feedback more transparent and accessible to people across the company - this becomes more and more of an issue as companies scale and ticket volumes becomes unmanageable for people to manually sift through. For example it’s a lot harder for a CEO to read all customer service tickets if they receive 100 an hour!

[0] https://en.wikipedia.org/wiki/Serial_position_effect

WhatsApp is an unusual case; a consumer app with low ARPU, very much a "take it or leave it" proposition.

Most other businesses need more touch with the customer.

If you are at the "hacker and/or hustler" phase, efficiently using the time of the principals matters a lot; a founder could probably train a system like that in a week or so and use that to reply to tickets with less time spent and then hand it off to other hires like a "playbook".

I think that's the wrong take - currently consumer companies don't ask for much feedback because they simply couldn't read all of it. Tech like this is bringing about a revolution allowing companies to solicit more feedback than they can directly read, and yet understand via text analytics tools.
At Monzo (a bank), we have approximately 10,000 customer contacts a week.

This would be valuable to us.

You also have a significant advantage over other banks in that your customers can easily and securely communicate with you over text (I once tried to send a secure message to my "other bank" and gave up after ten minutes on their site.)

This is great for me as a customer, but I suspect it creates for you a greater per customer text volume to comb through than most other banks.

Anyway, looks like this might become a sale - hope it goes well for both sides.

(A very satisfied Monzo customer)

(comment deleted)
We were drowning in feedback at my old job, this would've been super useful. Can I customize the tags at all?
Yes we can automatically apply established tags to incoming support tickets and there is the option to delete and merge similar tags that ScopeAI generates and assigns to tickets as well!
Would this work with tickets where customers mix a bit of local language (spanish in our case) with English?
Generally homogeneity in language works better for the analysis (accuracy is higher) but we do capture tickets with some mixed languages. In this case, we mark the non-english word as an irrelevant entity instead of translating. Spanish support however is next on our roadmap of languages to support!
Congrats on the launch Luciano and Natalie! The product sounds incredibly helpful and I'm eager to try it out.

I was wondering- if customer support tickets are coming into an inbox (like in gmail), can this also be integrated?

Thanks!

This is an interesting topic, at my first job I've developed some tool to help our customer support, it's a topic which is quite interesting for me.

On what data do you train your models? Do you train them individually for each customer with their own data, or do you take the data of all your customers and train "universal" model?

Another question about the broader categories, are they defined by you? I guess you're doing some supervised learning. Is it possible for the customer to add own categories?