AI-enabled speech automation
I believe NLP, speech-to-text and text-to-speech have all advanced to a stage where AI-enabled speech automation can really kick-off in India. While the applications for these are immense, I want to focus on only one use-case: call-center automation.
The problem: Let’s talk about inbound call-centers first. Managing customer service through these inbound call-centers is a costly affair. In consumer-tech companies, these costs might typically start at INR 4–5 per order or could go as high as INR 10–15 per order — oftentimes wiping out the entire margin and skewing the unit economics of the company. Apart from the high cost of training and manpower, this issue is magnified by the need to correctly predict demand, so as to balance between high-wait times and underutilized agent capacity. Not to mention, the added cost of underutilization due to hiring for “peak volumes”.
Companies have typically tried to solve this by (a) making the customer go-through a series of frustrating selections either on the app or IVR (b) making chat-bots do the job
While chat-bots work well for a certain set of customers, they are particularly ineffective with “vernacular customers” for two reasons
(1) Most chat-bots don’t provide services in multiple languages
(2) Even if they do, anyone who’s used a traditional keypad to type in another language will tell you — it is extremely inconvenient. So much so, that a large number of people use whatsapp by sending each other voice messages in vernacular languages, because they’re easier to compose.
Given all these factors, and the fact that most contact center employees also leverage standard scripts and flowcharts in their interactions, automating contact-center calls seems like a problem worth solving.
Size of the market: While many reports peg the Indian call-center market at upwards of $10Bn, let us see if we can size a particular segment of the market. Instead of answering the question “how big is the market”, let’s answer an easier one “is the market big enough?”.
Now for the purposes of VC funding, let’s assume a $100Mn revenue potential is “big enough”.
In the table below, I’ve added the orders-per-day for some of the top consumer tech companies (through public sources).
Now I’m assuming a conservative number of Rs 4 per order ($0.05) as call-center cost. This brings the total current revenue potential to $100Mn. Assuming only 25%, or Rs 1 ($0.015) per order is the value that can be captured through these companies, we have visibility on an addressable revenue potential for at least $25Mn. With this coming from just a handful of clients, this number would garner a lot of interest. But it is safe to say that after onboarding Banking and Financial Services clients, $100Mn looks very achievable.
Business Models & valuation: I see a SaaS model work here, with a fixed “per month” or “per interaction” (or hybrid) cost structure. Given the upfront fixed costs involved in cleaning terabytes of training data and customizing the product, Go-to-market will need to be with 1–2 large companies with sufficient call-center cost outlay. I see one of the consumer-tech companies (probably from the table above) to the be the first adopters of this — due to faster sales cycles, higher amount of co-creativity and sufficient pressure to manage that cost. This would be followed by moving “upmarket” and into BFSI/ IT services.
While the former is heavily dependent on the founder’s hustle & grit, building a strong sales function and adapting it as customer profile changes (as in most SaaS companies) will be key to future success.
Let’s also do a simple back-of-the envelope calculation for valuation for a company that has captured the $25Mn revenue in steady state. The big assumption here will be the size of the sales force and data science/ product teams — something I built through some linkedIn & website hunting. Secondly, applying a thumb rule of Sales & Product costs being ~40% of total costs for SaaS companies, we have a rough P&L.
Assuming a PE ratio of 20, we get a steady state valuation of ~$250 Mn. Now in order to make minimum 30% yield over 3 years, current valuation would become ~$120 Mn.
Competition & Moats: There are a good number of Indian companies trying to attack this space, a few notable ones being qualyon, vernacular.ai, yellow messenger, gnani.ai. Of course, many international players are trying this as well. While “data” might be the obvious moat given the industry, I feel a more nuanced view of moats will be in niches — two specific axes being
(1) A niche like “vernacular language” might give an advantage over global competitors
(2) Developing a vertical product for a large-enough industry might give it some temporary defensibility
Future view: While we covered “inbound” contact center automation here, there is a tremendous potential for “outbound” as well. Your next call offering you the best credit card might actually be a bot, and might even do a better job this time. The future of this market will be owner by vertical players catering to the large customers, and horizontal players offering off-the-shelf or mass-customizable products to the SMEs (once they have perfected the art of frugal sales and collections). Current contact-center giants might also try entering in this space, with a distinct advantage of having vast amounts of training data and existing sales channels already at their disposal. We should expect a few acquisitions also on those lines 2–3 years hence.
Overall, this looks like a very exciting industry — some dealflow has already happened over the past few months, but we should expect more.
