Sourabh Gupta

The through-line

Sourabh Gupta's decisions, from founding through to the collections business Skit.ai runs today, come out of a single framework he and his co-founder wrote on a whiteboard in Bangalore before they had a product: pick the technology first, then go looking for the problem. "We decided we want to work on the edge of technology," he says, "rather than you know just picking like an let's say an operational problem and then you know going deeper into that" [2]. The reasoning was competitive rather than romantic. Two 22-year-olds with no work history had no advantage over anyone experienced, except in a field so new that experience did not exist yet: "no one can help anyone uh so because it's all about the paper that came out yesterday" [2]. That is how they arrived at deep learning in 2016, then voice, then call centers, then debt collection [1][2].

What has shifted is the second half of the sentence. Early on the technology was the point and the application was a monetization question [2]; now the preoccupation is the unglamorous distance between a working model and a delivered outcome, which he locates in compliance, edge cases, human handoffs and whether the customer's own end users are treated decently [1]. The company he describes today is around 200 employees with over 100 customers across India, North America and Canada [1].

On working at the edge of technology

The framework was arrived at deliberately and is still in use. Fifteen days after graduation they moved to Bangalore with no idea what to build, and spent the first months whiteboarding daily, looking at US startups and at the Flipkart-follows-Amazon pattern, asking what framework should apply [2]. AI and blockchain were both candidates; AI won quickly, and within AI, vision, video and voice were weighed before voice was chosen [2]. The next step was to find something inside voice that did not exist: speech-to-text in Indian languages, where Google was the only company with anything and it did not work [2]. Only after building that did they ask where voice could be monetized, and call centers at large Indian banks were the answer [2]. That vernacular-first bet was the company's original identity, and the opportunity he was discussing publicly was the one where the overwhelming majority of Indian internet users prefer to consume content in their local language [10]. He also took part in the wider conversation about where AI could be applied as businesses reorganized under the pandemic [9].

The choice of call centers came from both ends of the phone line. Everyone has called one and "the experience is terrible," waiting an unreasonable time for a basic query, but when they spoke to the people running those centers, "they were equally frustrated their own reasons," and that two-sided frustration is what he calls the niche they went deep into [1].

On starting with no experience at all, and why that helped

He wanted to build a company from the first year of college, having grown up in a small northern Indian city where "95% of people around have some sort of a business" [1]. He tried to drop out in his third year and his parents refused, not believing he could build a company [1]. He found his co-founder Akshi in a college canteen, someone he knew from entrepreneurship-cell events rather than a close friend, and the offer was blunt: there is no money for the first two or three years, borrow from your parents [2]. Akshi turned down a paying job to accept, to both sets of parents' disappointment [2].

Not only was it a first startup, neither founder had ever held a job, which produced gaps he describes cheerfully: at twenty people they discovered "there's a concept of HR policies" because they had not known it existed [1]. The same ignorance cut the other way. They did not know sales cycles were long or that received wisdom said a 100k ARR deal was impossible in India; the first two deals were $300,000 annualized each with two of India's largest banks, won against Google in one case and Nuance in the other, while he was still "just learning how to wear a suit" and had walked into early bank meetings in a Pink Floyd t-shirt [2]. His account of it: "we just didn't know and we were just following like first principles" [2]. The runway questions resolved faster than expected too. They raised a first seed round from Accel Capital within nine months of graduating, enough to stop borrowing from parents and pay themselves roughly 50,000 a month, though serious money took a couple more years [2]. What they took to that raise was two months of reading and implementing papers, some demos, and a friend with a fintech company willing to do a pilot [2].

On hiring, and on onboarding as the real work

The early team was assembled around product, not go-to-market: they were "very clear we did not want to hire for GTM" until the product was right, and of the first twenty people, eighteen were writing code [2]. The first hires were friends from college, including one unhappy in his job and one whose US visa had been rejected, both still with the company; nobody was over 25, and the criterion was people ready to hustle and "ready to solve problems they've not seen before" [2]. Senior, experienced hiring started two or three years in, triggered by raising a few million dollars and the sudden arrival of board meetings, numbers to present, and the realization that hitting numbers requires systems and processes they did not have [2].

That transition is where he says they made "maybe all the mistakes possible in the world" [2], and his conclusion is unfashionable: since people become the only leverage as you scale, "you have to be incredibly hands-on" [2]. The specific failure mode he names is treating a senior hire as self-sufficient and giving them an hour a week; instead, "you have to spend like 100 hours in the first few months," after which the time can be reduced [2]. He handles the obvious objection, that an experienced person resents being managed by a younger founder, by settling it during the interview rather than after: the point is not hierarchy but alignment, and the arrangement he asks for is that even when both sides disagree, "we are aligned that we are not aligned on something," so it gets talked through [2]. This is now a rule rather than an instinct: every new hire, senior or junior, has a named owner of their onboarding who spends a large amount of structured time with them, with topics defined in advance rather than a drink and a chat [2].

On the debt problem, and why it became personal

The current business is a narrower cut of the same call center problem: debt management, whether credit card, medical or otherwise, and specifically how businesses engage with people who owe money [1]. He frames it as a gap between two aggrieved parties, the consumer or patient who is frustrated, full of questions and possibly unable to pay, and the collection agencies who "get a bad rep" [1]. It became personal after a two-hour appendicitis surgery at NYU with no overnight stay produced an initial invoice of $97,000, which he has so far negotiated down to $19,000 with the case still open [1]. His point is that the problem begins before collection: the first question is how the amount can possibly be that, and what its components are [1]. Talking to others convinced him the experience is near universal, which is why he treats it as "a real problem and it's a massive problem" now solvable given where AI has got to [1].

On customer centricity as the hard problem

Asked how a bot competes with a person when a caller goes off script, he says this is where the company spends most of its time, and reframes the difficulty: "the real problem or the hard problem is to actually be very customer ccentric or end user centric while trying to collect money" [1]. The argument he makes to buyers is self-interested rather than moral. A bank or a hospital wants those users back; he went to NYU for surgery and will still need routine checkups [1]. In practice this shows up as design defaults: if a caller asks for an agent or a representative, the call is transferred immediately, and if it is 3am with nobody available, the system captures the details and commits to calling back when offices open [1]. The conversations themselves are meant to work in "a very non-judgmental way," understanding someone's financial situation and helping with payment plans and settlements while answering the questions people actually have, such as the effect on their credit score and how much to pay this month versus next [1].

On how the product actually runs

Deployment is omni-channel with voice as the primary medium, plus SMS and email, and it works both outbound and inbound, so a person calling their bank or agency about a debt will likely reach a bot [1]. It starts before anything is overdue, with reminders ahead of the due date on the logic that a customer who can afford to pay should not be paying avoidable interest simply because they forgot, and then follows bucket-specific strategies as accounts age through 0 to 30 and 30 to 60 days [1]. Businesses are, he notes, very often willing to settle for less than the full debt or to stagger payment over one, two or three years, and the bot can hold those conversations [1]. Buyers are heads of collections in larger organizations, or owners and CXOs at companies whose main business is collections, and the addressable set is "anyone and everyone who has collections in their organization," including banks, hospitals, auto finance companies and captives [1].

He also insists the automation is partial by design: "we also believe AI is not at a stage where like oh you know 90% of people are not needed," so the work is "a very good marriage of humans with AI," which in turn makes the handoff interface a core engineering problem, including whether the human picking up the transfer has full context [1].

On being an outcome business, not a software business

Internally the line he repeats is that "people look at us as a software business but we are actually an outcomedriven business" [1]. The test he uses is the review meeting. Nobody runs an MBR or QBR with Slack asking how many messages came in and what the ROI was; his customers' reviews are exactly that conversation, how much am I paying you and how much are you collecting [1]. Results are contextual, depending on debt type and workflow and conversational complexity, but he cites agents moving from around 100 minutes to 70 or 75, with the freed portion either split between bot and agent or handled entirely by the bot [1]. On job displacement he is deliberately measured: it is hard to attribute headcount effects, because businesses have more work than they can get to, so freed agents end up on more complex accounts and larger debt amounts, while the cost saving from work AI now does is real [1].

On LLMs and the long arc of AI

He resists the story that everything changed with ChatGPT. There was AI of some sort in 2010 and before, deep learning arrived around 2014 or 2015, a wave of startups followed, and now LLMs have brought another wave of companies [1]. His assessment is that performance keeps improving with each new class of model and he is "very optimists about LLMs," which have improved his own product's performance, but this is not a dramatic break in the world: "it's not like things were not happening um before" [1].

On what enterprise deployment actually demands

His biggest lesson from being an early mover is the gap between demonstration and delivery: "it's one thing to you know create a demo and say that you know you have a bot uh it's a whole different thing to actually deliver value to the customer" [1]. The difference lives in the details and in a long tail of small edge cases, and above all in regulation. Collections is heavily compliant, which means constraints on whom you can call, when you can call them, how many times, what may and may not be said on the call, and what responses are permitted once the person answers in a particular way [1].

Takeaways

• Choose a field where inexperience is not a handicap: they picked deep learning in 2016 precisely because "no one can help anyone" and the advantage went to whoever read and implemented the newest paper first [2].
• Technology first, problem second: speech-to-text in Indian languages came before the decision to monetize it through bank call centers [2][10].
• Not knowing what was supposedly impossible let them close two $300,000 annualized bank deals in India, beating Google and Nuance, while advisers were saying 100k deals could not be done [2].
• Hire for product before go-to-market: eighteen of the first twenty employees wrote code, and nobody was over 25 [2].
• Onboarding is where senior hires succeed or fail; spend roughly 100 hours in the first few months rather than an hour a week, and give every hire a named owner of their onboarding [2].
• Treat alignment as an explicit agreement, including being "aligned that we are not aligned" on a given point [2].
• Sell outcomes, not seats: quarterly reviews with customers are about how much was collected against what was paid, unlike a software vendor such as Slack [1].
• Keep humans in the loop by design, transferring on request, capturing 3am calls for callback, and passing full context to the agent who picks up [1].
• The hard part of collections AI is being end-user centric while collecting money, and the hardest engineering is compliance and edge cases, not the demo [1].