In 1993 the CEO of a paper mill in Portland, Maine came down to IBM Research and asked for help. Alok Aggarwal’s team told him the truth, which was that they had no idea how paper gets made. They spent three months learning, then built a scheduling system out of agents and meta-learning, words that would not become fashionable for another thirty years. The mill improved by about 1.5 percent, worth roughly nine million dollars a year. Nobody at IBM Global Services understood the thing well enough to run it, so the client ran it themselves for the next twenty-four years. He calls himself an accidental entrepreneur. Sixteen years at IBM Watson Research, then he opened IBM’s India Research Lab in 1998, then co-founded Evalueserve in 2001 and coined the term knowledge process outsourcing on the way to 4,500 people. Scry AI is the company he started to get back to building products. From the paper mill he took the four parameters he still judges every project by: is it timely, is it high quality, does it make money, is it robust. His view is that generative AI scores well on the first one and is terrible on the third. In this episode of Founders & Empanadas he sits down with Joshua Eidelman to argue something most of LinkedIn will resist: generative AI is completely real, and several of the companies building it will not survive the decade. His case is about funding, debt, and the revenue required to service both at a 15 percent return, and he has already moved his own date for the shakeup forward twice, most recently to September 2027, largely because Chinese labs are far more financially disciplined than American ones. Along the way, the nine months of zero revenue that ended with 9/11, why a product company needs three co-founders when he only has one of himself, and the reason COBOL is still running your bank. Highlights: - Nine months into Evalueserve with no revenue at all, and the two-month freeze after 9/11 that forced the pivot - Why nobody was selling to small and mid-sized companies in 2001, and how thirty-to-fifty-thousand-dollar projects got them profitable - Doubling revenue every year for four years, and the moment Deutsche Bank found them - The 2004 prediction about data management services that he made and then failed to act on, which he still calls his error - Person-to-person offshoring, called in 2008, now going by the name gig economy - Two co-founders for a services company, three for a product company, and why running solo makes him “handicapped” - The four parameters from the paper mill, and the one where AI keeps failing - Why 5 percent edge cases in a demo become 12 percent in production, and CFOs stop signing - COBOL was created in 1960 and still runs most banks, insurers, and government offices - 400 million jobs gone by 2050, and the population decline that could have us begging robots to work - Why he would rather build after a crash than during a boom Subscribe for more founder conversations over empanadas. New episodes on YouTube, Spotify, and Apple Podcasts. Timestamps: 00:00 Intro 01:50 The Mountain View billboards remind him of 1999 02:37 Johns Hopkins, then 16 years at IBM Research 03:52 The paper mill that ran on agents for 24 years 05:19 Founding Evalueserve and coining KPO 08:01 Nine months of zero revenue, then 9/11 09:10 The $30k projects that got them profitable 11:25 Calling person-to-person offshoring in 2008 16:21 Two co-founders for services, three for product 19:38 Why the big AI labs may not survive 22:40 The funding, debt, and 15 percent return math 24:29 The four parameters from the paper mill 26:03 Moving the shakeup date to September 2027 29:43 COBOL is still running your bank 31:43 400 million jobs gone by 2050 33:17 The population decline nobody prices in 35:04 His real worry is that we forget how to talk 36:11 Why he would rather build after a crash 38:08 The belief 99 percent of founders would disagree with
