AI Can Prep the Meeting in 30 Seconds. It Still Can't Sign the Contract.

Two numbers frame the conversation. First, sellers spend 75% of their time on work that is not customer-facing: weekly reports, CRM updates, meeting prep. Put another way, three of every four dollars a company pays a salesperson buy admin labor. Second, before an important customer call, Gong founder Amit Bendov used to pull five people from the account team into a briefing, wait for a Q&A document, and burn about five hours of highly paid time. Gong now prepares him automatically. He reads for 30 seconds and asks if something is missing.

Those figures come from a May 2025 episode of Sequoia Capital’s Training Data, hosted by Sonya Huang and Pat Grady. Their guest is Amit Bendov, CEO and founder of Gong, a company that started in 2015 with meeting recordings and is now one of the best-known AI applications in sales. Huang opens with a sharp question: will AI take over the SDR, the outbound hunter who cold-calls and qualifies leads?

Amit’s answer is crisp. Completely, in the near future, with current technology? No. A lot of BDR work will be done by AI, and he thinks that is a good thing; the picture is not binary. Complex outbound SDRs, in particular, are not going to be fully replaced for years. The follow-up is sharper still: will CRM even exist in five years, once agents are running around doing sales work? He is sure CRM will still be there in 20 years. Too much is built into it, the way mainframes outlived better technology. What will change is the center of gravity. "We're moving from a CRM-centric world to a more AI-centric world." CRM manages customers and their data. It does not manage operations, activities, or how to engage buyers. That last stretch is AI’s.

On the human side, he is more uncertain than on the software side. The role of software is starting to come into focus; the role of humans is less clear, and new jobs will appear that nobody can picture yet. Prompt engineer did not exist three years ago. SDR did not exist 30 years ago; it is a modern invention and might go away. Anything based on transformers, he says, cannot replace humans wherever there is accountability, decision-making, or creativity. AI can already review a contract, redline the hot spots, even suggest corrections. Nobody should let it sign the contract, send the contract, or send a proposal. There will still be a lawyer in that loop. The shift he expects is toward humans who are augmented, using AI as a tool for a long list of tasks, while that 75% of non-selling activity gets replaced.

Plenty of people still think Gong began as a transcription company. Amit says the vision was never about transcripts. At his previous company he was already recording sales calls and paying a transcription service. The real pain was that he never had time to read them. He has a bit of ADHD, he says, and is not the person to listen or read at great length. What he wanted was something that would take all that conversation and turn it into structured data. Transcription was just a step in the pipeline: capture audio and video, transcribe, identify topics and themes, then use the result.

They started with meetings because that was the big black hole. Other customer activity left some trail. What happened inside the meeting, nobody knew, and nobody was doing anything about it. The first technical decision was just as revealing. Gong had about $6 million in the bank. Amit wanted to build a transcription engine because it felt core. His co-founder Eilon, "the smarter half," said no: do not spend that money for a 3% accuracy bump until you know what customers will actually buy. They duct-taped a web API onto an application, shipped something that worked, and watched competitors pour money into real-time transcription and other cool technology that never made a real difference for customers.

Hindsight made Eilon look right. Once Gong raised more money, the team could no longer live on Nuance, which was expensive and not great. They hired some of the best speech people in Israel, built their own engine, and got better quality at a much lower cost. Today they mix Amazon services and Whisper with in-house models for languages where the public options still fall short. If a vendor does a better, cheaper job, they will switch. Transcription, Amit says, is like a car battery: build it if you must, buy it if someone else is better. Gong’s competence is the AI application layer, not the low-level engine.

Huang notes that many peers at the time thought from the technology out, while Gong thought from the customer back. Real-time transcription sounded like the obvious next feature. Amit held the line. To this day he says the technology still does not really work, and the value is questionable. What customers wanted were coaching workflows and the ability to count how often a competitor came up after the fact. Nobody else was doing that.

Grady remembers Gong as one of many, then almost overnight as one of one. Amit pushes back on the nostalgia. He thinks they have more competition now, and he tells the team that more success will bring more of it. When they started they thought they were alone, scratching Amit’s own itch. Then they found two others, Chorus and ExecVision; neither is an independent company today. Now there are dozens, and he hopes for hundreds in a few years. CRM has hundreds of vendors. If the market is growing, that is opportunity. If you look around and see no competitors, you are in the wrong market.

The move after they discovered rivals mattered more than the discovery. The other company was better funded, a little ahead, with cooler technology. Amit still has the all-hands recording from that small-company moment. By year-end, he told them, there was only one goal that mattered: be number one, or nothing else counts. "In SaaS there is no second prize," at least none he is interested in. They even quantified winning: media share, LinkedIn, social, twice all the others combined. They hit it. First it was neck and neck, then a small win, then a 5% or 10% gap. Salespeople got more confident, customers started following the crowd, and the gap widened. The strong get stronger, but only after you take that first bite.

People have been predicting Gong’s death ever since. Outreach and Salesloft would eat them. Before they raised, Google and Amazon would eat them. Amit’s answer is a Beatles metaphor, which he admits is a function of his age: every two years the company should feel like a new band playing a different kind of music. If Gong still only recorded meetings, it would already be dead, because anybody can do that. Competition, in his telling, is not football, a zero-sum fight over ball possession. It is an Easter egg hunt. Do not watch where the other kids are running. Find the eggs nobody else sees, and take as many as you can.

Amit calls generative AI a game changer, and he has been waiting for it since the beginning. In 2015 and 2016 Gong used the five levels of self-driving as a product map: level one is alerts, level five is fully autonomous. The vision was an application that runs a revenue organization. You tell it the revenue goal; it breaks the goal down and helps everyone from leaders to individual sellers execute. That was science fiction then. The components still are not all there. What changed is that the team can now look at what is actually reliable, stand behind a customer experience, and ship it. Call summaries are the clearest example. In 2016 Amit hired a TV news editor, handed over a recording, and asked for a 30-second newsreel from a 60-minute call. After about 20 attempts he gave up. Summaries come out biased. Today a model does it easily.

The market surprise was almost as large as the technical one. Gong has always been an AI application company, but for years AI was third-level messaging. The buyers are CROs and other business users. Talk about AI and it sounded like scary science fiction, so you had to double-click a couple of times to find it. After ChatGPT, the calls reversed. CEOs were being asked on earnings calls how they planned to increase sales productivity with AI, and they turned around and asked Gong if it had something. Boards that used to fear the word now demand a productivity story.

On the numbers, some customers have seen as much as 60% more sales capacity. A person who handled 50 accounts can now handle 100, mostly by making tasks redundant. You no longer need as many internal "what happened, what am I going to do" conversations, because everything is visible. Amit’s own meeting prep is the personal version of that 75% statistic: five hours of highly paid people down to 30 seconds of reading. They are not at level-five driving yet, so he thinks the remaining opportunity is larger still.

Different pieces of Gong sit at different levels. Level four is the smart default: Gong drafts the follow-up email, the human clicks send. Level five is the system sending on your behalf, and that will take time. Cold outbound, the task everyone wants to automate, is a poor fit for the current state of AI. It is hard for people. If it goes wrong, it damages the brand. Amit cites a Cursor support agent changing a privacy policy and airline chatbot incidents as reasons he would not let AI talk to his customers except in confined cases like support tickets. Role-play with a virtual avatar is the opposite: if the practice call is imperfect, you are a little annoyed, not embarrassed in front of a buyer. Customer-facing actions stay at level four until the technology can be 100% trusted.

Grady presses the 100% line. Human reps are not 100% either, so why is that the bar for software? Amit says yes and no. People do hold AI to a higher standard. A Tesla crash is a headline; human drivers crash far more often and nobody writes it up. Early Gong actually hid the transcript. Accuracy was about 70%, three words in ten wrong, and a smart advisor who had shipped other products told them not to show it. They let people search, showed some highlights, and users were happy. Showing the raw text would have been an awful experience.

The other half of the answer is that ease of use buys forgiveness. Gong summaries are maybe 80 to 85% correct, and almost nobody complains, because they are fast. Amit reaches for a teenage hi-fi metaphor: people once insisted on proper stereo, then happily listened to a single cheap speaker because it was easier. Huang admits she does not know what a hi-fi system is. He laughs that he recently installed major speakers in the basement and his wife wanted to kill him.

On agents, he thinks they are real and that expectations are inflated. They are not taking your job anytime soon. They can take the parts you do not like. Gong’s approach is task-specific: no general "do anything" agent, no writing code. "Brief me for an account" is its own trained agent, with parameters, built into the workflow, and they have dozens of them. That, he believes, is what works in the enterprise. General-purpose agents are hard to nail because they can go in all sorts of directions. The line between simple automation and an agent is also blurrier than the marketing suggests.

Have reasoning models unlocked new capabilities? Not yet. The hard line is truthfulness: not 99%, 100%. Send a proposal that is wrong on price or over-discounted in 1% of cases and you have a hallucination incident. Commission calculations are the same: get them wrong and it is very bad. Reasoning makes more tasks viable. It is still a quantitative improvement, not a qualitative jump to something you can fully trust. He agrees with Yann LeCun and physicist David Deutsch that the statistical nature of transformers may never get there, the way an analog computer is not a digital one with error correction. Gong can already update every CRM opportunity record automatically. Some companies still want the seller to click yes, because otherwise they can blame the AI. Accountability is as real as the model error rate. That is why he does not expect AI to sign a million-dollar contract in the next three to five years. Virtual avatars selling simple products in two or three years? Yes.

The sales team of the future, Amit says, will change gradually rather than in a software-engineering-style collapse of headcount. Transactional work goes first: buying an iPhone, upgrading a cellular plan. Buying a house, a mortgage, or enterprise software is more complicated. What he does not see continuing is clerical form-filling, or the traditional ritual of manual forecasting. Sales stages themselves are a 50-year-old management tool invented for visibility. AI can run that process more smartly. The BDR role may not stick around, and not only because software replaces it. If sellers have the right tools, they can do more of their own development. BDR was invented for hypergrowth companies. Most companies in the world never had it.

That raises a packaging problem. Gong captures value today mostly through seats, and making each seller more productive is a force against seat growth. Amit is not sure there will be fewer sellers; the economy grows, and products nobody has imagined yet will need to be sold. They will start adding consumption and usage pricing, and this year they are shipping an orchestration product that fits that model. When Gong started, buyers wanted per-seat pricing because CRM was the anchor, and the company did not want to innovate on product and pricing at once. The Series B pitch already said AI would be bigger than cloud: cloud was an IT revolution, AI is a work revolution. Gong never saw itself as a software company so much as BPO, with a market ten times larger. Typical sales tech is about $1,000 a seat. A seller costs $200,000 a year; save 70% and you have $140,000; take 10% of that and it is still a bargain for the CFO, even at $1,000 a month.

Huang asks about the sales-tech winter. Early 2023 was a perfect storm. The market tanked, tech customers cut budgets and consolidated vendors, and Gong’s suite was not complete enough to win those bake-offs. Customers said they loved the product and still left. The early hunger from 2016 had gone dull. The company almost came to a halt. Amit’s response had four parts: keep the foot on the gas and hire while others laid off; ship Forecast and Engage at speed to fill out the suite; rebuild the sales motion around a bulletproof business case a CFO would approve, not "buy Gong because it is cool"; and run a lot of experiments, knowing some would take two or three quarters to show anything. The financial discipline underneath it was simple. They raised a lot of money and never burned a lot. Investors asked why they were still raising. The answer was a rainy day they knew would come. Then ChatGPT arrived. People predicted it would kill Gong. It became the biggest tailwind instead. By taping, Gong was on its seventh consecutive quarter of accelerating growth, the latest faster than the one before.

Rapid fire, he uses ChatGPT at least 15 times a day. Advice for founders building at the AI application layer: keep your eyes on the customer, talk to customers, and do not believe investors or influencers. Be the fastest thing that works now, not the thing that is true in two years. The book he recommends is David Deutsch’s *The Beginning of Infinity*, a harder read that is not only about AI but about what comes next; it changed how he thinks. Million-dollar contracts signed by AI will need something beyond transformers, something closer to a digital computer you can rely on 100%. He is happy to put that prediction on the calendar in three to five years and see if he is wrong.