Valuation of $9.5 billion: Interview with Gusto co-founder on returning to the front lines to write code and create a disruptive "AI co-founder" for logistics.

Gusto Eddie Kim: AI is turning software companies back into startup teams of five people If we only understand Eddie Kim's recent story as: "A CTO who hasn't written code for a decade is writing code again with Claude Code." We would seriously underestimate this matter. What is truly worth paying attention to is not: Eddie can code again. But rather: Why does a large software company with an annual revenue exceeding $1 billion and serving over 500,000 small businesses need to have its co-founder return to the front lines, leading a small team of only five people, to develop products as if it were the first day of the startup? By 2026, Gusto had crossed the $1 billion mark in actual revenue over the past 12 months, serving over 500,000 small businesses; its employee stock trading corresponding to a valuation of about $9.3 billion in 2025, while earlier funding rounds had valuations reaching about $9.5 billion. So strictly speaking, the more accurate statement today is not "current valuation $9.5 billion," but rather this is a mature tech company valued at around $9 billion with revenues exceeding $1 billion. Yet in such a highly mature company, Eddie Kim did something very "unlike a big company CTO": He started coding again. Ultimately, a core team of only 5 people, consisting of 3 engineers, 1 designer, and Eddie himself, built an entirely new first-level product line from scratch in just 10 weeks: Gusto Cofounder. In the process, there was almost no: Traditional PM; Long-term PRD; Complex Jira; Traditional Figma → Engineering handoff; A lot of meetings; Long-term Roadmap. They relied more on: Claude Code; Real-time prototypes; Perma-Zoom; Customer feedback; And a very short feedback loop of "write it out—try it—if it doesn't work, throw it away." The truly important aspect of this matter is: It may indicate that the organizational structure of software companies is undergoing a very profound change. For the past twenty years, the software industry has been solving: "How to enable more engineers to collaborate?" The new question in the AI era may become: When one person already has the execution capability of five or ten people, do we still need so many layers of coordination? This is the real point worth studying about Eddie Kim's return to the front lines. ──────────────── 1. The story of Gusto itself comes from a typical small business "logistics black hole" Eddie Kim founded Gusto not because: Payroll is sexy. On the contrary. Payroll, taxes, benefits, employee onboarding, and compliance are perhaps the least sexy types of work in the entrepreneurial world. But: The less sexy, more repetitive, and more necessary the work is, the more commercial value it often has. Because the boss cannot avoid it. Payroll cannot: "I'll send it next week when I have time." Taxes cannot: "I'm too busy this year, so I won't report." Employee benefits cannot: "Let's put it on hold for now." These tasks belong to: Mission-critical Back Office. Critical backend work. ──────────────── 2. What Eddie saw as a child was not a medical issue, but an SMB productivity issue His parents ran a medical business. What really impressed him was not just the doctors seeing patients. But rather the vast invisible labor behind a small business: Paying salaries; Scheduling; Taxes; Employee benefits; Insurance; Reimbursements; Various administrative tasks. This is the reality for many small business owners in America. A chef who opens a restaurant quickly finds that he not only has to cook. He also has to: Hire; Calculate salaries; Manage insurance; Handle taxes; Arrange employees. Thus: Founder → Chief Everything Officer. The boss ends up spending a lot of time doing: Work he didn't start the business for in the first place. Gusto's business opportunity comes from here. ──────────────── 3. This is also why "small business SaaS" is a huge market Large enterprises can have: HR departments; Payroll teams; Legal; Finance; IT. What about a spa with 15 employees? The boss is: HR. The boss is: Payroll. The boss is: Operations. So large enterprises can solve complexity through: People. Small businesses must solve complexity through: Software. This is the basic economics of Gusto. ──────────────── 4. But AI has pushed this logic forward a step The traditional SaaS answer is: Give the boss a better tool. The AI Agent answer may become: Don't make the boss learn the tool, just get the work done directly. This is the real difference between two generations of software. Traditional Gusto: Helps you manage Payroll. Gusto Cofounder: Attempts to directly replace you: Run Payroll. Gusto officially now defines Cofounder as an "always-on AI assistant" that can proactively identify problems, execute repetitive tasks, and interact with the boss via SMS, Slack, or Web. This is not a UI upgrade. But rather: Product Category Shift. ──────────────── 5. Software is truly beginning to shift from "tools" to "employees" The past business logic of software companies: Human + Software → Work. Humans operate software. Complete work. Agentic Software: Human → Goal. AI + Software → Work. Humans only provide: Goals; Rules; Approvals. The intermediate process is completed by machines. This means: Software is moving from interface to labor. ──────────────── 6. This may be the most important change in SaaS history Traditional SaaS sells: Seat. For example: One employee at $50 per month. AI Agents are more like selling: Outcome. Not: "Here’s a Payroll Dashboard." But rather: "Payroll has been calculated, anomalies have been flagged, and you need to confirm here." The product value shifts from: Access To: Completion. This is also why AI SaaS may ultimately penetrate: Labor Budget, And not just: IT Budget. ──────────────── 7. Eddie's personal experience is very interesting because he has actually experienced "one-person companies" long ago Before Gusto, he had worked on Android applications. That was an extremely unique time in the early smartphone era. The Android operating system was not mature. Many basic functions that seem obvious today did not exist at that time. Thus independent developers could fill these system gaps: Develop small tools; Sell directly in app stores. For example: Finding cars; Screenshots; Various system tools. Essentially belonging to: OS Gap Arbitrage. What the operating system lacks, Developers fill. ──────────────── 8. Why can this type of business allow one person to make a lot of money? Because the App Store / Android Market brought about a huge economic change: Distribution Cost → Near Zero. In the past, after writing software: You needed channels; Agents; CDs; Retail. After mobile app stores: One programmer Can directly reach: Global users. So one person suddenly has: Global Distribution Leverage. This is the super leverage of the mobile internet era. ──────────────── 9. But Eddie understood very early on the biggest risk of this type of business: platforms will absorb features Today, screenshots are an independent app. Next year: Android supports it itself. The company goes to zero. This is called: Platform Absorption Risk. All startups built on large platform feature gaps must face this. Browser plugins are like this. ChatGPT Wrappers are like this. Shopify Plugins are like this. If your only value is: "The platform doesn't have this button yet." Then when the platform adds a feature: The company is gone. ──────────────── 10. This is a lesson that today's AI entrepreneurs especially need to remember Currently, many AI startups are essentially: Model Gap Arbitrage. OpenAI hasn't done it yet. Anthropic hasn't done it yet. Google hasn't done it yet. So: Startups do it. The question is: What happens after the next generation model updates? So entrepreneurs really need to ask: If the foundational model provides my core functionality for free tomorrow, what do I have left? Customers? Data? Workflows? Brand? Distribution? Regulation? System integration? If the answer is: Nothing. Then it is no different from a "screenshot app" in 2010. ──────────────── 11. Why is Gusto much larger than those small tools? Because it has entered the System of Record Payroll is not: A small function. It connects: Employee identity; Bank accounts; Taxes; Benefits; Compliance; Payroll records; Historical data. This makes Gusto gradually become: System of Record. And System of Record has a huge advantage: Context. For AI to help businesses work, It must first know: Who the business is; Who the employees are; When payroll is issued; What has happened in the past; What the rules are. Gusto already has this data. ──────────────── 12. So the biggest advantage of Gusto Cofounder may not be the model, but rather "it already knows your company" This is exactly what Gusto officially emphasizes: Cofounder does not need users to tell it the company structure from scratch, because it already knows: - Team; - Payroll cycle; - Benefits; - Compliance calendar. This means: The competitive advantage of AI products is increasingly shifting from: Model Intelligence to: Context Ownership. Models can be purchased by everyone. Enterprise context cannot. ──────────────── Thirteen, this could be the biggest counter-weapon for all traditional SaaS in the future. Salesforce has: Customer data. Workday has: Employee data. ServiceNow has: IT workflows. Gusto has: Payroll + People Data. QuickBooks has: Financial data. Therefore, AI does not necessarily destroy these companies first. If they successfully transform, they may instead leverage: Existing Context to become the operating system for Agents. The real danger lies with those: Without core data; Without key workflows; Just a pretty UI of SaaS. ──────────────── Fourteen, why must Eddie return to coding after ten years of not writing code? Because he realized: You cannot understand the changes in technical paradigms through PPT. This is the most thought-provoking point for technical managers throughout the interview. A CTO can: Look at team weekly reports; Look at the Roadmap; Listen to model demonstrations; Look at Benchmarks. But: Until you personally work with Claude Code for a few hours, you may still not know: What is actually happening in software development. ──────────────── Fifteen, this is similar to CEOs in the 1990s not personally going online. Imagine 1996. A media CEO says: "I don’t use the internet, but my team reports internet trends to me weekly." Can he truly understand the internet? It’s hard. Today, AI may be in a similar stage. If a tech leader has never: Personally used a Coding Agent; Built an Agent; Designed a Workflow; Made models call tools; Conducted Evals, then their judgment on AI can easily remain at: Second-hand Knowledge. Second-hand cognition. ──────────────── Sixteen, during a technological revolution, the most dangerous thing for leaders is being too far from the means of production. In normal times: Management does not need to do everything personally. This is the inevitable result of organizational scaling. But when the technical paradigm changes: The management abstraction layer formed over the past decade may suddenly become distorted. Dashboards show: Engineering Velocity. But the way underlying engineers work has completely changed. At this time: Leadership must re: Touch the Metal. Engage with real technology. ──────────────── Seventeen, the biggest significance of Eddie returning to coding is not to prove that CTOs should also code every day. Not all CTOs should return to being full-time programmers. The real principle is: Those who decide the technical strategy must have first-hand experience with new production tools. This is similar to factory management. You don’t necessarily have to operate the machines every day. But if the production line upgrades from: Manual lathes to: Automated robots, managers must truly understand: What the new machines mean. ──────────────── Eighteen, the biggest organizational impact of AI Coding may not be "engineers are 30% faster." This is the easiest place to underestimate it. If it’s just: Originally writing 100 lines a day. Now writing 130 lines. That’s just: A Productivity Tool. But what Eddie’s team is showing is another change: Role Boundary Collapse. Role boundaries are disappearing. Designers are starting to submit production code. Technical leaders can independently create complete prototypes. Engineers are involved in product design. This is not: Increased speed. But rather: Organizational structure change. ──────────────── Nineteen, why did past software organizations produce PMs, Designers, Frontend, Backend, QA? Because each professional capability: Has a high learning cost. Product managers write requirements. Designers create Figma. Engineers translate designs into code. QA tests again. There is a: Translation Cost between each step. Requirements translate into designs. Designs translate into code. Code translates into tests. AI begins to reduce these translation steps. ──────────────── Twenty, every "handoff" is an organizational cost. Product managers: Write 20-page PRD. Designers: Understand. Create Figma. Engineers: Re-understand. Write code. Information is constantly lost in the process. This is called: Coordination Tax. When the cost of actually writing code is very high, these processes are valuable. Because: You must plan clearly in advance. But if: AI reduces implementation costs by 10 times, another method emerges: Don’t write a 20-page document explaining the product, just build the product directly. ──────────────── Twenty-one, this is where Eddie’s so-called "Trash-can Method" is truly revolutionary. In the past: Writing code was very expensive. So: First discuss; First approve; First write documents. Avoid: Wasting engineering resources. After AI: Code is getting cheaper. So teams can: Write first. See. Don’t like it? Just throw it away. How I AI summarizes it as Eddie’s "trash-can method": directly establish a complete PR, then use real implementations for product discussions, and if it’s not worth continuing, close the PR. The real change here is: Cost of Experimentation. The cost of experimentation plummets. ──────────────── Twenty-two, when code becomes cheap, the economic value of "planning" itself will decline. Traditional large software development: Making a wrong direction is very expensive. 100 engineers working for half a year: The cost is huge. So: Planning is important. But if 5 people + AI: Can produce a real version in three days, then why do you need to: Hold two weeks of meetings? Write thirty pages of documents? You can directly: Build to Think. Think through building. This is completely different from the past: Think → Plan → Build ──────────────── Twenty-three, this may change the entire product management profession. This does not mean PMs will disappear. Rather, the value of PMs will shift from: Writing Tickets; Maintaining Jira; Synchronizing status; Translating requirements to: Judging what is worth doing. That is: Customer Insight; Prioritization; Taste; Strategy. AI is easiest to automate: The coordination layer of project management. The hardest to replace: Real product judgment. ──────────────── Twenty-four, top PMs in the future may need to be more like Builders. It’s okay not to write traditional code. But they must be able to: Prototype themselves; Call models; Conduct real tests; Connect data; Demonstrate to users. Otherwise: All thinking must wait for engineers to translate. In the AI era: This is a significant speed loss. So: Prototype Literacy is likely to become a basic ability for future product managers. ──────────────── Twenty-five, why is the designer case from Gusto so important? Katie Kovalcin, the designer of the Cofounder project, publicly described how she worked with four engineers and used AI to directly migrate her prototype to production frontend, which was then reviewed by engineers and launched. This in the past organization meant: Designer Crossing Into Engineering. And the program even mentioned: She once reached a high percentile in internal code output metrics. What’s truly important is not the ranking. But rather: The translation layer between: Design → Production is shrinking. ──────────────── Twenty-six, this may give rise to a new profession: AI Builder. Not purely: Engineer. Not purely: Designer. Not purely: PM. But a person who can: Understand users; Design processes; Build prototypes; Write production code; Connect models; Test; Deploy. In the past, this was called: "Full-stack genius." In the future, it may gradually become: Ordinary high-level knowledge workers. This is one of the biggest organizational significances of AI: Generalist Returns. The value of generalists is rising again. ──────────────── Twenty-seven, the industrial revolution increased specialization, while AI may partially reverse specialization. In the past, as companies grew larger, one reason was: Specialized division of labor improved efficiency. Adam Smith’s pin factory: Everyone does one step. Internet companies are also: Frontend; Backend; Infra; PM; Design; Analytics. AI allows individuals to regain: Cross-disciplinary execution capabilities. This may lead to: Companies re-modularizing into smaller, high-capability teams. Not eliminating expertise. But rather: Reducing friction between professional boundaries. ──────────────── Twenty-eight, why might Cofounder having only five people be an advantage? The larger the team: Communication Channels grow exponentially. 5 people: 10 bidirectional relationships. 20 people: 190 relationships. 50 people: 1,225 relationships. This is why: As organizations grow large, processes naturally emerge. Jira; Standup; Manager; Program Manager; Roadmap. These things are not because large enterprises are clumsy. But rather: Coordination Complexity. ──────────────── 29. If AI allows five people to have the execution power of the past twenty people, it can directly eliminate a portion of coordination costs. This is the most important leap in productivity. Not: 20 people each 30% faster. But rather: Originally needing 20 people, now only needing 5. The latter's change in productivity is far greater. Because you simultaneously reduce: Salaries; Meetings; Management; Communication; Handovers; Organizational politics. Thus, the greatest improvement in corporate efficiency from AI may come from: Smaller Team Size. ──────────────── 30. One of the most valuable company metrics in the future may be Revenue per Employee. The internet has already improved this metric. AI will further enhance it. In the past: $1 billion in revenue might require: 10,000 people. In the future: it might only need: 1,000 people. In the further future: some software companies might achieve this with just a few hundred people. Thus: Company value is becoming increasingly decoupled from: Employee count. This will change: Offices; Management; Recruitment; Organizational design. ──────────────── 31. But do not misunderstand “5 people successfully developing” as meaning all company processes should be eliminated. This is the most common mistake. Cofounder projects can achieve: No Jira; No PM; Less documentation, with a few very special prerequisites: Senior talent; Very small teams; High trust; Founders directly involved; Short-term sprints; Technical exploration; Existing mature Gusto infrastructure. A: 500-person financial system migration project cannot be directly copied. So what should really be learned is: Process should match coordination cost. Not: Process = Bad. ──────────────── 32. Processes are essentially a "tax on organizational size." In the early stages of entrepreneurship: Everyone sits at one table. No need for Jira. After 1,000 people: Lack of processes will lead to collapse. So the issue is never: Whether there are processes. But rather: Does AI allow companies to maintain smaller teams long-term, thus delaying the emergence of processes? If the answer is: Yes. Then AI is not just improving productivity. It is changing: The Company Scaling Curve. ──────────────── 33. This will cause future startups to "mature later." Traditional startups: 5 people. 20 people. 100 people. 500 people. To grow, they must continuously hire. AI startups might be: 5 people; 15 people; 40 people; with significant revenue already. Thus: Organizational bureaucratization occurs later. Founders can maintain: Direct product control for longer. This may reinvigorate: The innovation speed of startups. ──────────────── 34. Another important layer to Eddie's story: Technical founders often get pushed away from technology by their own success. In the early stages: Founders write code. After company growth: Hiring; 1-on-1s; Performance; Budgets; Organizational design; Board meetings. Eventually: CTOs no longer touch code. This is normal. Because: The company's bottleneck has shifted from: Coding to: Organization. Eddie himself also experienced this transformation. ──────────────── 35. Many technical founders' real pain is not that they cannot manage, but rather "identity loss." Your original life identity is: Builder. Suddenly every day you become: Manager. Meetings; Hiring; Performance; Conflicts. This creates a deep psychological change: Who am I now? Past achievements came from: "I built something." Later achievements come from: "I made others build something." These two reward mechanisms are very different. ──────────────── 36. This is also why technical founders easily "snatch code to write." When the team has a bug. The founder sees it. Directly fixes it. In the short term: It's the fastest. In the long term: It's the most dangerous. Because: You forever become: The most expensive IC in the team. And your real job is: Building the team. This is: The Founder Delegation Trap. ──────────────── 37. So Eddie stopping coding back then may have been correct, and today starting to code again may also be correct. These two things are not contradictory at all. In 2015: Organizational expansion was the bottleneck. He should manage. In 2026: The technical paradigm has changed. Technical understanding becomes the bottleneck. He becomes a builder again. The greatest ability of an excellent leader is not: To always stick to one management style. But rather: To know where the current organizational bottleneck truly lies. ──────────────── 38. This is why the essence of leadership is not "what to do," but "what to do at which stage." In the early stages of entrepreneurship: Founder = Builder. As the company grows: Founder = Recruiter. Further: Founder = Manager. Maturity: Founder = Capital Allocator. With changes in technical paradigms: It may temporarily revert to: Builder. Truly excellent people can accomplish: Role Switching. Many founders fail: Not due to lack of ability. But rather: Staying in the role they were best at in the past. ──────────────── 39. Conflicts between co-founders are essentially also about roles and a sense of fairness. Early-stage technical founders often experience a psychology: I write code until dawn every day. Why does the CEO: Meet investors; Have meals; Chat? It seems: I am working. He is socializing. But this is a: Short-term Accounting of entrepreneurial division of labor. Early teams cannot calculate daily: Who worked more hours. What should really be looked at is: In the long term, who bears: What responsibilities; What risks; What results. ──────────────── 40. The most dangerous thing for a founder team is not unequal division of labor, but rather "unspoken feelings of unfairness." If this emotion accumulates: Technical founders feel: The CEO is not working. The CEO feels: The CTO does not understand the market. Product founders feel: They have no power. In the end: The company does not lose to competitors. But rather: Internally disintegrates. So Eddie's team clearly defined early on: Josh Reeves → CEO; Tomer London → Product; Eddie Kim → Technology, The real value lies in: Role Clarity. ──────────────── 41. Many startups truly die from "ambiguous control." Everyone is friends. So: Let’s not talk about it first. This is a mistake. The more friends you are: The more you should talk about it in advance. Who is the CEO? Who makes major decisions? Who is ultimately responsible for the product? Who is responsible for technology? Why is equity divided this way? What happens if the company fails? What happens if someone leaves? The earlier you eliminate: Ambiguity, The less emotional debt there will be in the future. ──────────────── 42. Emotional management is actually the most underestimated productivity tool for founders. Eddie mentioned: When people are attacked or challenged, There are obvious physiological responses. Increased heart rate; Changes in breathing; Fight or Flight. The most dangerous thing at such moments is: To respond immediately. Because: You think you are in a "rational debate." In fact: Your body has already entered a defensive state. ──────────────── 43. What excellent leaders truly need is "Meta-awareness." Not only knowing: What has happened. But also knowing: What is happening to me right now. Why am I angry? Is my ego being challenged? Am I protecting the company, Or protecting my own identity? This is called: Metacognition. Observing your own thoughts. The higher the leadership level, The more important it is. Because: A single uncontrolled word, Can affect: Dozens of employees; Funding; Co-founder relationships. ──────────────── 44. Why is pausing for a few minutes effective? Because truly advanced management is not: Never having emotions. That’s unrealistic. But rather: Emotion → Pause → Interpretation → Response. Ordinary people: Emotion → Reaction. Adding a Pause in between, Will create a huge difference. This is also the ability difference that often exists between excellent CEOs and ordinary managers: Response Latency. It’s not about being faster is better. But rather, in high emotional situations: Knowing when not to respond immediately. ──────────────── 45. AI programming assistants have another very underestimated value: they lower the "cost of learning shame." Eddie hasn’t coded in a decade. Today the tech stack has completely changed. When asking employees again: How do you configure this? How do you use Datadog? How do you use the framework? For a senior CTO, There is a psychological cost: Shouldn’t I already know this? AI has no such social judgment. You can ask: Extremely basic questions; Ask repeatedly; Let it explain ten times. No one thinks you are stupid. This is called: Zero-Judgment Tutoring. ──────────────── 46. This may change the way adults learn technology. The biggest obstacle for adults learning is often not: Intelligence. But rather: Ego. College students can ask: "What does this mean?" A VP of Engineering with 20 years of experience might feel embarrassed to ask. AI brings this cost down to zero. So one of the biggest markets for AI Tutors may not be: Students. But rather: Mid-career Reskilling. Mid-career professional retraining. ──────────────── 47. The "technological generational gap" issue for future executives may significantly decrease. In the past: Engineer → Manager → VP → CTO. Getting further away from coding. Ten years later: It’s hard to go back. AI makes re-entering easier. This may mean: Technical leaders can switch back and forth between: Management and: Building. This is a new: Career Elasticity. ──────────────── 48. Why is that five-hour wait at Heathrow Airport so symbolically strong? Because in the past, an executive had a new idea. First step: Find a PM. PM writes requirements. Arrange the team. Schedule it into the Roadmap. Possibly: Start three months later. Now: Pull out a computer. Claude Code. In five hours: A prototype emerges. This means: The distance between Idea → Artifact suddenly collapses. This is one of the most important changes in AI entrepreneurship. ──────────────── 49. The biggest bottleneck in tech entrepreneurship is shifting from "Can it be done?" to "What should be done?" In the past, one person had an idea. But couldn’t: Backend; Frontend; Database; Deployment. Could only: Find a technical co-founder. Today, AI can help accomplish a lot of implementations. Thus: Implementation Scarcity ↓ And: Judgment Scarcity ↑ Knowing: What problems are worth solving, is becoming increasingly important. ──────────────── 50. This does not mean that technical co-founders are useless. On the contrary. Truly excellent technical founders have greater leverage. In the past: One excellent engineer = 3 ordinary engineers. In the future: One excellent engineer + AI may equal: 10 or even more ordinary executors. So AI is not simply: "Replacing programmers with non-coders." It’s more likely to be: Amplifying the Best Builders. ──────────────── 51. This is also why small team startups may usher in a golden age. In the past, a SaaS startup: Needed: Frontend; Backend; Infra; Designer; PM; Data; DevOps. Today: A few comprehensive Builders can cover more areas. Thus, the earliest Burn Rate of startups decreases. Runway extends. The number of trial and error increases. This enhances: The Optionality of startup success. ──────────────── 52. The real smart point of Gusto Cofounder is that it did not stop at "letting bosses create apps themselves." Eddie's initial idea was closer to: Letting small business owners describe their needs, AI helps them create a: Mini App. Sounds cool. But after customer research, it was found that: Bosses really don’t want: Another app. What bosses really want is: You help me get things done. This is a very important product insight. Entrepreneurs often love: Solutions. Customers only love: Outcomes. ──────────────── 53. The biggest enemy of small businesses is not "lack of software," but rather too much software. Restaurants may have: POS. Scheduling software. Payroll software. Accounting software. Ordering platforms. Delivery. CRM. Review platforms. Every SaaS claims: "Increase efficiency." In the end, the boss becomes: A Human API. Manually copying from one system to another. This instead creates new work. ──────────────── 54. One of the biggest opportunities for AI Agents is to eliminate the "Human API." Assuming a spa uses Mindbody to manage appointments. It contains: Service items; Working hours; Tips; Commissions; Additional fees. Payroll is in Gusto. In the past, the boss needed to: Export CSV; Excel; Calculate; Re-enter Payroll. The boss actually bears: The Integration Layer. What the co-founder wants to do is: AI reads data directly; Understands rules; Calculates payroll; Confirms with the boss; Executes Payroll. That is: Machine-to-Machine Workflow. ──────────────── 55. This is where Agents truly add value: cross-system execution. A chatbot can only: Answer. A true Agent must: Read; Reason; Act; Check; Repeat. For example: Read Mindbody; Calculate commissions; Detect anomalies; Call Gusto; Wait for approval; Execute payroll. This is called: End-to-End Task Completion. ──────────────── 56. The biggest market for Agents may not be "knowledge Q&A," but rather those tedious tasks within companies that no one likes. For example: Chasing Timesheets; Employee onboarding; Checking Payroll; Preparing tax documents; Updating CRM; Handling invoices; Checking compliance; Customer follow-up. These tasks share a common characteristic: They are not particularly difficult. But: They are repetitive; Cross-system; Must be done. This is where AI can most easily create ROI. ──────────────── 57. The golden formula for enterprise AI may be: High Frequency × High Friction × Clear Rules × Digital Inputs × Measurable Outcome High frequency. High friction. Relatively clear rules. Digital inputs. Verifiable results. Payroll fits this very well. This is also why: Back Office may become the area where Agents are first commercially scaled. ──────────────── 58. Why is the "Heartbeat" concept so important? Traditional ChatGPT: If you don’t ask, it won’t do. A real employee is not like this. An excellent employee will: Know payday is Friday; Proactively check Timesheets on Thursday; Detect anomalies; Approach the boss in advance. This is: Proactivity. The significance of Heartbeat is: AI wakes up regularly, Checks: Is there anything that should be done now? Gusto has officially designed Cofounder to actively monitor: Payroll; Filings; Anomalies; Reminders as a system, rather than simply waiting for prompts. ──────────────── 59. The true qualitative change of Agents is not Intelligence, but Agency. Intelligence: Knowing the answer. Agency: Proactively completing tasks. The past LLM revolution mainly improved: Intelligence. The next stage: Tools + Memory + Scheduling + Permissions provides: Agency. This is what will truly integrate into enterprise workflows. ──────────────── 60. However, businesses like payroll reveal the biggest risk of Agents: mistakes can directly turn into real money. If a chatbot answers incorrectly: The user asks again. If a Payroll Agent answers incorrectly: Payroll may be wrong. Taxes may be reported incorrectly. Employees may not receive their money. Therefore, Agentic Software must have an additional layer compared to ordinary chatbots: Control Architecture. Permissions; Approvals; Logs; Rollbacks; Anomaly detection. Gusto also clearly indicates that Cofounder outputs may contain errors, and decisions regarding Payroll, taxes, and compliance remain the responsibility of the client; its terms also clearly incorporate automated actions and authorization mechanisms into the existing employer liability framework. ──────────────── 61. So the true moat of future Agent products may not be the model, but rather the "secure execution layer." Anyone can call: Claude; GPT; Gemini. What’s truly difficult is: How to let AI: Safely access bank data; Issue payroll; Change insurance; Submit taxes; Modify employee records. This requires: Permissions; Audit Logs; Human Approval; Identity; Compliance. That is: Enterprise Agent Infrastructure. ──────────────── 62. This is why old SaaS companies have a huge advantage. They already have: User trust; Permission systems; Enterprise data; Audit systems; Regulatory experience. A new AI startup can be smarter. But: Getting clients to allow it to issue payroll directly is another matter. So: The truly big companies in the AI era may not be: The ones with the strongest models. But rather: The ones most qualified to let AI take action. This is an extremely important judgment. ──────────────── 63. "Who owns Action Permission" may be more valuable than "Who has the strongest model." Imagine future enterprise Agents. They need to: Make payments. Change CRM. Issue contracts. Approve expenses. Issue payroll. If each step requires: New identity verification; New authorization, The experience will be very poor. Truly possessing: System of Record + Permission Layer platform, naturally occupies: Action Layer. Gusto has this position in Payroll. ──────────────── 64. This could become the best path for traditional SaaS to go from "being disrupted by AI" to "using AI to fight back." Don't just add: AI Chat. Instead, expose your: data; workflow; permissions as: Agent Actions. The future value of Salesforce: is not just CRM. But: AI Agents can safely execute Sales Work there. Gusto: safely executes Payroll Work. QuickBooks: executes Finance Work. This is: System of Record → System of Action. ──────────────── 65. Why do small businesses especially need an "AI Cofounder"? Large companies can hire: Chief of Staff; HR; Finance; Operations. Small business owners do not. So the greatest significance of AI for SMBs is not: replacing 10 people in a 100-person team. But rather: enabling a 5-person business to gain the management capabilities of a 50-person company. This is called: Capability Democratization. ──────────────── 66. This could lead to a further explosion in the number of small businesses. The cost of starting a business has already decreased significantly due to: Cloud; Shopify; Stripe; Social Media. AI further reduces the costs of: Coding; Marketing; Accounting; Operations; Customer Support. Thus: Minimum Efficient Firm Size continues to decrease. In other words: the minimum scale required to maintain a competitive company is getting smaller. ──────────────── 67. Macroeconomically, this could lead to a "small business revival." In the past few decades: large companies have enjoyed: economies of scale. Because: HR; Legal; IT; Software these fixed costs are cheaper for large companies to spread out. If AI turns backend capabilities into cheap digital services: small companies begin to gain similar capabilities. This could reduce: Scale Advantage. Allowing small businesses to regain competitiveness. In Gusto's own 2026 survey, about one-third of small business owners said AI has made starting a business easier, and 62% believe AI has the potential to narrow the competitive gap between small and large businesses. Of course, this is attitude data obtained from Gusto's own survey and does not equate to proven macro gap narrowing. ──────────────── 68. The biggest change for a person starting a company in the future may not be "one person completing the product." The real difficulty is: operating the company. Sales; Payroll; Tax; Customer Service; Finance; Compliance. If AI Agents gradually take on these backend tasks: Solo Founders may truly exist long-term. So: AI Coding is just the first layer for Solo Companies. AI Back Office is the second layer. ──────────────── 69. Ultimately, a "company operating system" may emerge. Founders only need to say: This month's revenue target. Add three people. Control marketing costs. Ensure payroll on Friday. The system automatically coordinates: Payroll Agent; Recruiting Agent; Finance Agent; Sales Agent. What the boss really does: Strategy + Judgment. This could be: Autonomous Enterprise OS. Gusto Cofounder can be seen as an early prototype of this future. ──────────────── 70. But don't be misled by the name "Cofounder." A true co-founder has: value judgment; risk-bearing; creativity; responsibility; trust relationships. AI currently cannot truly bear: legal and economic responsibilities. So Cofounder is more accurately: Agentic Operations Teammate. The name is product positioning. Not a co-founder in the legal sense. This distinction must be made. ──────────────── 71. From an investment perspective, why is Gusto Cofounder more worthy of attention than an independent Agent Startup? Because Gusto has: 500,000+ customers; over $1 billion in revenue over the past 12 months; Payroll System of Record; enterprise permissions; long-term customer relationships. This means: once the new product is effective, Distribution Cost is extremely low. No need to: find new customers. Directly: Upsell / Deepen Engagement. This is: Installed Base Advantage. ──────────────── 72. The real biggest AI asset of mature SaaS may not be code, but rather the Installed Base. Startups have: technological speed. Incumbents have: customers. In the AI era, the outcome depends on: who can more quickly combine: Technology and: Distribution. If traditional companies are too slow: Startups will steal customers. If traditional companies are fast enough: they can directly distribute AI to hundreds of thousands of customers. Gusto is trying to take the second path. ──────────────── 73. This also explains why it is important for founders to personally lead this product. A common problem with innovation in large companies is: new products must go through: budget; committees; management; roadmaps; quarterly planning. The more disruptive the product, the easier it is to be killed by the existing organization. Because: it does not meet current KPIs. When the founder personally leads the team: it is equivalent to giving the project an: Organizational Shield. ──────────────── 74. This is the real version of what Christensen calls the Innovator’s Dilemma. Mature products are already making money. Why take the risk to create: new products that may cannibalize existing models? Traditional managers will naturally protect: the current business. Founders find it easier to say: If someone is going to disrupt Gusto, it is best: Gusto disrupts Gusto itself. This is the hardest thing for excellent mature tech companies to do. ──────────────── 75. So Eddie's return to the front line truly represents another version of "Founder Mode." Not: the founder managing everything. But rather: when a sufficiently large technological paradigm shift occurs, the founder should: bypass organizational inertia; relearn directly; engage directly with customers; prototype directly. Because: this change cannot be understood solely through: Management Chain. ──────────────── 76. What should "No Jira, No Docs" really be translated to? Not: management tools are useless. The correct translation is: minimize the distance between Build and Feedback as much as possible. If: today's idea. today's writing. tomorrow's customer use. The day after tomorrow's modification. This is: Learning Velocity. The real competition in entrepreneurship is not: Code Velocity. But rather: who learns the fastest. ──────────────── 77. Thus, the ultimate competitive advantage for software companies may increasingly shift from "development speed" to "learning speed." All companies can use: Claude Code. Thus: Coding Advantage is compressed. The real distinction becomes: who can more quickly: find customers; understand problems; experiment; obtain feedback; change direction. The formula becomes: Build → Observe → Learn → Rebuild. Whoever has the most cycles, is more likely to find: Product-Market Fit. ──────────────── 78. Therefore, AI ultimately does not make product judgment unimportant, but rather exposes erroneous judgments more quickly. In the past: making the wrong product. Only knowing after six months. With AI: producing it in two weeks. Customers say: no. That's fine. Because: it only wastes two weeks. Thus, one of the business values of AI is: Increase the Number of Bets. Increase the number of trial and error. The VC world likes Optionality. Product development does too. More cheap experiments increase the probability of finding big opportunities. ──────────────── 79. In the future, the core asset of product teams may be the "experiment portfolio." Simultaneously trying: Option A; B; C; D. Because code is cheap. However: user attention is still expensive. So the bottleneck shifts from: Engineering Capacity to: Customer Attention. This is also why user research and real distribution are becoming increasingly important. ──────────────── 80. The most dangerous trap for products in the AI era: because anything can be done, everything is done. Claude can: generate 20 features in a day. Very satisfying. But customers: may not want any of them. This is called: Capability Trap. After technological capabilities expand infinitely: the biggest risk is: lack of focus. So the value of Taste is rising. Know: What not to do. ──────────────── Eighty-one, the term "Focus" during the Steve Jobs era is even more important in the AI era. Things not to do in the past: Due to limited engineering resources. In the future, engineering resources will become cheaper. Thus, constraints will disappear. After constraints disappear: The most likely occurrence: Product bloat. So a truly excellent Product Leader must create: Artificial Scarcity of Attention. Artificial limitations: What is worth investing in. Otherwise, AI will create: Feature Explosion. ──────────────── Eighty-two, what Eddie is truly relearning this time may not be code, but rather "technical intuition." After ten years of management, the easiest thing to lose is not: Syntax. But rather: Feel. What is easy now? What is difficult? Where will the Agent fail? What requires human approval? What can the model really do? These cannot be obtained through: Quarterly reports. Only by personally building, can one restore: Technical Intuition. ──────────────── Eighty-three, this is a very important warning for all technical managers. The higher your position, the easier it is to obtain: Abstract information. Dashboard; Metric; Summary; Slide. But: The technological revolution first occurs in: Details. Why does the prompt fail? Why does the Agent loop? Where does the Tool Call break? When does the model hallucinate? Only by getting close to: Ground Truth can one know. So: Executive Abstraction Debt is also a real existence. ──────────────── Eighty-four, the future excellent CTO may have to switch between two extremes. Macro: Architecture; Organization; Capital; Talent; Long-term strategy. Micro: Personally running models; Writing prototypes; Looking at traces; Doing evaluations. In the past, CTOs could mainly stay at the Macro level. In the rapidly changing AI era: One must maintain: Macro-Micro Oscillation. Constantly switching between high-level and low-level. ──────────────── Eighty-five, the biggest insight for large companies from Gusto is that AI transformation cannot just be "buying Copilot for employees." Many companies' so-called AI strategy: Purchase ChatGPT Enterprise. Train employees. End. This is merely: Tool Adoption. True AI-native transformation must ask: Does the organization still need these roles? Should workflows be redesigned? Which processes can be entirely eliminated? Which decisions must be retained manually? This is: Operating Model Redesign. The difficulty is completely different. ──────────────── Eighty-six, the most important thing is not to make old processes 30% faster, but to ask whether old processes still need to exist. For example: Traditional process: PM writes requirements; Designer makes drafts; Engineer writes code; QA; Release. In the AI era, it may be: Product Builders directly create real versions; Engineering reviews; User testing. This is not: Making each step of the old process 30% faster. But rather: Eliminating steps. Truly significant efficiency gains almost always come from: Process Elimination. ──────────────── Eighty-seven, this is very similar to the thinking of Toyota and Lean Manufacturing. The most important thing in Lean is not: To make everyone work faster. But rather: To eliminate Waste. Waiting; Inventory; Handling; Rework; Unnecessary processes. AI software organizations also need to ask: Which meetings are Waste? Which documents are just for handover? Which tickets exist just because people cannot express directly? AI is redefining: Software Lean. ──────────────── Eighty-eight, if Gusto's 10-week experiment is expanded to the entire economy, the significance is enormous. There are millions of small and medium-sized enterprises in the United States. Each has: Background friction. Assuming AI saves the boss: 5 hours per week. 1 million bosses: 5 million hours/week. In a year: 260 million hours. This is: Aggregate Productivity. The true economic value of AI will not come only from: A few large model companies. But rather from: Countless small time savings accumulated. ──────────────── Eighty-nine, this is also why "small business Agents" may be a very large market. Consumer AI: May be willing to pay: $20 per person per month. Enterprise AI: If it helps the boss save: $50,000 a year, it can charge: $5,000; $10,000. Because: ROI is clear. The true pricing anchor for excellent B2B AI is not: Model cost. But rather: Economic Value Created. ──────────────── Ninety, Gusto has already proven this logic of "hidden value automatic discovery." In the spring of 2026, Gusto claimed its AI tax function helped small businesses identify and claim about $70 million in federal tax credits. The company also clearly reminded that "the claimed amount" does not equal the actual refund, and the final benefit depends on the business's own tax situation. But this case still illustrates: One of the areas where AI can most easily create value is: Complex rules + users do not know what they have missed. Tax; Payroll; Compliance are particularly suitable. ──────────────── Ninety-one, a huge future function of AI is not "answering questions," but rather "discovering the questions you don’t know you should ask." The boss does not know: There is a tax incentive. So they won’t ask. Traditional Chatbots: If no one asks, they don’t answer. Agents: Discover: You qualify. Proactively tell you. This is: Latent Need Discovery. The discovery of latent needs. The value is one level higher than search engines. ──────────────── Ninety-two, this is also why proactive Agents will resemble real employees more than Chatbots. A truly excellent employee does not wait for the boss to say: "Remind me to file taxes." They will say: "The deadline is next week, I am ready." So the true maturity of Agent products is marked by: From: Prompt-driven to: Event-driven + Goal-driven. Event-driven. Goal-driven. ──────────────── Ninety-three, but the risks of proactive Agents also increase exponentially. The more proactive: The easier it is: To make mistakes. So Agent products have a core trade-off: Autonomy vs. Control. The higher the degree of autonomy: The easier it is for users. But the higher the risk. Truly good systems must design: Which actions: Are executed automatically. Which: Are suggestions. Which: Must be manually confirmed. ──────────────── Ninety-four, therefore, the core of future Agent UX may not be Chat, but rather Permission Design. Today everyone is designing: Chat boxes. But the most important UI for truly mature Agents may be: Permissions. For example: Can read: Payroll. Can calculate: Bonuses. Can draft: Salaries. But: Final transfers must be confirmed by me. This is called: Progressive Autonomy. Gradual autonomy. As user trust increases: Permissions gradually expand. ──────────────── Ninety-five, this is very similar to the Level system of autonomous driving. L1: AI suggests. L2: AI executes part of the steps. L3: AI executes most steps, supervised by humans. L4: Automatically completes in a limited environment. Enterprise Agents may also form similar levels. Different tasks: Different levels of autonomy. Payroll: Cautious. Birthday Reminder: Can be automatic. This will become: An important foundation for enterprise Agent product design. ──────────────── Ninety-six, the moat that Gusto co-founder may truly form is "each small business's own way of working." For example: This Spa: Technician A takes a 40% cut. Project B adds $15. Weekend rules differ. Certain employees have special bonuses. These are all: Tacit Business Logic. Implicit business rules. In the past: They existed in the boss's mind. If the Agent learns these: It begins to have: Business Memory. ──────────────── Ninety-seven, Business Memory may become the most valuable data in future enterprise software. System of Record preserves: What happened. Business Memory preserves: Why we do this. Including: Rules; Exceptions; Preferences; Historical decisions. Once the Agent accumulates this Memory: It understands the enterprise more and more. This will form a very strong: Switching Cost. Because switching Agents: Is equivalent to retraining an employee. ──────────────── Ninety-eight, this may be stronger than traditional SaaS data lock-in. Traditional SaaS Switching Cost: Data migration. Agent Switching Cost: Data + habits + judgments + Memory. This is like: Replacing a COO who has worked for ten years. So truly successful Enterprise Agents: Customer lifecycles may be very long. This is also why: "AI Teammate" It has the potential to be stronger than ordinary SaaS in terms of economic structure. ──────────────── Ninety-nine, but this also sharply increases the importance of data governance. If the Agent knows: Salary; Employee performance; Banking; Taxation; Company strategy. It possesses: Highly sensitive information. So: Security; Privacy; Access Control; Audit will become: Core product features. Not: Back-end Compliance. Whoever handles it poorly: One incident could destroy trust. ──────────────── One hundred, what does Eddie Kim's return really tell us? Not: "The boss must also know Claude Code." The real answer is deeper. It tells us: The technological change is significant enough for a software company with fourteen years of history and over $1 billion in revenue to rethink like a startup team of five. This is very rare. When a mature company is willing to: Delete processes; Rewrite code; Re-understand customers; Let designers build; Let co-founders return to the front line, it indicates that it feels: The old software development paradigm itself is failing. ──────────────── One hundred and one, what is the biggest revelation for technology managers? Not: You must write code every day. But: Do not let your position separate you from new production materials. If AI changes: Coding; Design; Research; Management, you must experience it personally. Otherwise: Your strategic judgment is based on: Others' descriptions of the new world. ──────────────── One hundred and two, what is the biggest revelation for entrepreneurs? First: Do not start by making tools. Find: The work users truly want to be "done away with." ──────────────── Second: Look for: System of Record. Data + Permissions + Workflow is more valuable than the model itself. ──────────────── Third: Keep the team small. The biggest dividend of AI may not be: Paying a few less engineers' salaries. But rather: Reducing coordination complexity. ──────────────── Fourth: Get things to customers as soon as possible. After AI makes experiments cheaper: Do not waste time on: Hypotheses. ──────────────── Fifth: Build Memory. The truly powerful future Agents: Must understand customers more and more. ──────────────── One hundred and three, what is the biggest revelation for mature enterprises? Do not place AI in: Innovation Lab. You should choose a few truly: Mission-critical Workflows, and let senior leaders participate personally. Because: AI transformation is not: Adding a feature. But rather: Redesigning: Who does the work. Who actually completes the work. ──────────────── One hundred and four, what should investors observe the most? Do not just ask: Which model does this company use? Ask: Does it have: Customer distribution? Proprietary Context? Execution permissions? Long-term Memory? Quantifiable ROI? Real Automation? These are the: Agent Economy long-term moats. ──────────────── One hundred and five, what are the truly dangerous software companies of the future? Not: Those without AI features. But rather: Those whose software only exists to: Make humans click many buttons. If Agents can bypass the UI, Directly call underlying functions: The value of UI will decline. So future software companies must possess: Data + Workflow + Permission + Action. Otherwise: They may be turned into: Back-end APIs by Agents. ──────────────── One hundred and six, the truly valuable aspects of software may increasingly move to the underlying layers. In the past: Beautiful UI was the product. In the future: Agent Interface may unify the front end. Thus differentiation will shift to: Data; Rules; Business logic; Execution permissions; Trust. This is why: System of Record will not easily disappear. But: Interface SaaS may face tremendous pressure. ──────────────── One hundred and seven, this also means that the software industry may see a new profit redistribution. Model layer: Charge Intelligence Rent. System of Record: Charge Data / Workflow Rent. Agent: Charge Labor Replacement Value. Traditional pure UI software: Value may be compressed. The future SaaS war, is not: Who has the prettier page. But rather: Who controls the work execution chain. ──────────────── One hundred and eight, stepping further: AI may redefine the "company" itself. Coase's theory of the firm suggests that: Companies exist, partly because: The internal coordination costs are lower than: The costs of trading everything through the market. If AI simultaneously lowers: External transaction costs and: Internal coordination costs, the optimal boundaries of firms may change. In the future: A core team • AI Agents • Numerous external APIs / Contractors could accomplish the work of past large companies. This is called: Firm Boundary Compression. ──────────────── One hundred and nine, thus AI may ultimately not only improve company efficiency but also change the average size of companies. Industrial Revolution: Large factories. 20th Century: Large enterprises. Internet: Light asset companies. AI: May further produce: Tiny Giants. Micro giants. Dozens of people. Hundreds of millions in revenue. Or even more. If such companies emerge in large numbers, The entire: Labor market; VC; Management; Offices; Career paths will all be redesigned. ──────────────── One hundred and ten, this is the most memorable aspect of Eddie Kim rewriting code. A technology founder who has already achieved great commercial success, reopens the code editor. The significance lies not in: Nostalgia. Nor is it: To prove he can still code. But rather he realizes: The new software production methods have changed to the point where a management perspective alone is insufficient to understand. In the past decade: Management was his leverage. Today: AI makes Building itself a high-leverage activity again. An excellent founder must accept a difficult truth: The way that made you successful in the past may not be the most correct way in the next stage. True growth, is not about constantly upgrading your position. But rather: Willingness to become a: Beginner again after the world changes. ──────────────── The most memorable sentence The story of Eddie Kim and the Gusto co-founder, on the surface, is a: AI Coding revolution. But at a deeper level, three revolutions are happening simultaneously. The first: Software revolution. Software is moving from: Providing tools to humans to: Completing work for people. ──────────────── The second: Organizational revolution. AI is causing: Designer; Engineer; PM; Technical Leader to begin overlapping in boundaries, small teams regain extremely strong productivity. ──────────────── The third: Entrepreneurial revolution. In the past, establishing a truly operational company required: Software; HR; Payroll; Finance; Operations; Compliance. In the future, more and more work may be undertaken by: AI Agents. Entrepreneurs will be truly responsible for: Product; Customers; Judgment; Culture. Thus, what the Gusto co-founder truly represents is not just: "A smarter payroll software." But a bigger question: If in the future an entrepreneur has dozens of always-online AI teammates that can understand the company and execute work across systems, how small can a company's minimum scale shrink? In the past, the internet brought: Distribution costs down to nearly zero. Cloud brought: The startup costs of computing infrastructure down to very low levels. AI Coding is lowering: Product manufacturing costs. And what the Gusto co-founder represents as the next step is lowering: Company operational costs. When these four types of costs simultaneously decrease: Distribution • Computing • Building • Operations are all repriced. What may truly happen then is not: "Starting a business has become a bit easier." But rather: The organizational form of companies, which has existed for hundreds of years, begins to be redesigned. And Eddie Kim sitting back down at the computer to code ten years later, may be the most symbolic scene of this new era: A company with over $1 billion in revenue, in order to understand the future, has to relearn how to work like a startup of five people. The three insights that are most worth remembering long-term are: First, the greatest value of AI Coding is not "code faster," but allowing organizations to become smaller, directly reducing coordination costs. Second, the true disruption of SaaS by Agents is not adding a chat box, but transforming Human + Software into Human → Goal → AI completing the work. Third, the most dangerous aspect of mature companies' AI transformation is not falling behind in technology, but the management being too far from the real new production methods. Eddie Kim returning to code is essentially regaining first-hand technical intuition.