Eliot Durbin invests in founders who have a strong, sometimes irrational opinion about how a product should exist and a mission that can continue across companies.
2
Agents are moving enterprise software from copilots toward more autonomous systems, with trust, interfaces, evaluation, pricing, and maintenance still unresolved.
3
Founders should start with a narrow product that proves something useful, ship quickly, learn from customers, and turn repeated human-assisted work into a product.
Summary
Eliot Durbin describes Boldstart Ventures' inception-stage approach and the founder traits he looks for after years of making first-check investments. He backs people with a strong product intuition, empathy for users, and an almost irrational drive to make something exist. He connects the current agent wave to earlier automation and cloud shifts. Agents can automate repetitive enterprise work, improve existing products, or support entirely new applications, though companies still need to solve interface design, user trust, evaluation, pricing, and variable model costs. Durbin sees a path from copilots, where humans remain in the loop, through agents, toward autopilot systems. For founders, he recommends starting narrowly, shipping fast, learning from demand, and productizing onboarding and deployment rather than building bespoke consulting work. His investment test is whether a product proves something useful and whether the founder has a credible opinion about how the problem should be solved.
Investment regrets often come from ignoring a founder's unusual product intuition
Eliot Durbin says he repeatedly regrets passing on people because their product focus was not sharp enough or because he did not believe their idea. The successful founders he missed later figured out how to productize something new and empathize with users. He wants to get better at recognizing that ability. His example is Rahul Vohra, whose first company, Reportive, put a sidebar in Gmail to make people more productive. After learning that a free-platform product was difficult to monetize, Vohra built Superhuman as a focused, fast email client. Durbin says the best founders could choose logical, well-paid jobs, yet feel that a particular product has to exist.
A founder's mission can continue across several companies
Durbin looks for founders he can back repeatedly, regardless of whether it is their first company or their fifth. He says a founder's mission sometimes takes the form of multiple products. Rahul Vohra first tried to make people more effective inside Gmail, then pursued the same goal with a dedicated client. Superhuman narrowed its audience and focused on making every action feel immediate. Durbin treats this kind of continuity as more useful than a single company idea. The investor's role, in his description, is to recognize product intuition and help with parts of company building that may not interest the founder.
Agents make old enterprise automation problems easier to handle
Durbin compares current agents with robotic process automation, screen scraping, and desktop automation. Earlier systems automated repetitive work while keeping a human in the loop, but they depended on many APIs and broke when inputs varied. He recalls companies processing faxed purchase orders, including handwritten fields and codes, with OCR before entering the data into SAP. Large language models can handle handwriting and other messy inputs more easily, in his view. The remaining problem is predictability. Durbin says agents initially hallucinated and were far from production quality, until CrewAI showed how separating agents into crews, assigning roles, and giving them discrete tasks could improve accuracy and support automation at scale.
Agents can be the foundation of new software or an addition to existing products
Durbin sees two paths for agent companies. A new application can be built around agents from the start. An existing product can use agents to make its current workflow more useful. He cites Clay, which began as a data-enrichment tool and later added agents that search the web for information, strengthening the product for sales and revenue-operations users. He also warns against adding AI to every available feature. Some products work better with less intelligence. Durbin prefers rapid experiments where a team can try an AI feature, discard it quickly when it fails, and move on. Granola is an example he likes because its AI stays out of the user's way.
Enterprise agent systems need a path from copilots to autopilot
Durbin describes a progression from copilots to agents to autopilot. Copilots keep people involved while making repetitive work faster and more accurate. He says quality assurance is already a strong example, since agents can automate tests and may be more consistent than human QA work. Multi-agent systems then break a workflow into separate tasks, such as researching, drafting, editing, and publishing a blog post. Each part can be evaluated before more automation is allowed. Autopilot is the further step, where trusted systems act with little human involvement. Durbin says the first interfaces will need clear application-level signals that work was completed correctly, giving users a simple mental check before they delegate more.
Chat can become an orchestration layer when teams already work there
Demetrios Brinkmann describes firing agents from Slack threads as a form of ChatOps. Durbin agrees that teams could tag an agent after making a decision in a conversation and ask it to carry out the work. He connects this to customer support, where Kustomer is exploring virtual agents that use graph and vector databases to handle more of the work a human support agent performs. The limit is the exception queue, which still needs people. Durbin says user trust determines when a company can move from one level of automation to the next. The interface is therefore tied to confidence, since users need to know what an agent did before they allow it to do more.
Agent pricing has to control unpredictable usage and support deployment
Durbin says enterprise agent pricing is difficult because customers may hesitate to use a system if each action adds cost, while an unexpected burst of activity can create a very large bill. Customers may also demand a particular model after a product has been built around another provider. He expects early companies to use forward-deployed engineers who install systems, work with customers, and teach teams how to use the product. Palantir is his example of a company that made this approach work. He advises founders to turn repeated deployment work into a productized onboarding process. Human assistance can remain part of a repeatable product, as long as the answer is not a bespoke service for every customer.
Strong opinions matter more than starting with a company-shaped idea
When evaluating a proposed company, Durbin asks what the product proves and why the problem matters. He prefers founders who have lived with a problem and hold a strong opinion about the right solution. He gives Snyk as an example: its founder saw that faster software delivery would create security problems in dependencies and built for developers before the product became a security tool. Durbin says founders should think about the product and the user rather than beginning with an abstract company plan. Large platforms may copy or absorb a product, but a focused team can move faster and establish a best practice for a narrow group of users.
"If you're going to start a company just be ready to be really narrow but when it runs ship right into it lean into it and run fast deliver what the customers want."Eliot Durbin13:53
Who should watch
Founders deciding whether an agent idea should become a standalone product or an AI feature inside an existing application.
Investors and early startup operators who need a practical way to assess founder conviction, product intuition, and founder-market fit.
Teams selling agents to enterprises that are working through evaluation, trust, pricing, deployment, and customer onboarding.
