For accounting, the speaker lays out three approaches: sell AI software to existing firms, start a full-service firm, or buy a firm and introduce AI. Starting a firm means taking on manual work, so founders should track the share of work automated and avoid scaling revenue by adding staff faster than automation improves. The speaker points to Vessence, whose founders learned from a Stockholm law firm, and recommends finding early customers with the authority and motivation to adopt new software.
Startup Advice: AI GTM, Pivoting & How To Hire
A Y Combinator speaker joins the host and community members for Office Hours to discuss AI go-to-market, hiring, pivots, and open-source strategy.
Early companies should prioritize fast learning about customer needs and the sharpest problems. Start with the smallest customer segment that has the problem; a narrow product or a few users inside a large company can shorten the path to useful feedback. AI sales tools work best after founders have learned whom to target and how to reach them, while premature sales hiring and growth tactics can obscure what is actually working.
The Firecrawl team pivoted from Mendable after noticing that its crawler solved a more valuable problem for AI agent companies. The speaker says conviction can come from repeated customer conversations and evidence that users truly value a product, rather than revenue alone. Exploring several ideas can make a pivot easier, and technical difficulty can be an advantage when the team believes the problem matters; founders can reduce scope with a rough first version while continuing to talk to customers.
Hiring is timely when a specific area such as engineering, sales, or onboarding is close to breaking, since recruiting can take months. Before product-market fit, hire sparingly; after it, teams may wait too long, but hiring itself is not a measure of success. Founders should be wary of adding people in the hope that it will speed things up, with rare opportunistic hires making sense when an unusually strong candidate is available.
Open source can help enterprise SaaS sales by making a product easier to trust, inspect, extend, or self-host. The host cites Medplum's open-source EHR as a way to address enterprise trust and potentially shorten sales cycles, and Twenty as an open-source CRM that is not aimed at developers. Self-hosting can ease data and security concerns, but it carries costs that companies need to price for.
Takeaways
- For AI services, track automation progress alongside revenue and staffing.
- Sell first to buyers who have authority, urgency, and a reason to adopt.
- Use customer conversations and evidence of real user value to assess a pivot.
- Test a rough, smaller version of a hard product while staying close to customers.
- Hire when a real bottleneck is emerging, not to signal company growth.
Mentioned in this talk
Search inside this talk
What was claimed 14
- metric
Y Combinator: Tracked the percentage of technical employees; the speaker recalled a target around 30%“The percentage of technical people working at the company... I forget whether it was 30% or something like that”
YouTube Unverified1 source· Talk
- metricautomation rate: The trajectory of automation can matter more than overall revenue for an AI-enabled service firm
“I would actually care more about the trajectory of the automation rate than the overall revenue”
YouTube Unverified1 source· Talk
- feature
AISDR: Work best when plugged into an already effective sales process“AI SDRs tend to work well when they’re plugged into a sales process that’s already working well”
YouTube Unverified1 source· Talk
- metricAISDR: Customers who cannot sell their own product are likely to churn
“if they go after people who can’t sell their own product... those customers... will mostly churn”
YouTube Unverified1 source· Talk
- feature
GPT-5: May make products better if improving models enhance their capabilities“where improving models will make your product better”
YouTube Unverified1 source· Talk
- feature
Claude Sonnet: Enabled some internal tools to work that had not worked before“When that model came out, a lot of companies building internal tools suddenly had things working that hadn't really worked before”
YouTube Unverified1 source· Talk
- featureMendable: Q&A on top of documentation
“they were doing Q&A on top of documentation”
YouTube Unverified1 source· Talk
- feature
Firecrawl: pivoted from Mendable to focus on a crawler that AI agent companies neededpivoted from Mendable to focus on a crawler that AI agent companies needed“they realized every AI agent company needed that crawling function”
YouTube Unverified1 source· Talk
- metricAlgolia: A few thousand dollars a month in revenue from its initial product
“We got some revenue—not as much as Mendable, but a few thousand dollars a month.”
YouTube Unverified1 source· Talk
- metricGreile: A few thousand dollars of MRR before focusing the product
“They had a few thousand dollars of MRR”
YouTube Unverified1 source· Talk
- featureAutumn: Open-source billing framework
“to an open-source billing framework called Autumn”
YouTube Unverified1 source· Talk
- featureBrahante Biologics: Microfactories to manufacture drugs in small volumes
“they’re building microfactories to manufacture drugs in small volumes”
YouTube Unverified1 source· Talk
- featureopen source: Can build enterprise trust and shorten sales cycles when privacy or security concerns matter
“being open source for them wasn't about go-to-market... It was really about creating trust with enterprise customers and shortening the sales cycle”
YouTube Unverified1 source· Talk
- featureself-hosting: Lets customers run software locally to avoid sending sensitive data to a startup’s cloud
“If you open source it, you can self-host, so there's no question about sending your data to some random startup in the cloud.”
YouTube Unverified1 source· Talk
