Vertical AI, vendor risk, and what founders are prompting at 11 PM
Five pieces worth reading this week across vertical AI, vendor risk, prompt behavior, and the quiet return of the full-stack app. Here is what stood out.
What founders actually prompt
PostHog published data on how people use its AI. A quarter of prompts come in after 6 PM. Tuesday is the highest volume day, peaking at 11 AM. Engineers skew toward debugging queries. Marketers skew toward web traffic and revenue. Users become more informal over time, with more swearing and fewer please-and-thank-yous as they get comfortable. The most common prompt targets: dashboards, session replay, and database queries.
The takeaway for anyone building an AI product is practical. Your power users are not polite, they are not on a 9-to-5 schedule, and their prompts reveal their actual job function more than any persona doc will.
How to build vertical AI that actually holds up
Monaco’s principles post on vertical AI argues that the real work is codifying the workflows of the best practitioners into agentic systems. The claim: that kind of institutional knowledge delivers capability and efficiency a third-party frontier model can’t match on its own.
The hard part is not the model. It is finding the workflows worth encoding, then building the context layers, agentic harnesses, data systems, and feedback loops around them. Monaco also flags an org-design problem: GTM and engineering need to share a vocabulary and work from the same table daily. That is rarer than it sounds.
AI agents will fire your vendors for you
A SaaStr piece makes a pointed argument about outages in an AI-enabled world. When a core vendor went down for 24 hours, the team’s AI agent had identified and begun replacing the vendor before the original vendor came back online. The piece argues that agents make less emotional and more calculating decisions than humans. Humans tend to wait for a fix. Agents tend to evaluate alternatives and act.
The operator implication is direct: if you are a vendor, poor incident communication is now a retention risk, not just a customer satisfaction issue. By the time renewal comes up, an agent may have already built the replacement.
The full-stack app is back
Ashu Garg’s piece argues that leading AI app startups now resemble infrastructure companies. They train their own models, manage complex and costly agent workflows, and treat security and governance as first-class concerns rather than afterthoughts. Four factors favor this approach: model ownership, handling expensive agent operations, compliance depth, and customer trust built through accountability.
The pattern that stands out: successful AI app companies focus on outcomes in high-stakes verticals, not features. They integrate organization-wide workflows rather than point solutions.
Engineers building agent factories
Cutlefish draws a parallel between today’s engineers building AI agent factories and operations teams at Netflix and Etsy in 2012, when cloud made traditional ops roles look expendable. Someone still had to build the platforms. The engineers controlling quality standards and deployment systems for agent output are making the same move now.
Quick links from the same issue
- Gemini in Google Slides can now edit multiple slides or an entire presentation from a single prompt, updating content, refreshing visual styles in bulk, or matching the style of another deck. Users keep full control and get a change summary in the side panel.
- Deel Security launched a new product that connects identity signals across the hiring process to flag fraud before access is granted. It can spot fake applications, detect deepfakes, identify proxy candidates in live interviews, and verify government ID.
- Cloudflare services for private origins is now generally available. It connects internal services to the public internet without exposing a public IP, opening a firewall port, or running cloudflared alongside the origin. Note: Argo Smart Routing and Dedicated CDN Egress IPs are not supported.
- Poppi’s $300 million earnout from its May 2025 acquisition by Pepsi has dropped to $12 million.
- Claude Code removed over 80% of its system prompt for newer models with no measurable degradation on coding evaluations. If you have been writing long prompts for Claude, the data suggests you may be padding.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
