This newsletter is about applied AI use cases: what operators are actually building, what it cost, and what changed once it shipped. We just launched the AI Use Case Library to collect those cases in one place, organized around your problem instead of a vendor's pipeline.

Last week, we looked at Chris Taylor of Fractional AI, recently acquired by the Anthropic/PE JV, who cut an e-commerce company's document-processing cost 84% by pulling the work back from an offshore BPO, and Osman Ghandour of Soal Labs, who encoded six senior MDs as queryable personas inside a PE firm's investment committee.

This week, two builds about the same underlying move: finding expensive work that nobody was measuring. Jordan Gurrieri and BlueLabel rebuilt the dispatch desk at a regional fiber ISP and took the monthly cost of running the AI down to about $2. Adam Schwartz and Parable gave Sunrun's leadership a quantified picture of where its engineering time was actually going, and surfaced $80M in waste in the process.

A couple of other things worth your time below: my interview with Alex Lirtsman of CorralData, who has some sharp opinions about why a decade of dashboards left most companies no more data-driven than they started, and a piece I published on LinkedIn, Proof of Work, on why the case study is the only real signal in a market where brand names lag the actual skill. That argument is also why you'll see this newsletter get a new name in the next few weeks.

Key takeaways from this week's cases:

My thoughts: The number everyone will fixate on is the $1.78 a month. The one that actually matters is the 30 to 45 minutes per technician it gave back, five visits a day, across the whole field force. That idle time was the real cost, and it was invisible until someone built the thing that removed it. 

The setup. A regional fiber ISP on the East Coast ran a 10-person back-office dispatch team that had become a chronic bottleneck. Field technicians had to call in for equipment health checks, troubleshooting guidance, and work order closure, which left them sitting idle in their vehicles for 30 to 45 minutes after 15 minutes of actual work. With a field force handling up to five site visits a day, the compounding downtime degraded both customer experience and throughput.

What BlueLabel built. They re-architected the client's existing Custom GPT pilot at the API layer using OpenAI's Agents API, building two integrated tools. The first is a voice-enabled assistant that gives technicians natural-language access to troubleshooting documentation and equipment data in the field. The second is a dispatch automation workflow that triages incoming service orders, runs hardware health checks against the telecom's OSS/BSS platforms, and closes completed work orders without a dispatcher in the loop.

The return, and where it came from. Removing the call-the-office step cut dispatch calls in half and returned more than $11K a month in labor, while the field force stopped bleeding idle time between jobs. The hard engineering was connecting modern AI tooling to older OSS/BSS infrastructure without disrupting live operations. The decision to move off the Custom GPT tier onto the Agents API is what pulled the monthly infrastructure cost down to $1.78 while serving 150+ technicians. Encoding senior dispatchers' institutional knowledge took the most care, and became the part competitors would struggle to replicate.

Best fit for: Operators in field-service-intensive industries — telecom, utilities, logistics, facilities — where back-office bottlenecks create measurable idle time for frontline workers, and where an internal AI pilot has proven the concept but failed to scale reliably across the team.

→ Want to talk to Jordan? Reply to this email, and I'll make the intro. Or request one directly on the case study page.

My thoughts: You can't manage what you can't measure, and until recently, seeing how work actually moves across a dozen systems was a manual, near-impossible job, so most leaders just guessed. AI is what makes that observation possible at scale, and Parable was one of the first to aim it there instead of at the flashier use cases. The $80M is what that visibility turned out to be worth.

The setup. Sunrun, the largest U.S. residential solar company, had grown fast through the 2010s but struggled to ship product inside its R&D organization. Leadership suspected technical debt was the cause but had no data to prove it or act with conviction. Across the company there was a deep gap between how leadership believed resources were being spent and how work was actually happening, which blocked AI transformation planning entirely.

What Parable built. Rather than another data warehouse, Parable constructed a contextual knowledge graph: a graph database designed to capture not just what happened, but why, who was involved, which systems were touched, and how much time each path consumed. Seven to ten core systems were integrated into a dedicated virtual private cloud — M365, Salesforce, ERP, Jira, Git, and more — with the graph capturing the flow of work across all of them at once.

The return, and where it came from. The graph gave Sunrun's CEO and leadership a quantitative view of how R&D time was actually allocated, and revealed that far more engineering capacity was going to servicing technical debt than to roadmap work. That single evidence-based insight drove both direct cost decisions and AI prioritization. The combination generated over $80 million in savings, and Sunrun's stock 2.5x'd over the course of the engagement.

Best fit for: Enterprises of 500–1,000+ employees (PE-backed, late-stage venture, or Fortune 500) whose executive teams are responsible for operational efficiency and AI deployment at scale — especially companies coming out of hyper-growth that are starting to interrogate how work is actually organized and where AI investment should go.

→ Want to talk to Adam? Reply to this email, and I'll make the intro. Or request one directly on the case study page.

Latest podcast: Alex Lirtsman, CorralData, on why most companies are "data-adjacent," not data-driven. Alex spent 15 years solving data problems by hand for brands like the NBA and Pfizer, then sold his agency and built CorralData to automate the whole thing. He now works with PE firms like KKR, Shore Capital Partners, and Leon Capital across healthcare and consumer. His most contrarian claims from our conversation:

On who to trust with your AI. I published a piece on LinkedIn, Proof of Work: How to Tell Who's Actually Good at AI. The argument: in a market this new, brand names tell you who was excellent at the last game, not this one. The only signal that survives scrutiny is the work itself — a specific, checkable account of what a firm built, what it cost, and what changed. Software always saved the case study for the close of a sale. This market needs it moved to the very start, as a tool for discovery. That's the thesis behind the Library, and the reason this newsletter is about to get a new name.

Background reading on the experts featured above: both Jordan and Adam have been on the podcast — Jordan Gurrieri on turning AI experiments into enterprise systems · Adam Schwartz on how Sunrun found $80M in hidden costs.

Working on something where one of these patterns might fit? Reply to this email. I'll either make the intro to the right expert directly, or open a blinded, Mini RFP brief and come back with 3-5 expert perspectives and pricing in a few days. Both are free, and I read every reply.

— Stu

Just Curious: Applied AI Case Studies. The newsletter of Pluris.