A bit of news before the cases: we're renaming this newsletter from Just Curious to Proof of Work: Applied AI Case Studies. The name borrows from the idea behind Bitcoin — proof of work, a credential that's expensive to produce and trivial to verify, where the effort itself is what you trust. In a market where no one yet knows who's genuinely good at AI, brand names only tell you who won the last game. An honest case study, with real numbers a stranger can check, is the signal that actually holds. That's what we collect here, so we named the newsletter for it. The full argument is in our essay.

This is the newsletter of Pluris, a network of pre-vetted applied AI experts and AI-native service organizations that specialize in helping businesses design and implement AI solutions that drive value.

Two free ways to use us. Browse the AI Use Case Library for real builds from our network; what got built, what it cost, and what changed once it shipped, organized around the problem you're solving. Or submit an AI need through a Mini RFP. We'll anonymize it, match it to experts with relevant experience, and return 3–4 perspectives with pricing.

Two case studies this week, both in financial services, both facing the same high-level challenge: making AI reliable enough to trust with a regulated, high-stakes decision. One is a bank deciding who to lend to; the other is a venture-backed software company whose finance copilot answers CFOs' questions about their numbers. The two builds are very different. Dean Sloves and Bolt Group rebuilt a Tier-1 bank's credit-risk desk and took a three-day assessment down to about 30 seconds. Unmukt Raizada and TrustEvals.ai wrapped an eval harness around a finance copilot that was stuck at 60% accuracy, and lifted its net revenue retention from 82% to 144%.

Key takeaways from this week's cases:

My thoughts: The ratio is the story. A $1–5M build set against the ~$40M a year in lending revenue this bank was losing to slow decisions is a seemingly rare payback, and it's worth holding up against how your portcos scope AI spend. The gain came from decision speed rather than cutting headcount, which is part of why it holds.

The setup. A Tier-1 bank managing over $2 trillion in assets was losing deals to speed. Credit-risk assessments took three days, Basel III/IV compliance was manual, and 15-plus data sources each sat behind their own interface. Roughly $40M a year in lending revenue was leaking out through delayed decisions.

What Bolt Group built. Four compounding moves. They consolidated 15+ data sources — mainframes, Bloomberg, credit bureaus, CRM — behind a real-time API layer on Kafka and Snowflake, so every downstream agent worked off one clean feed instead of fifteen interfaces. On top of that mesh they ran a swarm of specialized GPT-4o agents (via Azure OpenAI) that split each assessment into parallel tracks: data extraction, financial analysis, regulatory checks, and report generation. A proprietary PyTorch scoring engine trained on a decade of credit history produced the actual risk grade. Basel III/IV stress tests and audit trails were generated automatically on every assessment.

The return, and where it came from. A credit assessment that took three days now lands in about 30 seconds, and the lending revenue the bank had been losing to slow decisions — roughly $40M a year — came back within reach. Compliance overhead fell about 80%, because Basel III/IV stress tests and audit trails are now generated on every assessment instead of by hand. The reason it could ship at all in a regulated setting: the risk grade came from the purpose-built ML engine, not the agents, so it stayed explainable to regulators, where a black-box score is a non-starter. The whole build reached production in sixteen weeks, principal-led, with senior engineers on the keyboard from week one.

Best fit for: Large, regulated lenders — Tier-1 and mid-size banks — with slow, manual, multi-source credit-risk workflows and heavy Basel III/IV compliance overhead, where decision latency rather than analyst capacity is the constraint on lending revenue.

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

My thoughts: If you're buying AI, or backing a company that sells it, one question cuts through the noise: how do you measure accuracy, and on what data? TrustEvals graded every answer on held-out data, the way a CFO grades a forecast, and that reliability is what moved the product from churning to 144% net revenue retention. Demos are not evals.

The setup. An AI-native FP&A software company sold a natural-language finance copilot into finance teams, including Fortune 500. It stalled at about 60% accuracy. Wrong answers were costing deals, and prospects were choosing more AI-native rivals.

What TrustEvals built. An AI workbench and evaluation harness wrapped around the copilot. A semantic data dictionary mapped the customer's financial data. A router classified each incoming question and sent routine ones through templated fast paths built for that question type. The hardest asks — forecasts and what-if scenarios — went to a separate predictive engine rather than trusting the model with math. Around all of it, a testing layer scored every answer against custom evals on held-out data, ran security checks, and compared OpenAI, Claude, and Gemini head-to-head to pick the best model per job.

Start the evals on day one. Nothing reached a customer until it passed the eval, and accuracy was judged on held-out data the way a CFO judges a forecast, not on a one-time demo. Validating the data dictionary up front surfaced gaps that were fine for a human analyst but left the model short on context. Human-in-the-loop flows let the team update the eval set and remediate weak answers immediately. The harness took accuracy from 60% to 95% and NRR from 82% to 144%, delivered in 4–8 weeks, now live with 100+ enterprise customers.

Best fit for: Any team building a customer-facing platform with natural-language-to-SQL at its core — sharpest in finance — where the data sits in a real structured database with a deep semantic layer, multiple use cases or tenants need support, and accuracy is the bar rather than a nice-to-have.

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

🎙️ Next week: my interview with Doyl Burkett, Managing Partner of Integrity Growth Partners (~$800M AUM). Doyl's thesis is that AI won't compress private equity evenly, it sorts it: companies fall into three buckets — AI-native winners, defensible-moat survivors, and undifferentiated losers — and the same sorting hits the GPs investing in them. He's running that experiment inside his own firm with an AI-native sourcing and reporting stack he calls CAPE.

It's the investor-side companion to my conversation with Tom Scott, CEO of PE-backed Wrike, who made the same argument from the operator's seat: this is a rebuild-the-company moment for software businesses, and the next two years are about execution variance between firms, not uniform disruption. Two sides of a similar coin, Doyl sorting from the GP's chair, Tom rebuilding from the CEO's. → Listen to the Tom Scott episode

📚 Both cases this week are financial services. If you missed last week, two builds in very different industries: a regional fiber ISP cleared a dispatch bottleneck that left field technicians idle 30–45 minutes between jobs (BlueLabel's Jordan Gurrieri), and Sunrun, the residential-solar company, used a knowledge graph to surface $80M in hidden R&D waste being eaten by technical debt (Parable's Adam Schwartz). → Read last week's issue

💬 Every week I publish more case studies on LinkedIn and X. This week: how a 60-person SaaS marketing team rebuilt its workflow, featuring Tahnee Perry. → Read it

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 Mini RFP brief and come back with 3–4 expert perspectives and pricing in a few days. Both are free, and I read every reply.

— Stu

Proof of Work: Applied AI Case Studies. The newsletter of Pluris.