Proof of Work is the newsletter of Pluris, a vetted network of AI implementation partners that help companies build faster, better, and cheaper than the big firms.

Pluris is always free for our clients. Two ways to use us: Browse the Pluris AI Use Case Library, hundreds of real AI implementations from our network: what got built, what it cost, and what changed once it shipped, organized around the problem you're solving. Or submit a need; we'll anonymize it, match it to experts with relevant experience, and return 3–4 perspectives with pricing and timelines.

Axia Partners is a Utah firm that buys RV parks and outdoor hospitality assets, and before it can model a deal it needs a comp analysis, the survey of what comparable parks charge and how full they run. That took 80 analyst hours per park, roughly three weeks of elapsed time, and at this year's acquisition pace the math stopped working. The plan was to hire two analysts just to keep comps ahead of the pipeline. SageCreek built a comp agent instead, side by side with Jake Jones, Axia's acquisitions lead, and Jake ran it against an 80-hour comp he had just finished by hand. It found every comparable he had and about twice as many he hadn't. A park comp now comes back the same day, the full portfolio refreshes overnight for about $200 a month, and the two hires came off the plan.

I asked Connor for a line from Dave Allred, Axia's CEO, on how he got comfortable calling off the hires. Connor's answer was that the call was never Dave's. Jake owned the analyst budget and made the decision himself after checking the agent against his own work. "Two analysts would have been two people to train and check," he says. "This I check in an hour."

SageCreek has a second Axia case in the library. When Axia buys a park it rebrands it under its Anthem umbrella, with a new website, ad accounts, and creative, and that used to take 20 weeks. It now takes 4. Both cases are below. I sent Connor five questions about how they did it, so a lot of what follows is in his and Jake's words. One disclosure up front, which also sits on the case study pages: Greg Butterfield is a partner at SageCreek and a general partner at Axia Partners.

Key takeaways from this week's cases:

If your firm has one analysis everybody has to wait on before the real underwriting starts, and a hiring plan built around it, this is the case for you.

Axia Partners underwrites RV park and outdoor hospitality acquisitions, and a full comparable analysis took 80 analyst hours per park, roughly three weeks of elapsed time. At the firm's acquisition pace that stopped working, and Axia had planned to hire two analysts purely to keep comps ahead of the pipeline. A portfolio-wide refresh across every park took 21 days, so in practice it rarely got run.

SageCreek's method starts before anyone writes code. "Every build starts with a workflow map," Connor told me. "We list the workflows that touch the P&L and rank them by what they cost or what they hold back. Comps came first at Axia because 80 hours per park was gating acquisitions, and that is a revenue number." Then the engineer sits next to the person who runs that workflow and builds each step as he does it. His reason for building beside Jake rather than from a spec: "The judgment calls that make a comp good, like which listings to throw out when the booking data looks off, never make it into a spec. Doing it this way costs Jake a few extra hours up front and saves the months you lose building the wrong thing and finding out in review."

The agent runs on Claude Opus 4.6. Purpose-built browser automation collects booking and occupancy data for the target market from the sources the analysts had been visiting by hand, the agent runs the same analytical sequence they ran, and it drafts the written study. Mapping the workflow took about a week and the build took roughly two months.

The proof came from Jake. He had just finished an 80-hour comp analysis by hand and ran the agent against it as a check. It returned every comparable he had found plus about twice as many he had not, with more information on each. His reaction at the time: "I don't want to be corny, but this is life changing." I asked Connor what actually changed in Jake's week once he could run comps himself. Jake's answer: "Comps used to be the two weeks I had to clear before I could touch the model. Now I run one in the morning and I am underwriting by the afternoon. The time went to the part that actually needs me, the twenty revisions it takes to get the risk right, and to deals I would have passed on because I did not have the hours."

The hiring decision was Jake's as well. He owned the analyst budget, and he made the call after running the agent against his own work. "The two hires were for comp throughput, and that was the bottleneck. The agent does the throughput and I still make every call about what goes in the model." The rest of that answer is in the opening above.

What nearly didn't work, in Connor's words: "The AI was the easy part. Most of the two months went to data. The booking and occupancy numbers live on booking sites, so we started by pulling them with browser automation. It worked, and it was the fragile piece of the whole build. If I did it again I would go to the booking system's API on day one." One more design choice worth knowing about: when the agent cannot fully source a comp, it flags the gap instead of filling it in. Connor calls that the right behavior, and it means some runs come back with holes that Jake fills by hand.

I also asked where he wouldn't use this. "It works when the comp inputs are visible from the outside. RV parks publish rates and availability, so an agent can go get them. If your comps live in broker relationships or behind a paywalled database, the agent has nothing to pull. It is also wrong for a firm doing two deals a year. Eighty hours twice a year does not justify a build. And it needs a Jake, someone who knows the workflow cold and will check the first ten outputs. Without that person you get a tool nobody trusts."

The playbook: build it beside the person who runs the workflow. Four steps, portable to any firm with an analysis that gates its pipeline.

An idea. Ask your acquisitions lead which analysis they have to clear before they can touch the model, and how many hours it took the last time. If those hours are holding up deals, that's the workflow to build against.

Best fit for: Real estate and outdoor hospitality investors running an active acquisition pipeline, where comparable analysis is the bottleneck and the underlying data sits in booking and occupancy platforms rather than a licensed comps database. Strongest where an experienced analyst can sit with the build, and where a completed manual comp exists to validate against.

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

Axia rebrands the RV parks it acquires under its Anthem umbrella, and taking a newly acquired park through brand integration took 20 weeks: website, ad accounts, creative, and positioning. One park could not take an online booking at all. Every week a park spent in that queue was occupancy it never got back.

SageCreek started by replacing the diagnosis step. A scoring tool running on Claude Opus 4.6 grades each park across four areas, revenue management, dynamic pricing, digital marketing, and guest experience, then calculates the revenue upside of fixing what it finds. It replaced the fixed checklist, and a list of defects became a ranked work order. A full run on one park takes about an hour, and it surfaced problems the old process had been missing, including websites built quickly on flash-style tools and analytics installations that had been dead for months.

The scores then drive the work. Ad creative and site copy are drafted for the individual park, against what the scoring says that park actually needs, in place of one master template adapted per property. Integration now takes 4 weeks instead of 20. The ads are park-specific, the no-booking park takes reservations online, and because each acquisition reaches full marketing ramp 16 weeks earlier, every park gains 16 additional weeks of ramped occupancy. Three parks have been scored so far, and Axia runs the tool on new acquisitions as they close.

Best fit for: Multi-site operators and roll-up sponsors that acquire consumer-facing locations and have to bring each one onto a brand system: RV parks, campgrounds, outdoor hospitality, and comparable asset classes where every week of integration delay is lost occupancy or revenue.

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

You can learn a lot about what companies are actually trying to do with AI just by watching which case studies people open.

These are the most-read use cases in our AI Use Case Library this week. Real problems, real numbers, real companies.

Three of the six are PE firms fixing their own deal work. The other three are large companies getting one workflow out of a process that ran on people.

$12.5M of value on one deal. A $3.6B PE firm was assembling company profiles by hand, thousands of hours a year, with two decades of deal history sitting in DealCloud, SharePoint and personal folders. Eliza built a sourcing engine that scores 2,000+ companies against the firm's thesis, a search layer over the archive, and data room review with page-level citations. (Brian Benedict, Eliza)

400+ hours a month of patient education. An ophthalmology practice had staff repeating the same explanation of cataract surgery to patient after patient, and anything built to take that over had to clear HIPAA first. PressW built Sophia, a chatbot that handles patient education by text and voice, fine-tuned to each practice with its data kept separate. (Bryson Greenwood, PressW)

CIM write-ups in under two minutes. A $10B PE firm had ChatGPT Enterprise and an earlier failed CIM project that had cost the team's trust, so usage stayed at email rewrites. Eliza built a set of custom GPTs led by a CIM write-up tool, wired to DealCloud and tested with human evals on real CIMs. 92% of the firm uses it weekly, which is 3,000+ hours a year. (Brian Benedict, Eliza)

Credit risk, three days to 30 seconds. A Tier-1 bank with $2 trillion in assets was losing deals because credit-risk assessment took three days across fifteen-plus systems, with Basel III/IV compliance done by hand. Bolt Group put those sources behind one real-time layer, then ran specialized agents over it in parallel, with scoring regulators can follow. Worth $40M a year. (Dean Sloves, Bolt Group)

Engineering delivery, ten weeks to one. A $1BN specialty insurer had thirty engineers on legacy tooling and AI picked up ad hoc by whoever reached for it. Eliza standardized the environments on GitHub Enterprise first, then embedded AI across coding, testing, documentation and review. Output is up about 5x, and 80 to 90% of work passes on the first go. (Brian Benedict, Eliza)

Ten-plus automations live for 144,000 employees. A $50BN consumer goods company ran thousands of repetitive workflows a day across six geographic zones, each region handling them differently. Eliza built Nexus, a modular agent platform tied into SAP, Salesforce and internal data lakes, and rolled it out zone by zone, with the first of six zones in production since March. (Brian Benedict, Eliza)

If one of these is your problem too, the library is free, and every case has a button that goes straight to the person who built it. Browse the full AI Use Case Library →

💬 Your pediatrician is a great doctor. You still wouldn't let him replace your hip. A PE firm hired an outside team with no experience in its domain, at 50 to 75% above what the work should cost, and couldn't tell. Overpaying is the smaller cost. The bigger one is a competent answer to a problem adjacent to yours, which becomes the ceiling on everything built on top of it. → Read it

💬 Proof of Work is back this fall. This we’ve got a great slate of investors, operators, and advisors. Teaser: I just sat down with David Yoken, who runs AI transformation at Frontline Road Safety Group, a Bain Capital business built on trucks, paint, and crews. His argument is that the least glamorous, most physical industries are where the most AI value is still sitting untouched, and that the labs overshooting toward autonomous agents is the opportunity for everyone else. → Read it

💬 The next 10 to 15 years will be the biggest value-creation window most of us will ever work in. Every company is sitting on a pile of problems it gave up on because the fix was impossible, too expensive, or too tangled to understand. All three excuses are disappearing at the same time. → Read it

Working on something where one of these patterns might fit? Reply to this email and I'll either make the intro to the right expert directly, or submit a need and I’ll 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.