Cutting a Nine-Month Process to Thirty Days

Joe Johnson’s in the world of AI, but he’s busy talking about fire trucks, commercial refrigerators, and surveyor PDFs: the specific, grinding problem of getting institutional knowledge out of people's heads and into software. He's been a VP of Cloud Operations, a Director of IT, and someone who's run M&A integrations. He's currently building the software and data function at Perry Homes, one of Texas's largest private homebuilders, 50+ years in business, roughly 4,500 homes a year across Texas and Florida. Before that, he was at Revalize, a commercial SaaS company serving complex manufacturers, and that's where his most interesting AI work happened.

You can find Joe on LinkedIn.

$50k of Software, But It Needed $200k Of Services To Set Up

Revalize sells CPQ (Configure, Price, Quote) and PLM software to complex manufacturers: companies making commercial kitchen equipment, fire trucks, specialized industrial machinery, where a single unit can have 1,200 configurable options and certain combinations simply can't work—a two-inch flange cannot go into a four-inch pipe on a fire truck; a left-opening door doesn't fit on a right-opening fridge.

Revalize's software handled the complexity. Getting customers into it was the problem.

Buy $50,000 of software from Revalize and you'd often need $200,000–$300,000 of consulting to properly implement it, because there are all those little rules. The implementation process meant sitting with a customer's most experienced salespeople and downloading everything they knew: every option combination, every incompatibility, every edge case. Six to nine months, and that process was the single biggest obstacle between a sale and a live, revenue-generating customer.

Most of the knowledge wasn't written down. It lived in the heads of salespeople who had quotas to hit. "Am I gonna sit down and spend eight hours downloading my brain to these consultants from the software company," Joe says, "or am I gonna go out and sell fire trucks? I'm gonna go out and sell fire trucks."

The Solution: Give the LLM the PDFs

Joe's team started with a simple observation: all of this stuff is in the product PDF. So they fed the PDFs into an LLM, asked it to generate an Excel option compatibility matrix (checkbox for valid combinations, X for impossible ones), and handed it back to customers to validate.

A nine-month process became roughly thirty days. "From an adoption standpoint, I can tell you, like, it was instant hockey stick," Joe says. "Everybody started using it."

Sales capacity became the new constraint. "The biggest problem we had now is we couldn't sell fast enough to keep up with our ability," he says. "Sales, go sell 500 more customers, because now we have the capacity."

Before a customer had even provided their own data, the team could pull publicly available product information and get to 60–65% accuracy, enough to preload a demo environment with the customer's actual catalog before the first meeting. "It blows the customer's mind that we put that much effort into it," Joe says, "because they're thinking we sat someone down to click all the if-then statements in the app instead of just using AI to cheat it."

But Then: The ROI Question

Joe doesn't stop the story there.

The stated goal was to get a nine-month implementation down to three months, prove it with a real customer, and show results. They hit it: kickoff call to the customer processing orders, three months and two days. "Close enough for government work," he says.

Then came the harder question: what did that actually save? Revalize was doing four implementations a year. They'd spent $70–80K of people's time proving out the process. "It didn't really feel like a big payoff in the end," Joe says, "because we picked something that sure seems bright red flashing emergency, let's get rid of this, let's solve this problem, but at the end of the day, the level of effort versus the return was not there."

His advice to builders: find the pain point, then put a number on it. "What is the pain point and the number associated with the pain point of your success? What is the happy path and in that number?" A technically impressive result on a problem that only affects four customers a year is still the wrong result. "Is it worth $60,000 worth of developer time to save something that's going to maybe save $5,000 a year?"

Joe's instinct throughout the project was to keep pushing beyond the 65% pre-sales accuracy, toward something close to a zero-human implementation process. His colleagues wanted to call it done. "You have to aim for the 10X goal, knowing you're gonna miss and hit the 2X goal," he says. "But you have to be audacious."

Another part of the ROI math: as model pricing has risen, the math has started inverting for companies that over-indexed on expensive frontier models for tasks that don't require them. "It's now cheaper for me to outsource development back to the Philippines than it is to pay my token budget to go do development using AI," Joe says. His prescription: be intentional about model selection. "Maybe Sonnet 4.5 is good enough. We don't need to use Opus for everything."

The 96% Problem

Joe also has strong takes on Copilot rollouts and the meaningfulness of “adoption.”

Here’s what he saw: under pressure from private equity backers or aggressive CEOs to "have AI,” companies bought Copilot. The logic was defensible: they already had significant Microsoft spend, and the capability gap between Copilot and ChatGPT or Claude wasn't enormous a couple years ago. They didn't thinking about the lock-in: a three-year contract, Microsoft ecosystem only, no clean path to switch as models evolved. "They later come to regret that," Joe says.

A year in, 96% of employees are using Copilot weekly, averaging about 30 interactions per employee per week. Microsoft's reporting showed roughly $80,000/month in time savings.

"And that translates to $0 of savings. People aren’t more productive. Nor are they reducing headcount. Nothing really changed."

Here’s how he puts it: "96% of people use the bathrooms at the office every day. Does that mean that makes us a more valuable company because we have bathrooms? Maybe not having them would make us a less valuable company. But it's meaningless."

A platform that reports on usage will always show usage. Whether anything about how the organization operates has actually changed is a different question, and it's the one that matters. Joe measures by workflow change.

The Dirt

Three and a half months into Joe's tenure at Perry Homes, the most concrete AI win is about ordering dirt.

Before construction can begin on any new home, the land has to be level. Surveyors come out, measure the grade across the entire plot, and produce a PDF: a picture of the land with numbers scattered across it showing height differentials. Someone takes those numbers and runs them through a calculator to figure out how much infill dirt to order. Get it wrong and you either don't have enough to finish the pad or you've paid for a truckload you can't use. The error rate was running at about 5–6%.

Joe's team built a process using Azure Document Intelligence to extract the numbers from the surveyor PDFs and pipe them directly into the calculator. The error rate dropped to 0.5%.

"People imagine robots building 3D-printed houses," Joe says. "And it's like, no, we just order dirt better than anybody else right now."

Perry Homes closes 4,500 homes a year, and every construction delay costs money on two fronts: carrying costs while the company waits for closing, and the very real customer pain of a home that isn't ready when promised. A homebuyer expecting to close on July 15th who gets pushed to July 25th may have already ended their lease, scheduled movers, arranged bridge financing. A dirt ordering error costs more than the price of an extra truckload.

"The most successful things," Joe says, "are probably the least sexy of all of it."

The Training Pyramid

Joe built his AI training program at Revalize and is now rebuilding it at Perry Homes. Three tiers: Foundation, Fluency, and Mastery.

Foundation is mandatory for everyone who touches a computer as part of their job: what is AI, how do LLMs work, how do you structure a prompt, how do you recognize when the model is lying to you. At Revalize, the goal was 100% completion across roughly 800 employees globally, about 20 hours of training in total, by May 31, 2026. Joe says his former team hit it.

Fluency is for managers and people leaders: how to integrate AI into team workflows and identify where it actually creates leverage.

Mastery is for a very small number of people, around five at an 800-person company. External graduate programs, AI certificate coursework, the ones who become internal experts and force multipliers for everyone else.

The design is deliberately asymmetric. "You don't need 800 AI masters," Joe says. "You need 800 people who aren't afraid of it, and five who are obsessed."

He frames AI literacy the same way he frames web search: "I trust people who know how to use a search engine. Now I need to trust people who know how to talk to AI and to do it efficiently." Foundation gets field construction workers to the point where they can use the tools the five obsessives build for them.

Citizen Development at Perry Homes

Joe is also standing up what he calls "citizen development" at Perry Homes, a term he prefers over "vibe coding," which he says undersells the rigor involved.

Perry Homes' product owners (their term for IT project managers, reflecting a philosophy of ownership over execution) are now using AI-assisted tools to build working UI prototypes directly with business stakeholders: actual clickable interfaces, built in real time during the conversation.

"They don't know what they don't know," Joe says of business stakeholders. Getting good product requirements from someone who's never built software is hard, because they can't easily articulate what they need from a blank page. Show them something on screen, even rough, even obviously prototype, and the feedback gets specific fast.

From there, those requirements flow into a structured pipeline: BSAs generate PRDs, ERDs, and API contracts formatted specifically to serve as context for Cursor (the AI-powered development environment Joe's team uses). Developers then use that documentation as the input for code generation, with significantly less back-and-forth between requirements and implementation.

"We're front-loading that investment of time into getting documentation and context really well architected," Joe says, "to create high-quality software."

The principle is the same as the Revalize work: the quality of what the AI produces is almost entirely determined by the quality of what you give it. Joe's team gives Cursor a fully structured, human-reviewed specification document before a line of code is written.

Find Joe Johnson on LinkedIn. Learn more about Perry Homes at perryhomes.com.