From AI Prototype to Production: How Wrench Group Decides What Is Worth Scaling
About the guest:Tony Ferreira is Senior Director of Technology, Data, and Product at Wrench Group, where he runs technology, data, product and AI work for a national group of HVAC, plumbing and electrical brands.
AI has cut the time between an idea and a working prototype. A team can now build in a few weeks what used to take most of a year.
That speed brings a new problem. When building is easy, companies end up with more demos, internal tools and app proposals than they can responsibly keep running. A prototype can look convincing without showing that employees will use it, that it works with the company’s real systems, or that it’s worth what it costs.
Tony Ferreira, Senior Director of Technology, Data, and Product at Wrench Group, has worked out a step-by-step path from experiment to production. Teams are free to explore, but an idea has to prove itself before it becomes part of how the business runs.
A promising idea first goes into an isolated innovation lab set up to look like Wrench Group’s production environment. If it works there, the team may show it to three users and watch how they actually use it. Next comes a pilot at one operating brand. After roughly 90 days, Wrench Group should know enough to expand the product, change it, keep only part of the idea, or stop putting money into it.
This lets the company learn from real use without treating every AI-built prototype as a permanent product. One of the best examples is a live operating map Tony’s team built to tie together activity across roughly 30 home-service brands.
The Map That Proved the Process
Wrench Group is a $2 billion-plus, private-equity-backed group of HVAC, plumbing and electrical businesses in markets across the United States. Each local company keeps its own president, dispatchers, call center and operating team. Tony’s organization provides the shared technology, the data infrastructure and an AI operating model that help those businesses run better.
That setup makes data hard. Wrench Group has customers, booked jobs, service histories, technicians, inventory and trucks spread across many brands that run independently. The information existed, but it sat in separate systems and was hard to pull into one view of the operation.
Tony’s team started with a simple prototype: put customer locations and service histories on a map. The first version showed where customers lived by ZIP code and flagged people who hadn’t had routine maintenance in a year or more. The team showed it to internal users, asked whether they’d use it and how it should work, and changed the interface based on what they heard.
Once the idea proved useful, the team added more operating data. The map started showing booked work, capacity, technician routes and the live location of trucks across every brand. A prototype for looking at customer activity turned into a shared operating tool for several parts of the business.
The same system now does different jobs for different people. A dispatcher sees trucks and upcoming appointments. A technician sees the day’s assigned jobs. A marketer finds customers who are overdue for service or live near an active route. A business leader looks at growth and market coverage by ZIP code.
One of the clearest uses starts with a canceled appointment. The dispatcher can see where the technician is, find nearby customers who are due for maintenance, and offer one of them an opening right now. The truck and technician don’t sit idle, and a customer who already needs the work gets faster service.
In home services that matters, because availability often decides who gets the job. When a pipe is leaking or the air conditioning dies in a heat wave, nobody waits a week for the company with the best-known brand. Whoever can show up today usually wins.
Better coordination can also cut down on costly repeat truck rolls. If the company knows the day’s jobs, the likely equipment problems and the parts needed, it has a better shot at sending the right technician with the right part the first time. One extra trip is manageable. Repeat visits across hundreds of technicians add up to real money.
Further out, the map could become proactive. Wrench Group is adding weather information to its customer data platform. Later connections to equipment diagnostics and connected-device data could help the company spot likely problems sooner. A local team might know that a bad storm is coming, that a customer hasn’t had a generator or heating system checked, or that connected equipment is reporting something unusual.
The technician still does the work in the customer’s home. AI pulls together the information that gets that technician to the right place before a small problem turns into an emergency.
Tony said the first version of the map took a few weeks to build, alongside the team’s regular work. A similar project might once have taken nine months and still left important data disconnected. The team then tested it with business leaders, improved it, and added to it based on what operations actually needed.
That path is why the map matters. It showed how fast AI-assisted building can be, and also the discipline it takes to turn a quick prototype into something the business can rely on. Its success also set off a wave of new ideas from technicians, dispatchers, marketers and business leaders across Wrench Group.
The Rest of the Discussion
- AI access should expand with demonstrated need. Wrench Group’s board wants AI used widely, but Tony has seen how fast costs climb when everyone gets the most expensive tools by default. Employees start with basic access, try it on a real task, and explain the limit they’ve hit before they move up to a more capable tool.
- The observability layer is still incomplete. Wrench Group uses tools from Microsoft, Anthropic and others. Their admin views show some usage, but Tony can’t get the full picture across platforms that he wants without buying expensive enterprise packages or stitching several systems together. Cost data alone can’t tell you whether the work is worth anything, or whether someone is moving AI-written code into an environment nobody approved.
- Cheap software creation increases the importance of business judgment. With AI it’s easy to build something technically impressive. Tony’s first question is what practical good it does. A production system needs a clear user, a real need and an impact you can measure. Being interesting or looking good doesn’t count.
- Established workflows move to less expensive models. When Tony sets up a new task, he often uses a very capable model while he learns what a good result looks like. Once the workflow is stable, he tries cheaper and cheaper models until the output stops meeting the bar. That way the company spends on its most expensive models only for work that’s still unclear.
- Home-service businesses increasingly operate like technology companies. The service is still physical, but technology decides how technicians are scheduled, what information they get, how fast the company responds and how reliably it keeps its promise to the customer. That layer can set a company apart from its competitors.
- Wrench Group is becoming an internal AI incubator. Once the map showed that a useful app could go from idea to working product in weeks, employees started suggesting more ideas for technicians, dispatchers, customers and business leaders. Tony’s team now has to choose which ones go into the same testing process.
- Builders should begin with the outcome, not the novelty. Tony’s closing advice was simple: customers rarely care that a product is cool. They care that it helps them get a task done or makes their work easier. AI has made building faster, so understanding the business value matters even more than it used to.