In July 2026, Anthropic launched Ode with Anthropic , a $1.5 billion AI implementation company that embeds engineers inside mid-sized enterprises to build custom Claude systems. Similarly, OpenAI launched the OpenAI Deployment Company , a $4 billion joint venture whose 150 forward-deployed engineers do the same work. Both AI labs are building standalone deployment companies, a sign that value capture is moving up the stack, from training frontier models to integrating them inside real-economy businesses.

That work opens a part of the economy the labs could not reach through API access alone. Of the roughly 36 million companies in the U.S., only about 17,000 are software companies . The rest are dominated by services businesses; in fact, the services sector accounts for more than three-quarters of U.S. GDP , the largest share of the economy, and it represents the next opportunity for AI value creation.

Over the past two years, I have had a standing biweekly call with Luke Hedlund of Tucannon Partners , a search fund built to acquire and run services businesses. He has evaluated literally hundreds of these companies as a buyer, increasingly based on the question of how they are (or are not) using AI to run their operations.

As Luke describes them, many services businesses still keep 30+ years’ worth of financials and customer records "held entirely on paper in filing cabinets." Large language models, trained mostly on public web data, have never seen those records: the maintenance logs of a regional HVAC operator, the claims patterns of a rural health network, the pricing methodology a family-owned distributor keeps in its ledgers, and countless other examples. These are decades of quotes, margins, customer outcomes and other written records for any given sub-sector, representing a proprietary record of how that business actually operates, and no model has been trained on any of it.

One would naturally be wary of calling that a moat, because digitizing still carries a cost. Analog businesses traditionally have traded at a discount to their digital peers, because converting their paper records into a usable, digitized system has been a manual and multi-year expense a buyer would have to underwrite before extracting the additional value. AI solutions now handle much of those digitization workflows at a fraction of the time and cost, which lets many of these companies move straight from analog to AI-powered, skipping the interim software era entirely. That should compress the discount they sell at, but to date it has not.

As Luke reads it: “The investors, search funds, and sponsors who know how to buy and operate these companies are just beginning to build AI capabilities, and the world-class engineers they need are drawn to software rather than services companies. The opportunity belongs to whomever can bring the technical skill and the process knowledge together, and that combination is still rare.”

In most services businesses, technical skill is necessary but rarely the deciding factor. Consider HVAC, one of the largest of these trades: any established operator can fix the unit well, so the repair itself is not what wins the customer. What separates them is the operation around it, how quickly they answer a service request, how densely they schedule, how sharply they price, how reliably they follow through. As Luke puts it, "How fast do you pick up when someone’s AC dies in July? What surrounds the craft, the answering, the scheduling, the follow-up, is where a services company wins or loses. Getting the back-office operations right all the time is hard, but it isn't what makes customers choose you, though it can make them lose you."

Only recently have these coordination tasks become areas where companies can truly compete, not simply reduce costs: low enough latency to hold a live phone call, and high enough accuracy to run with minimal human supervision. The companies pulling ahead are deploying voice AI agents to answer and triage inbound calls around the clock, LLM-based schedulers that optimize dispatch and routing against technician skill, location, and job urgency, and automated quote-to-invoice pipelines that generate line-item estimates for parts and labor. Each is grounded through retrieval on the company's own job history, so the system prices and answers from how this business has actually operated, not from a generic model.

Luke has worked with the founders of many "guys and trucks" businesses, and the ones pulling ahead share this trait: an internal champion who owns the AI implementation rather than outsourcing it to a vendor.

What all of this removes is the significant dependence on the business owner. In these firms, the pricing, the scheduling and the customer relationships run through the founder and a few key management hires, so the company cannot scale or change hands without them. That is a classic key-man risk that surfaces in any buyer’s diligence, and it drives one of the largest discounts on an asset. The same AI buildout that makes the business run better is also what reduces or eliminates that discount, making the company much more readily transferable to the next owner.

For the last few years, the returns on AI investment have been concentrated in two places: the labs that build the models and the companies that make the chips. The next leg comes from the services businesses that run on those models, once the technology is integrated deeply enough to change how they operate. That integration is early, and the operators who can execute it are few, which is what keeps the opportunity open. As Luke puts it, “The technology is available to everyone. What decides its impact is whether the people running the company can do the hard work of evolving decades-old processes to embrace the opportunity.”