AI Ops Agents Just Landed in Field-Service Software: What a Trade Company Should Own vs Rent
Rent the models and the bundled agents. Own the job record, the rules, the tests and the accounts. A founder's own-vs-rent test for AI in the trades.
Short answer: AI agents are starting to ship inside the software trade companies already pay for. That changes the question owners should ask. It is no longer "should we use AI?" It is "which parts of it should we own, and which should we rent?" My rule is simple: rent the intelligence, own the memory. Rent the models, the agents bundled into your field-service software and the new booking channels, because their makers improve them faster than any trade company could. Own four things no vendor should hold for you: your job record, your rules, your tests, and the accounts your agents run on. That is my operating rule, not a proven law. The best evidence I found favors renting over building, says almost nothing about the hybrid I describe here, and none of it comes from small home service companies.
What just changed in field-service software
In early November 2026, several vendors announced AI agents for software trade companies already use or can sign up for. What follows is how the vendors describe their own products. I have not tested any of them.
Simpro opened early access to JustDo on November 9. It is an AI operations agent inside Simpro's Lightning platform. Simpro says it drafts jobs, customers and quotes from emails, photos, PDFs and spreadsheets, works through Simpro's own APIs within each user's permissions, and waits for approval before sensitive actions. Two details matter for this essay: JustDo comes only with the Lightning platform upgrade, and it draws on a credit allotment. When the credits run out, it stops until you buy more.
Motive launched Agents for field services teams, including HVAC and plumbing, on November 9. Its chief product officer explained the logic in the release: "The data to prevent these issues already lives in Motive." The company says that early next year managers will be able to set a goal in plain language and let the agents work out the path.
Handle opened a waitlist on November 10 for an agent built for local service businesses. Its site says the agent "remembers the details, preferences and decisions that matter to your business" and that "your business data stays yours."
Jobber's help page, dated November 4, says Google's AI Mode can recommend a Jobber customer's business and let people request or book work directly, with no lead fees, for paid U.S. accounts. The page also says reviews help AI Mode decide which businesses to recommend.
I covered which workflows to hand to AI first in an earlier essay on dispatch, pricing, diagnostics and the job record. This one is about what these launches force on every owner: who owns what.
Why the own-versus-rent line just moved
For a long time the decision was easy for a trade company: you rented software. The rented system held your records and switching was painful, but your people did the thinking.
Agents change what you are renting. When an agent drafts your quotes, learns your preferences and works on data that already lives inside the vendor's product, you are no longer renting only screens and storage. You are renting a piece of judgment that has been tuned on your business. If that judgment lives only inside one product, the cost of switching is no longer just moving records. It is losing the part of your operation that learned.
That is my reading of the launches, not something any vendor says. The vendors above describe approvals, permissions and data policies that look reasonable. The risk I see is quieter: a company that cannot leave a product in two years, because everything that makes the agent good at its business exists only there.
What I decided for my own company
I began my career in the trades as a technician. Today I run a home service company, and I have built much of our AI setup myself. So this is a decision I have had to make, not a theory. Three positions came out of it.
Our operating data is not for sale. In a recent conversation about AI, I put it this way: handing your data to people who want to build AI on it is like sitting on a branch and cutting it off, because your data is what they need to replace you. I would rather use that data to make our own company more valuable. That is a founder's conviction, not an analysis. Some owners will reasonably decide that selling or sharing data is worth it.
AI is not something you casually share. On a call with other service business owners, I said you cannot really share one AI between people the way you share a tool, because it gets confused. When someone asked how a team could work together in one chat, my answer was levels: separate accounts shaped to each person's role, which can still share the same memory. Then control and monitor how they are used, including how other people use them.
The agents that do real work belong to the company, not to whoever set them up. When someone leaves, their AI work history goes into a shared company archive before the account is closed. The direction I am leaning for the longer term is a split. Managers experiment on rented tools, under company control. The agents that run production sit on infrastructure the company controls, with permissions, cost tracking, and the ability to switch the underlying model. That is a direction, not a finished build, and I may be wrong about how much of it a company our size should own.
Notice what is missing: I did not decide to build our own models or to replace the field-service software we rent. The ownership I care about is narrower than "build it all yourself."
What the evidence says about building versus buying
The best evidence I found cuts against founders who want to build everything, me included.
Buying reached deployment more often than building, in one sample. MIT's Project NANDA published a preliminary report on generative AI in business in July 2025, covering research from January to June 2025. In its interviews with 52 organizations, "external partnerships with learning-capable, customized tools reached deployment ~67% of the time, compared to ~33% for internally built tools." The authors state the limits themselves: the outcomes are self-reported, success was defined differently across organizations, and the gap "may reflect organizational capabilities rather than implementation approach alone." For the hybrid, where an internal team co-develops with a vendor, they report "insufficient data to quantify." So, on getting a tool into real use, the study favors renting over building. Note what it measured, though: partnerships on customized tools that learn, not off-the-shelf features switched on inside existing software. It is silent on the split I lean toward. And the organizations are not small home service companies.
The price of intelligence is falling fast. Stanford's AI Index 2025 reports that the cost of querying a model that scores at the level of GPT-3.5 on a standard benchmark (MMLU) fell from $20 per million tokens in November 2022 to $0.07 per million tokens in October 2024, a more than 280-fold drop in less than two years. That is one benchmark threshold, measured on list prices, with data ending in 2024, but the direction is clear. Building your own model would be a bet against a falling price curve. That is why I rent the model.
The model you rent can change under you. Researchers at Stanford and Berkeley compared two dated versions of the same AI service in 2023. On one task, identifying whether a number is prime, GPT-4's accuracy went from 84% in March 2023 to 51% in June 2023. Those are 2023 models and a narrow test, but the lesson holds: what you rent is not fixed, so you need your own way to check it.
And there is a gap. I found no published study of own-versus-rent decisions for AI in small home service companies. Everything above comes from organizations outside the trades or from AI in general. The rest of this essay is my reasoning from that evidence, not a finding.
The line: rent the intelligence, own the memory
Here is where I draw the line, as an operating rule.
Rent:
The models. They keep improving and their price has been falling. A trade company has no edge in building them.
Commodity agents inside software you already run. Turning emails into draft jobs, reading fleet fault codes, making reminder calls. When a vendor already holds the data and the task looks the same in every company, in my view a bundled agent will usually be cheaper and better maintained than anything you would build. Motive's own line explains why: the data already lives there.
New demand channels. You cannot own Google's AI Mode or a marketplace. You can own what they read about you: an accurate business profile, real reviews, honest availability and a booking path that works.
Own:
The job record. Every job, the equipment, what was found, quoted, done, charged and how it turned out, in a form you can export and read without the vendor. I argued in The Agentic CRM that the record, not the screens, is the part of a CRM that matters. Agents make that more true, because the record is what they learn from.
The rules. Your price policy, your discount limits, what an agent may do alone and what waits for a person. Write them in your own documents and load them into whichever product you use. If your rules exist only as settings inside one vendor's product, you have rented your own judgment.
The tests. A set of real past cases with known right answers: calls, quotes, diagnoses. Run any agent against them before you switch it on and again after every vendor update. It is a cheap protection against a model that quietly changes.
The accounts and the owners. Agents, their memory and their access keys should sit on company accounts, not on one employee's login. Each one needs named owners, one for the result and one who keeps it working, as I wrote in Every System Needs Two Owners.
An illustrative example, not company data: an HVAC company switches on a quoting agent that comes bundled with its field-service software. Renting that agent is fine. What the company owns is the price book and discount rule the agent follows, kept in its own document, plus a test set of, say, 50 past quotes with the prices a senior estimator approved. If the vendor raises prices or the agent gets worse after an update, the company loads the same rules into another product and runs the same 50 quotes. It loses weeks, not its pricing knowledge.
A three-question test for any AI decision
When a new agent shows up, ask three questions.
Is this where we are different? Data entry from emails is much the same in most companies. Your diagnostic knowledge, your pricing and your customer history are not.
If this vendor doubled its price or shut down next year, what would we lose that we could not rebuild in a month? If the answer is "nothing much," rent it without worry.
Can we check its work without the vendor's help? If not, build the test before you trust the agent.
If the task is common, nothing important would be lost, and you can check the work, rent it. If any answer points the other way, own the memory around it: the record, the rules and the tests. Owning rarely means building the agent. Usually it means keeping the agent replaceable.
The strongest case against owning anything
The best counterargument is not "AI is hype." It is "rent everything, and stop pretending a trade company is a software company."
A typical contractor has no engineers, and anything you own, you must maintain. Engineers at Google argued in 2015, from their own experience, that machine learning systems carry hidden technical debt and commonly incur massive ongoing maintenance costs. The MIT sample says internal builds reached deployment about half as often as external partnerships. An owned layer could become a half-maintained shadow system nobody understands once its builder leaves, the very key-person risk I say I am avoiding. Data has little value if it is messy, so owning a dirty record is not an asset. And on the demand side, platforms may favor businesses that stay inside their partner networks.
Most of that is right, and it changes my conclusion in one important way. The case against is strong for building agents. It is much weaker for owning the record, the rules and the tests, because those are mostly documents and discipline, not software. So my rule narrows: own the memory, not the machinery. If owning something requires engineers you do not have, rent it and keep the exit clean.
When this rule fails
Very small shops. For an owner with a few technicians, writing rules and maintaining test sets may cost more than it saves. Rent the bundle, keep an export and a written price list, and revisit later.
When the vendor's data is better than yours. A telematics company holds sensor data from its own devices. Owning a copy may not mean much.
When contracts block the exit. If an agreement limits exports or the data comes out in a format nobody can use, ownership is a legal question before it is a technical one.
When your record is poor. Owning a record full of gaps and duplicates protects nothing. Clean it first.
When the technology outruns your rules. Rules and tests written for today's agents will need rewriting. Owning them is ongoing work.
What this changes for owners and investors
For owners, AI spending becomes a capital allocation decision, not just a list of subscriptions. The rented part shows up on a bill. The owned part, the hours spent writing rules, cleaning the record and keeping tests current, shows up on no invoice, so it gets skipped. Budget for it on purpose.
For investors and acquirers, one diligence question gets more important: which AI assets actually transfer with the company? Agents that run on a founder's personal account, rules that live in one person's head or inside one vendor's settings, and records that cannot be exported are hard to transfer. I am not claiming that owning the memory raises a valuation. I found no evidence on that. I am saying it decides what a buyer is actually getting.
How to test this in 30 days
Week 1. List every AI tool and agent in the company. For each one, write down the account it runs on, its owner, and what it would take to leave.
Week 2. Run an export test on your field-service software or CRM. Can you get jobs, customers, equipment, quotes and notes out in a file someone else could use?
Week 3. Write your most important rules in a document the company owns: price policy, discount limits and what needs a person's approval.
Week 4. Build a small test set from real past cases, for example 20 quotes or calls (an illustrative number), and run your current agent against it. Repeat after every vendor update.
At the end of the month, two numbers tell you where you stand: how many agents run on company accounts, and how long the exit question takes to answer for each.
Frequently asked questions
Should a home service company build its own AI agents or buy them?
For most home service companies, buy or rent the agents, especially for common office tasks. The best evidence I found, a 2025 MIT report based on interviews with 52 organizations, found that external partnerships reached deployment about twice as often as internal builds, though the results are self-reported and do not come from the trades. What a company should keep for itself is the record, the rules and the tests the agents depend on.
What should a trade company own when it uses AI?
Four things: the job record in an exportable form, the written rules agents follow (price policy, discount limits, approval thresholds), a test set of real past cases to check any agent, and the accounts the agents run on, each with named owners. That is my operating rule, not an industry standard.
What is AI vendor lock-in for a field-service business?
It is when the knowledge that makes an AI agent useful to your business, such as your preferences, rules and history, exists only inside one vendor's product, so leaving means starting over. It goes beyond the older lock-in of moving records from one system to another.
Does owning your AI mean running your own servers?
No. For most trade companies, owning means control and portability: company accounts, written rules, exportable records and your own tests. Where the software runs matters less than whether you can leave with what makes it work.
What is supported, what is uncertain, what to test next
Supported: vendors have announced AI agents inside field-service software, and booking is moving into AI search (by the vendors' own descriptions); in one 2025 study, external partnerships reached deployment more often than internal builds; the price of a fixed level of AI capability fell sharply through 2024 on the benchmark Stanford tracked; and the same rented model can change its behavior within months.
Uncertain: whether the hybrid I am leaning toward works better than renting everything. The MIT study had too little data on hybrids to say, and I found no study in home services. It is also uncertain whether owning the memory adds measurable value for a buyer.
What to test next: the exit question. Pick one agent you rely on and work out what it would take to replace it within a month. If the answer scares you, you have found the part you should own.
Rent the intelligence. Own the memory. And keep checking which is which, because the line will keep moving.