The Two System Test: when a service business needs custom AI, and the questions to ask an AI developer before you sign

Most small service businesses don't need custom AI. Some do. The difference usually comes down to one question, and once you know which side you're on, a short list of questions sorts the good developers from the ones selling a demo.

The Two System Test

Ask one question about the job you want help with: does it cross more than one system?

1. One system: switch it on. If your field software already chases estimates the way you want, turn that feature on and spend nothing extra.
2. Two systems: build for the gap. If the job starts in one system and ends in another, like a missed call in your phone log and an unpaid invoice in QuickBooks, that gap is where custom earns its keep.

You'd think custom AI means training your own model. For a service business it almost never does. Most of the work is connecting the software you already run and deciding what each agent may do alone.

Why the test comes first

Finding the right tool is the hard part for small firms. In the Federal Reserve Banks' 2026 survey of small employer firms, 54% of firms planning to use AI said finding tools that meet their business needs was an obstacle, and 37% named the time it takes to implement and train. The same report found only 7% of firms using AI had it fully integrated into how they work.

The Two System Test saves you from both mistakes: paying for custom work your software already does, and forcing a single-system tool onto a job it can't see.

Who this is for and what you need first

This is for an owner who has a quote for custom AI in hand, or is about to ask for one. Before any call, write down:

• The job, in one sentence ("text back missed callers whose last invoice is unpaid").
• Every system it touches: field software, books, phones, CRM, calendar.
• Who in your office does it today, and what they'd still need to approve.

If your sentence names only one system, check that system's own features first. Our guide to the best AI for service businesses walks through that. You're not late, either: the Census Bureau found under 20% of firms with four or fewer employees were using AI as of May 2026.

What a custom build actually rides on

A custom agent connects through each system's own access, not through a person's password. Two examples of what that looks like:

• ServiceTitan requires an admin with the permissions to generate an API application key and manage API application access, and each app gets the scopes chosen for it, with an integration environment available on request.
• HubSpot private apps get their own token and scopes, should be rotated every six months, and can break if the user who created them is removed.

A developer who can't explain how they'll connect to your systems in those terms hasn't done this before.

The questions to ask any AI developer before you sign

About access 1. Which of my systems will it read, and which will it write to? 2. How will it connect to each one: the software's own API with chosen scopes, or a shared login? 3. Whose account creates those connections, and what happens when that person leaves? 4. Will you test against a copy of my data first? (ServiceTitan offers an integration environment for this.)

About what it does alone 5. What will the agent do without a person approving it? 6. Which actions always need a yes from my office? (Money, prices, refunds and promises should be on that list.) 7. Where do I see what it did yesterday? 8. How do I pause it, and who can?

About the build 9. What's the first job you'd build, and why that one? 10. How long until the first one is running, and what does "running" mean? 11. What do you need from my team during the build, in hours? 12. What does it cost to run each month after launch, and what makes that number change?

About leaving 13. If I stop working with you, what do I keep: the code, the prompts, the connections, the logs? 14. Who holds the credentials, and how do I take them back? 15. Will anything stop working the day the contract ends?

About claims 16. Can I talk to someone you've built this for? (If they can't name anyone, ask why.) 17. Will you put any promised result in writing? (Be wary of anyone who promises a return.)

How to read the answers

Good answers are specific and a little boring. "We connect through ServiceTitan's API with read access to jobs and invoices, your admin creates the app, the agent drafts and your office approves, and you keep everything if you leave." That's a developer who has shipped.

Warning signs:

• "It'll handle everything." No agent should handle everything.
• "We'll just need your login." That's a shared password, not access.
• "We promise results." Nobody can promise what AI will do in your business.
• No clear answer on what you keep if you leave.

How we answer them

Our own answers, for comparison. We start with one or two workflows found in the audit, give each agent only the access its job needs, connect through official APIs first, and run on the cloud, private hardware or a mix. The first agents usually take weeks, and we scope the timeline after the audit rather than before it. You can read the details on how we work and AI implementation, and our trust page lists the defaults: approval for sensitive actions and nothing sent unless you allow it.

What to watch out for

• Custom for a one-system job. If your field software already does it, a custom build is paying twice.
• No approval step. A build where the agent sends customer messages from day one is a build to walk away from.
• Lock-in. If you can't take the connections and the logs with you, you don't own the system.

How to tell you chose well

Three months in, you should be able to say what the agent did last week, show the log, point to what a person approved, and explain how you'd turn it off. If any of those is fuzzy, ask the developer to fix it before you add a second job. And if you'd like a second opinion on a quote, your free AI Operations Audit is a good place to get one.