If you run a field service business
For you if: five to a hundred people, some in trucks. Equipment sales and service, or electrical, mechanical, similar trades. A field service system or a CRM, maybe both, probably not talking to each other. Someone in the office still retypes things between them. You still get pulled into quoting.
If that’s not you, the other paths will fit better.
The problem is not the AI
Section titled “The problem is not the AI”You’ve been pitched an AI receptionist, an AI quoting tool, an agent that answers customer questions. Some of those products work. Almost none will work in your business yet, and the reason has nothing to do with the technology.
Your systems don’t agree on basic facts.
Ask a simple question — how many times have we been to this address? The service system knows the jobs. The CRM knows the deal that started them. Accounting knows what was invoiced. The customer exists in all three under slightly different names, one of them a typo from 2019. There isn’t one answer.
Anything you bolt on top inherits that. An AI that quotes from bad history quotes badly. An agent answering from three contradictory records picks one confidently, and it’s a coin flip which. The failure is quiet, so you hear about it from a customer.
So the first move isn’t an AI feature. It’s making one question answerable.
Step 1 — Pick one question you can’t answer today
Section titled “Step 1 — Pick one question you can’t answer today”One question. Not a dashboard, not a data strategy.
Good ones sound like:
- Which service contracts are losing money?
- How long from first call to cash collected?
- Which customers have equipment coming due, and did anyone call them?
- What share of quotes become work, by whoever wrote them?
The test: someone wants the answer, and a decision is waiting on it. If nobody would act differently once they knew, pick another question.
This takes an afternoon. It’s the step people skip.
Step 2 — Make that one question answerable. Nothing else.
Section titled “Step 2 — Make that one question answerable. Nothing else.”Now the plumbing — only as far as the question needs.
Usually that means deciding which system owns the customer, reconciling the records that disagree, and keeping the two in step going forward. It’s tedious, it doesn’t demo, and skipping it is why the demo you saw won’t survive your data.
Do not fix everything while you’re in there. Choosing one question is what gives you a boundary. Businesses that try to clean all their data at once don’t finish. Businesses that clean enough to answer one question do — and the second question comes much faster, because the hard part, deciding who a customer is, is already done.
Step 3 — The AI part, which is smaller than you think
Section titled “Step 3 — The AI part, which is smaller than you think”With one question answerable, the useful applications are obvious and unexciting: summarize what happened on a job, draft the follow-up nobody has time for, flag the quote sitting eleven days, pull the history a tech needs before knocking on the door.
They have something in common. Each one reads what you already have and produces a small piece of text a person checks. That’s the shape of AI work that pays in a business like yours right now. Not autonomous, not customer-facing on day one, and it doesn’t need to be to be worth real money.
Buy this rather than build it. Your existing systems ship AI features now, and they’re cheapest to try because they already have your data.
What to ignore for now
Section titled “What to ignore for now”Being direct, because nobody selling to you will be.
Customer-facing voice agents. The technology is real. The risk is that the first person mishandled is a customer, and your reputation is the asset. Revisit when your internal uses are boring and reliable.
Anything needing a data warehouse before it does something useful. You may need one eventually. Not to answer your first question — and projects that start there tend to end there.
Custom-built AI tools, at your size. If it can be done inside systems you already pay for, do that. Build only what’s genuinely yours, and only after the bought version failed for a reason you can name.
Replacing anybody. Practical, not moral: the returns available to you are in office throughput and drive-time prep, not headcount. Lead with headcount and you’ll get resistance from the exact people whose adoption decides whether any of it works.
Where to learn the tools
Section titled “Where to learn the tools”We’re not going to re-explain what the vendors document well themselves, or maintain a copy that goes stale. Their material is free, current, and better than ours.
Start with whichever assistant your business already pays for.
What we learned doing this
Section titled “What we learned doing this”What we got wrong first, or learned expensively.
The process everyone describes isn’t the process that runs. An owner once walked us through a sales funnel he’d genuinely systematized — and he had. Weeks later we found the handoff underneath it: someone typing the same customer into two systems by hand, then looking up a tax rate on a government website and typing that in too. Miss a step and a salesperson arrives at a house with the wrong number on the estimate. Nobody lied to us. The manual step had been there so long it stopped being visible.
Go watch someone do the job. Don’t ask them to describe it.
Your software’s dropdown is not your business’s vocabulary. A pipeline report can show almost no deals won, forever, because it counts a status nobody uses while the stage your team actually marks is something else. Margin can be computed two ways in one system and disagree. Before trusting a number, make someone who knows the business say out loud what “won” means here. Then check the report agrees.
Never match customer records on email alone. Families share an address and often an email. Match on phone, then email, then address, then name, then company. In trades the address is frequently your strongest identity, because the work happens at a place.
Nothing counts until it’s switched on. We’ve built automation that tested clean and then sat idle for weeks — while the business went on answering new leads hours later by hand — because the go-ahead never got chased. We won’t change how your staff works without your say-so, which means that decision is yours, and it’s the step most likely to quietly not happen. Put a date on it when you start.
Your staff won’t believe it works until they watch it work. Told a new system is running, people carry on the old way and report that it didn’t work. They’re not being difficult; they’ve been promised things before. The fix isn’t a memo. Put one real record through while they watch, then point at two more they can go verify.
Name a champion, and don’t let it be you. You won’t be the daily user and shouldn’t be training everyone. Separately: find out who the project actually depends on day to day. It’s often not the person on the org chart, and if they leave, everything stops that week.
Agree the one number before anyone builds anything. The mistake we’re least proud of. It’s entirely possible to finish a technically sound project, data tested and dashboards correct, and be unable to tell the owner what it was worth. Pick the number in week one — hours in the office, days to collect, quotes answered same day — and write down what it is today.
Ask what happens if your vendor disappears. Where does the code live, whose account, can you get in without them, what have you signed. Anyone worth hiring answers specifically and doesn’t mind being asked. Ask us too.
Want to argue with something here, or have a situation this doesn’t cover? That’s useful to us — tell us. If you’d rather someone look at your setup and tell you what they see, that’s what we do for a living over at flowmatrixai.com.