Most automation projects that fail do it expensively. The scope covered half the business, the contract ran a year, and by the time anyone asked whether the thing was working, too much had been spent for the answer to matter. If other agencies have already burned your time with proposals like that, your distrust of the big project is not cynicism. It is pattern recognition, and this page agrees with it.
Three properties, none optional
A first project should be small, measurable, and reversible, and each of those words has a specific meaning.
Small means one workflow. Not the phones and the invoices and the review requests — one of them. A project that touches a single seam of the business can be understood by every person it affects, and when something goes wrong, there is exactly one place to look.
Measurable means a single number was agreed before launch, and the number comes from a source you can check without asking the vendor. If the vendor's dashboard is the only place the result exists, you have a slideshow, not a measurement.
Reversible means the whole thing can be shut off in an afternoon without breaking anything it touched. Your phone number stays yours. Your calendar stays your calendar. Nothing gets migrated that would have to be migrated back.
The anatomy of one, concretely
The cleanest first project we know of is missed-call capture on a single phone line, so it is worth walking through in full.
A programmable number sits in front of the line you already publish. Every call forwards to your real phone immediately, and callers notice nothing. But every call now ends in a recorded status: answered, no answer, busy, or hung up before connecting. When the status is anything but answered, a text goes to the caller within about a minute — your business name, an acknowledgment that the call was missed, a question about what they need. Their reply lands on a phone someone already checks, usually the owner's. And every one of these events is written, timestamped, into a simple lead record: the number, the time, what was said, what happened next.
That is the entire system. For an HVAC shop it earns its keep in the first hot week of June, when call volume doubles overnight and every tech is in an attic with both hands inside a unit. For a roofer it earns it the afternoon after hail, when four companies get called about the same shingles and the one whose text arrives in a minute is having the conversation while the other three are still on ladders.
Week one is paperwork, and a vendor should tell you so
Since early 2025, US carriers require business texting to be registered — a legal-identity filing and a use-case filing with an industry registry — before messages will reliably deliver. Approval takes days at best and a couple of weeks in busy stretches. A vendor who sells you on Friday and promises a Monday launch is either skipping registration or does not know it exists, and unregistered messages fail in the worst possible way: they report as sent and never arrive, so nobody notices for weeks.
A second constraint is worth knowing before you sign anything. A text replying to someone who just called you is transactional, and the law treats it kindly. A text promoting a spring special to that same list is marketing, and it requires written consent you do not have. A vendor who offers to "blast your list" from the same system is proposing a violation with your business name attached to it.
Measure the week before, not the week after
Before anything launches, pull a month of call logs from your carrier's portal and count the calls that ended without a connection — and note when they happened. This is the baseline-first rule, and skipping it makes the whole pilot unjudgeable. "The system captured eleven conversations" means nothing on its own. Measured against a baseline, the same sentence becomes checkable: eleven captured, against nineteen that went unconnected in the baseline month, most of them after five in the evening.
The baseline also protects you from being fooled in your own favor. A system that launches the same week as a hailstorm will look like a miracle. The storm did that, not the software, and only the before-number lets you tell the difference.
Kill criteria, in writing, before launch
Agree — in the engagement letter, not verbally — on the condition under which the project ends. A workable form has three parts: a number, a period, and who reads the data. For example: if after sixty days the log shows fewer than an agreed count of conversations that would otherwise have been missed, the system comes down and the engagement is over. The exact threshold matters less than the fact that both parties knew it going in. A vendor confident in the mechanism will accept this readily. A vendor who resists it has told you what they expect the data to show.
What to refuse in a first engagement
Some terms are reasonable in year two and wrong in month one. In a first engagement, refuse:
- An annual contract. Sixty to ninety days is enough to know.
- A tracking number held in the vendor's account. If that number ends up on your trucks and your Google listing, firing the vendor means losing it. The number belongs in an account you control.
- Replacing your CRM, your scheduling software, or anything your staff touches daily. A first project should add one thread to what already exists.
- Any metric only the vendor can pull.
- Retainer line items for strategy attached to a sixty-day pilot.
- Marketing texts to the captured list, for the consent reasons above.
None of these refusals is adversarial. They are how the project stays small, measurable, and reversible, and a vendor structured to deliver those three things will not mind any of them.
What a pilot will not tell you
A sixty-day test on one phone line proves the phone-line mechanism and nothing else. It says nothing about whether a system can answer your warranty questions or chase your invoices; those are different builds with different failure modes. It can mislead on season — a roofing pilot run in November proves little about June, and the reverse. It does not solve the true emergency call: someone with water coming through a ceiling needs an immediate call path, not a text, and the message has to offer one. And it will not fix a shop that answers every call but never sends the estimate. That is a different seam, and pretending one system covers both is how the big proposals get written.
Where to start
Pull last month's call log from your carrier portal — most show it under call history — and count the calls that ended without a connection. If the count is near zero, you do not need this, and nobody should sell it to you. If it is not near zero, that count is your baseline, and it is worth an hour's conversation.