The Texting Tool We Built Doesn't Send the Texts

A client came to us with a simple ask. They wanted to text their customers more often. The job, as they saw it, was for us to build the thing that did the texting for them.

So we built a tool that won't send a single text on its own.

That wasn't us being difficult. It was the whole point. Let me walk you through why, because the reasoning applies to a lot more than text messages.

The setup

The client is an insurance agency whose customers are farmers and ranchers. It's a relationship business in the oldest sense of the phrase. Their advisors know clients by name, know what they grow, know which storm just rolled through which county. When bad weather is coming, a good advisor is exactly the person a farmer wants to hear from.

The agency wanted to stay in front of those clients more consistently, and texting made sense. People actually read texts. The problem was volume. One advisor might have hundreds of clients, and nobody is going to hand-type a hundred personal messages on a Tuesday afternoon.

So the ask landed on our desk: build us something that sends texts.

Where we almost built the wrong thing

Here's the fork in the road, and we came close to taking the wrong path.

The easy version of this tool is a mass blast. You load up a list, write one message, hit go, and a system fires it out to everyone from some number nobody recognizes. It's fast. It's cheap. And for this particular business, it would have quietly done real damage.

Picture it from the farmer's side. You've got an advisor you trust, someone you've talked to on the phone, maybe shaken hands with at a co-op. Then one day a text shows up from a random number that opens with "Dear valued client." You know in half a second it's a blast. The thing that made that relationship worth something just got a little cheaper.

So the real problem was never "how do we send more texts." It was "how do we send more texts without turning a personal relationship into marketing spam." Those are two very different problems, and only one of them is a software problem.

What we actually built

We flipped the design. Instead of a machine that sends, we built a machine that prepares and then steps out of the way.

Here's how it works in practice:

• First, the advisor filters the client list down to who the message is really for, by what they grow and the counties they farm in.
• Then the tool drafts each text with the personal details already filled in.
• Those messages line up one at a time inside the texting app already on the advisor's phone.
• He reads each one and taps send, from his own number, the same number his clients already have saved under his name.
• He can skip anyone who shouldn't get it, and the tool keeps track of where he is and moves to the next.

What used to be an afternoon of copy, paste, and fat-thumbed typos becomes a few minutes of tapping through messages that are already written and already personal.

The AI does the boring ninety percent. The human does the ten percent that actually carries the relationship. That is the design in one sentence.

We didn't build a tool that replaces the advisor. We built one that clears the busywork out of his way so he can do more of the part only he can do.

The messy middle (the part that isn't in the brochure)

Every honest build has a stretch in the middle where things get complicated. Here's ours.

The data was a mess, and we let it show. The first time we ran the preview on a real list, the tool flagged that about a third of the selected contacts were missing the detail needed to personalize their message. We could have hidden that and let the texts go out with a blank space where a client's county should be. We didn't. Surfacing the gaps was far more useful than pretending the data was clean, because the real first project here was never the fun AI part. It was cleaning up years of half-filled contact records. That is almost always true. The exciting build usually sits on top of a boring data cleanup nobody wants to pay for.

The off-the-shelf AI gave generic answers, so we stopped trusting it blindly. We'd also set up a resource tool that pulled market information using a general AI assistant. It sounded smart and said almost nothing, just broad market takes you could get from any quick search. So we reworked it to lean on the agency's own in-house market expert first, to give recent guidance priority, and to clearly flag older information as older. We added a guardrail so the AI wouldn't hand out advice that contradicted the firm's actual strategy. An AI that confidently tells your clients the opposite of what your experts believe is worse than no AI at all. Knowing when the tool is likely to be wrong is a skill in itself, and we'd rather build that caution into the system than hope for it.

We shipped the honest version, not the finished-sounding one. The weather-alert emails are a good example. We started with a small, cleaned list instead of pointing the system at the entire client database and crossing our fingers. Some of the nice-to-haves, like sorting people into different alert types, aren't built yet, so for now everyone on the list gets everything. That's a deliberate call. A smaller thing that works beats a bigger thing that misfires in front of your clients.

What we'd tell any owner thinking about this

If you take one idea from all of this, make it the question we kept coming back to.

When you bring AI into your business, don't start with "what can we automate." Start with "where is my team doing something repetitive around a moment that actually matters, and how do I strip out the busywork without stripping out the human."

For this agency, the moment that mattered was a real person texting a real client from a number they recognize. The busywork was everything around it: the filtering, the typing, the personalizing, the keeping track of who got what. AI is genuinely great at the busywork. It's terrible at being the person your client trusts. A good build knows the difference and protects it.

This is really the line between automation and delegation. Automation runs start to finish with no human in the middle. Delegation lets the tool do the first pass and keeps a person on the final call. For a relationship business, the send button is almost always a delegation job, not an automation one. And for a few tasks, the right answer is still to not hand them to AI at all.

There's a quieter lesson in here too. The reason a blast would have backfired is the same reason this agency has a business worth protecting: being real is getting rarer, and more valuable, right as everyone else rushes to automate the humanity out of their outreach.

A soft pitch, since you read this far

If your team is sitting on a pile of outreach everyone knows they should be doing but never finds the time for, and you've held back because "automating" it felt like it would make you sound like a robot, that exact tension is the kind of thing we work on at Humanity AI. The goal is never to take the person out of the loop. It's to clear away the repetitive work so there's more room for the person to do what only they can.

One last thing worth saying plainly: the hardest part of a build like this is almost never the AI. It's the data. Years of contact records with missing fields will sink a personalization tool faster than anything else, and most of the value only shows up once AI is working from your real, specific information. The unglamorous cleanup is usually the true first project.