How I Built an AI Assistant in 3 Hours (And Why That Matters for Your Business)

A few months ago, one of my pilot clients — a small contracting firm — told me their estimators were losing deals because nobody was logging the details after site visits. Great meetings, strong rapport, solid quotes delivered on the spot — and then nothing. No contact in the CRM. No follow-up scheduled. No deal tracked. By Friday, the details from Tuesday's visit were gone.

I'd heard this problem before. Not from one client — from every client. Different industries, different sizes, same story: the gap between the conversation and the CRM was killing their pipeline.

So I sat down and built Dispatch. Three hours, start to finish. A working SMS assistant that takes a text message from a phone, extracts the contact details, the deal information, and the follow-up timing, and writes it all to HubSpot. Contact created. Deal opened. Task scheduled. Confirmation texted back.

Why three hours, not three months

This is the part that matters for your business, so let me be specific about what "three hours" means and doesn't mean.

It doesn't mean I wrote the AI from scratch. I'm not training models or building machine learning infrastructure. I'm using existing AI capabilities — large language models that are very good at understanding natural language and extracting structured data from it — and connecting them to the systems service businesses already use.

The three hours went to: designing the conversation flow (what happens when someone texts in, what questions the assistant asks if information is missing), connecting the SMS layer (Twilio — the same infrastructure every major business text system uses), building the extraction logic (telling the AI what to look for: names, emails, phone numbers, deal amounts, dates), and wiring it to HubSpot's API (creating contacts, opening deals, scheduling tasks).

Each of those pieces is well-understood technology. The insight isn't technical — it's operational. I knew what to build because I've spent twenty years watching service teams lose deals to this exact problem. The AI is the tool. The domain knowledge is the product.

What this means for you

The lesson isn't "AI is easy." It isn't, in general. The lesson is: the gap between your team's real workflow and the system that should be capturing their work is shorter than you think. For most service businesses, the problems that eat the most time are also the simplest to solve with AI — because they're structured, repetitive, and pattern-based.

Your dispatcher copies job details from an email to a scheduling system. Pattern-based. Automatable.

Your office manager reconciles invoices against completed jobs every week. Structured. Automatable.

Your account manager rebuilds the same pipeline report every Monday morning. Repetitive. Automatable.

None of these require a custom machine learning model. They require someone who understands your workflow well enough to build the right connector between where the work happens and where the data needs to go.

The "three-hour" test

Here's a useful way to think about whether AI can help your business: describe the problem in one sentence. If you can explain the inputs (what information exists), the outputs (what should happen with it), and the trigger (when it needs to happen), you're describing something that can probably be built in hours or days, not months.

"When an estimator finishes a site visit, their notes should become a CRM record." Inputs: the notes. Outputs: a contact, a deal, a task. Trigger: a text message. Three hours.

"When a job is completed, the invoice should be generated from the work order." Inputs: the work order. Outputs: the invoice. Trigger: job status change. A few days.

"When a lead calls and nobody answers, they should get a text back within 60 seconds." Inputs: the missed call. Outputs: a text message. Trigger: the missed call event. A day.

If you're sitting there thinking "yeah, but my problem is more complicated than that" — maybe. But probably not as much as you think. The complexity is usually in the integration (connecting to your specific systems) not in the AI (understanding what the data means). And integration is engineering work with a known timeline, not research with an unknown one.

What I'm actually selling

I want to be direct about this: I'm not selling you AI hype. I'm not telling you AI will transform your business overnight or replace your team. I'm telling you there are specific, concrete workflows in your business right now that are eating hours every week, and that an operator who understands your business can build AI systems to handle them — fast, and for less than you'd pay a part-time admin to do the same work manually.

Dispatch is one of those systems. It started as a three-hour build for one client's problem. It's now in private beta because the pattern turned out to be universal — every service business has the CRM gap.

Your business has its own version of the three-hour build. The question is whether someone has sat down and identified it yet.

That's the conversation I want to have. Tell me what your team's week looks like, and I'll tell you what I'd build first.