What We’ve Learned Building AI for Service Businesses

Five patterns hold across all three of our products. I wanted to write them down because they’ve completely reshaped how I think about building software for service businesses, and I think most of the AI industry is still chasing the wrong end of the problem.

We build three things at Velzyx. Aria Dental AI handles front-office voice and intake for dental and service practices. AgentCentric is the website and lead system for luxury real estate. AnalytixCRE underwrites commercial real estate deals for analyst teams. On the surface these have nothing in common. Underneath, they keep teaching me the same lessons. The lessons are not glamorous. They are not the things I would have said mattered most when I started. But they are the things that have actually moved the needle, and I’d rather share them than pretend the work has been about something more impressive.

1. Integration depth beats model sophistication

If you’d asked me in early 2024 what would determine whether our products worked, I would have said the quality of the underlying model. I would have been wrong by a wide margin.

The model is, in 2026, mostly a commodity. The capability gap at the top of the field has narrowed considerably, and what actually matters in production lives one layer up. What is not a commodity is the wiring. When Aria takes a call at a dental practice, the value isn’t the agent’s ability to hold a conversation. It’s the fact that the conversation ends with an appointment written into the right column of the right operatory in the practice management system, with the correct duration, with the new patient’s insurance verified, with the chart note in the right format. The conversation is the easy part. The integration is the entire job.

I notice this most when we onboard a new practice. The model setup takes a day. The integration with the specific practice management system, with its specific quirks and version, takes a week or two. And the integration is where the value lives. When something breaks in the field, it’s almost never the model. It’s the integration drifting because the practice updated their software, or an edge case in the schedule that wasn’t in the mapping, or a field that the system requires but we weren’t writing.

Same thing on AnalytixCRE. The model can read a lease. What makes the product useful is that the lease gets parsed, normalized, written into an underwriting model that the analyst can edit, with assumptions flagged where they came from the document and where they came from defaults. That’s engineering, not intelligence.

The implication for how we hire and how we spend time is heavy. We invest more in integration engineering than in model work, by a significant ratio. The companies in our category that get the ratio backwards tend to produce impressive demos and disappointing deployments.

2. The boring 80% matters more than the impressive 20%

Every one of our products has a shiny 20% that’s easy to demo and a boring 80% that’s where the customer actually gets value.

On Aria, the shiny part is the natural-sounding voice. The boring part is the routing logic that decides whether to take the message, book the appointment, transfer to a human, or escalate to the practice owner’s cell. The boring part is the call summary that has to be in the right format in the right inbox within ninety seconds. The boring part is the do-not-call list management, the after-hours handling, the snow-day handling, the voicemail-only handling for the days the front desk wants to triage themselves.

On AgentCentric, the shiny part is the design of the site. The boring part is the lead routing rules, the auto-response timing, the way leads tagged as serious-buyer get a different cadence than leads tagged as window-shoppers, the analytics that show the agent which listings are pulling traffic and which are not.

On AnalytixCRE, the shiny part is the underwriting model that comes out at the end. The boring part is the document handling, the rent roll normalization, the way we surface the assumptions we made versus the ones the user needs to confirm, the audit trail that lets a senior reviewer trace why we projected a certain expense ratio.

The shiny 20% gets you the meeting. The boring 80% gets you renewed.

I’ve started thinking of the shiny part as the entry fee and the boring part as the product. When we’ve let the team get pulled toward the shiny work, we’ve underdelivered. When we’ve put our best engineers on the boring infrastructure, the customers have stayed.

3. Trust is built through transparency about failures

This is the one I had to learn the hard way, and I wrote about a piece of it in an earlier post. The version that holds across all three products is simple: customers do not trust a product that claims to do everything. They trust a product that tells them, clearly, what it does well and what it doesn’t.

The first version of our Aria sales conversation in 2024 was a pretty standard pitch. The agent handles your calls, frees up your front desk, captures more new patients. We landed customers. The customers churned faster than we wanted, because they hit failure modes we hadn’t set expectations around, and they felt misled.

We rewrote the entire pitch. Now we start by saying what the agent doesn’t do. We don’t handle emergency calls; we escalate them in under five seconds. We don’t replace your team for complex clinical questions. We don’t handle billing disputes; we route them to whoever you designate. We don’t cover the calls where the patient is clearly upset; we hand those off immediately. The pitch is now mostly about the limits, and only secondarily about the capabilities. Conversion rate dropped. Retention rate rose by a much larger amount. Customers who knew the limits going in were dramatically more forgiving when the limits showed up in practice.

Same lesson on AnalytixCRE. We tell analysts, in the first conversation, that our underwriting is not a replacement for their judgment. We tell them where we make assumptions and where they need to confirm. We surface uncertainty in the model output, not just confidence. The analysts who use it heavily are the ones who trust that the tool will tell them when it’s guessing. The ones who tried products that hid their uncertainty stopped using them within weeks.

Transparency about failures is not just good ethics. It is the only durable sales motion in a category where every product will eventually fail in front of every customer.

4. Vertical-specific vocabulary matters more than I expected

I underrated this. The first version of Aria spoke generic, polite English. It worked. It also felt foreign in dental offices, and dental practice owners noticed before patients did. The agent would say “appointment” when the practice says “visit.” It would say “cleaning” when the practice says “prophy” or “hygiene.” It would say “sorry, can you repeat that” when a real front desk would say “I didn’t catch that, one more time.” The cumulative effect made the agent feel like an outsider, even when it was doing the work correctly.

The fix wasn’t a better model. The fix was a vocabulary layer, tuned per practice, that taught the agent the specific terms that practice uses for its operations. The same agent, with the right vocabulary, started to feel like part of the office instead of a vendor on the line.

AnalytixCRE has the same dynamic with much higher stakes. CRE analysts have a specific way of describing leases, rent rolls, expenses, and underwriting assumptions. An analyst will lose trust in a tool that uses the wrong terms within five minutes. A tool that uses the right terms gets the benefit of the doubt for much longer.

AgentCentric is the same: luxury real estate has its own vocabulary, and a site that uses generic real estate language reads, to a $20M buyer, as cheap. We learned to tune the copy by tier, by region, by agent, because that’s how the agents themselves speak when they’re doing their best work.

Vocabulary is the cheapest, most underused lever in vertical AI. It costs almost nothing to get right and almost everything to get wrong.

5. The human handoff is the product

This is the lesson I’d emphasize most if I could only emphasize one. After three years of building, I think the actual product we sell across all three lines is the handoff, not the AI.

On Aria, the handoff is what happens when the agent decides it cannot or should not handle this call. The transcript that lands with the front desk. The timing of the alert. The information already captured so the human picks up with context, not from zero. When we’ve gotten the handoff right, the front desk team uses Aria as a partner. When we’ve gotten it wrong, they treat it as a black box that creates work.

On AgentCentric, the handoff is the moment a lead transitions from auto-response to live agent contact. The information passed to the agent. The timing of the alert. The lead score and the rationale. When we get the handoff right, the agent walks into the conversation with leverage. When we get it wrong, the agent looks unprepared in front of a buyer who is paying close attention.

On AnalytixCRE, the handoff is the moment the underwriting model gets to the senior analyst for review. The flags on the assumptions we made. The places we’re uncertain. The supporting documents linked. When the handoff is clean, the analyst saves real time. When it isn’t, the analyst redoes the work to be sure.

The framing I now use internally is that we are not building an AI that does the job. We are building a system that does the bounded portion of the job and hands the rest off, well, to the human who has to finish it. The hand-off is the product. The AI is the engine. The customer cares about the engine only insofar as the handoff is clean.

What this changes about operational AI

If those five lessons hold, the priorities for any serious operational AI vendor change in ways that are worth being explicit about.

Integration depth and workflow design matter more than raw model capability. The model layer matters, but it is not where the leverage is.

Per-operator onboarding has to be the rule, not an upsell. The per-practice or per-firm tuning is where the product becomes useful. Skipping it saves money in the short term and costs the customer in the medium term.

Sales material has to lead with limits. Not because it is noble, but because it is the only way customer expectations align with what the product will actually do in their environment.

Vocabulary tuning and per-vertical language work matter more than they look. No obvious technical glamour. Outsized practical impact.

The human handoff is the central design problem of any operational AI product, not an afterthought to be solved last.

None of this is novel in the abstract. Plenty of writers in the space have made versions of these points. What surprised me is how directly the lessons translate into resource allocation. Whoever spends the most time on the parts of AI that aren’t the AI tends to ship the better product. That is, increasingly, the operating principle I want this company to be known for.

Why I’m sharing this

Partly because I think the AI industry, in 2026, is still over-indexed on model capability and under-indexed on workflow depth. Partly because the operators I respect are the ones who write down what they’ve actually learned, not the ones who pitch the most polished version of what they wish were true. And partly because if you’re a service business owner reading this, I’d rather have you know how we think about building before you talk to us than after.

If any of this resonates, the easiest next step is a conversation. We’ll tell you the same things we tell every prospective customer: where we fit, where we don’t, and how we’d know within thirty days whether the product was earning its keep in your business. That’s the only sales pitch I’m willing to put my name on.