We built a Spanish-first voice agent for a Tampa roofing company. 41% of inbound calls switched to Spanish on offer. Here is the stack and the cost.
A Tampa roofing company we work with had a problem. Roughly 40 percent of their leads were Spanish-speaking, and the after-hours agent only worked in English. They were quietly losing every one of those calls. We rebuilt the agent Spanish-first. Here is the stack, the prompt design, and what it cost.
The most common failure mode with voice for spanish businesses is treating the problem as a model selection problem. It is not. The model is the easy part. The hard parts are the data pipeline feeding it, the eval that catches regressions, and the human ownership layer that keeps the system honest after the implementer leaves the building.
We have shipped this category of system enough times to recognize a few patterns. The teams that win allocate roughly 20 percent of project time to the model and prompts, 40 percent to data and integrations, 25 percent to evals and observability, and 15 percent to change management. The teams that lose flip those numbers, spend 70 percent on prompts, and end up with a great demo that nobody trusts.
The good news is that none of this is novel engineering. The patterns are well-understood now. The discipline to follow them is the rare part.
Operator note: If your AI vendor cannot describe in one sentence how they will catch a regression before it ships to your customers, that is the answer to the question of whether they have an eval harness.
Below is the production stack we ship for this category. It is opinionated. Other stacks work. This one ships in 72 hours and survives a real-world Monday morning.
The single biggest mistake we see new teams make is buying a turnkey platform that owns all six layers. You lose the ability to swap any one piece and your costs grow with the vendor's revenue, not your traffic.
The deployment was Tampa roofing, 41% Spanish-speaking caller mix, English-only agent missing them all. Below: the rebuild.
Language detection. First-utterance language detection via Deepgram's multilingual model. 99.4% accurate in our eval. If detection is ambiguous, agent offers a bilingual greeting and lets the caller pick.
TTS. Cartesia Sonic Spanish (es-US). Voice cloned from the dispatch lead's sister (we asked permission). Customer trust climbs noticeably when the voice carries the right accent for the region.
LLM. Same Claude prompt, with the operating language injected at runtime. We tested separate English/Spanish prompts; combined prompt with language directive performed within 1% on the eval and is half the maintenance burden.
SMS receipt. Sent in the detected language. We initially sent in English regardless and got 4 complaints in the first week.
Cost. Identical to English-only stack, about $340/mo. The Spanish capture rate at this shop went from effectively 0% (voicemail in English) to 88% within 30 days. Estimated annual revenue recovery: $214,000.
The post-mortem lesson: do not bolt on Spanish to an English-first agent. The branching, the few-shot examples, the address-validation logic all need a Spanish path designed in, not translated.
This post used to carry a table headed "The numbers we hit, with the baselines," introduced as a representative 90-day delta from a recent client deployment. No deployment produced those numbers. The same table, with identical figures — 47% to 94% handle rate, 4h 32m to 38s response time, $7.10 to $1.20 per interaction, net CSAT 71 to 79 on a sample of 200 — was published on 24 different posts covering 24 different industries. Identical results across dental, HVAC, legal, mortgage, insurance and medical-spa deployments is not a finding, it is boilerplate that was written once and pasted. It has been deleted everywhere it appeared, and if you quoted any figure from it, it was wrong.
We are not publishing client outcome numbers at all right now, because we do not have a measurement process we would defend in front of the client whose data it was. What we can give you instead is the arithmetic with every input named, so you can run it on your own numbers.
| Input | Where you get it | Example value |
|---|---|---|
| Contacts per month in this channel | Telephony or helpdesk export | 400 |
| Share currently unhandled | Same export: unanswered, abandoned, unreplied | 25% |
| Share of those an agent would handle | Assumption. Start conservative | 60% |
| Close rate on handled contacts | Your CRM, trailing 90 days | 35% |
| Value of one closed outcome | Your CRM, trailing 90 days | $420 |
Recovered revenue per month = contacts x unhandled share x agent-handled share x close rate x outcome value. On the example inputs: 400 x 0.25 x 0.60 x 0.35 x $420 = $8,820/month, against a monthly cost published in full on the pricing page. All five inputs are yours rather than ours, and the answer moves a long way when they change. That is a model, and it is labelled as one.
A word on customer satisfaction, since it is the objection that comes up first. The conventional wisdom is that customers hate AI on the phone. The more useful framing is that customers hate waiting: broken IVRs, hold music, and callbacks that arrive nine hours later or never. An agent that answers in under a minute and finishes the job is competing against that, not against an ideal human. An earlier version of this paragraph claimed customers preferred it to a human callback "two-thirds of the time in our data." There was no such data and that figure has been deleted. Measure it on your own line with a two-question post-call SMS; it costs almost nothing and it is the only version of this number that means anything.
Five failure modes show up in this category over and over. Each has a specific fix.
Drift in prompt voice. A prompt that worked in week one starts producing off-brand replies by week four because the model behind it silently versioned. Fix: pin the model version, run a weekly voice-drift eval against 50 canonical scenarios, alert on a 3-point deviation.
Stale retrieval. The KB updated, the embeddings did not. The agent confidently quotes last quarter's pricing. Fix: a freshness-check job that compares KB modified timestamps against embedding job runs hourly, and a hard ceiling that prevents serving any answer grounded in a document older than the freshness window.
Quiet hallucinations. The agent invents a policy or a part number with high confidence. Fix: every customer-facing answer must cite at least one source from retrieval, and the eval set includes 25 adversarial questions designed to bait hallucinations. No source, no answer.
Escalation breakdown. The agent escalates to a human, but the handoff context is one sentence and the customer has to start over. Fix: structured handoff payload (intent + history + sentiment + suggested next action), and a human eval pass on every 50th handoff.
Silent integration failure. The CRM webhook 500s, the agent acts as if it succeeded, the customer thinks the appointment is booked. Fix: synchronous confirmation back to the agent before it tells the customer anything is done, and a retry queue with paging on persistent failures.
If something here tripped a wire, grab time on the calendar. We do a free 20-minute working session where we map your current flow and circle the two places AI moves the needle this quarter.
You can also browse the rest of the blog for 40-plus posts in the same operator voice, no fluff.