CHOOSEBusiness workflows
TypeSafe AI Jev
Faddy@faddyia
Jev is an AI model that analyzes incoming customer communications like voicemails or texts and provides answers to predefined questions with associated confidence scores, enabling rapid decision-making for service businesses.
SignalService type: pick from your six.
A plumber in Mesa lost a $4,200 water heater job last month.
Not because his auto-responder failed.
Because it took six seconds to decide the text was an emergency, and by then the customer had called the next guy on the list.
Six seconds.
WHAT LAUNCHED THIS WEEK
TypeSafe AI came out of stealth on September 15 with a model called Jev.
It does not write text. At all.
You hand it a mess. A voicemail transcript, a text, a form fill. Along with a list of questions that have fixed possible answers.
It hands back every answer in one pass, each with a confidence number attached.
Is this an emergency? Yes, 0.94.
What service type? Water heater, 0.88.
Is this an existing customer? Yes, 0.97.
Does a human need to see this? No, 0.91.
70 to 500 milliseconds. All four answers, one round trip.
WHY A SERVICE BUSINESS SHOULD CARE
Every automation you have ever been sold has the same shape.
Something comes in. Something decides what it is. Something responds.
The middle step has always been the expensive one.
You were paying a text-generating model to answer a yes or no question, then waiting several seconds while it wrote a paragraph explaining itself to nobody.
THE ARITHMETIC
Three-van HVAC shop.
22 missed calls a week.
40 inbound texts a week.
15 web form fills a week.
That is 77 a week. Call it 330 a month.
5 questions per message.
1,100 input tokens per message, including the customer history you attach.
330 x 1,100 = 363,000 input tokens a month.
Jev is priced at $0.042 per million input tokens. Output costs nothing.
363,000 / 1,000,000 x $0.042 = $0.015
One and a half cents a month to triage every single thing that comes at your business.
RUN YOUR OWN VERSION
- Count your inbound items per month. Calls, texts, forms, emails, reviews.
- Decide how many questions you need answered about each one.
- Estimate input tokens per item. A text plus job history runs 800 to 1,500.
- Multiply items by tokens, divide by a million, multiply by 0.042.
- Ignore output cost. There isn't any.
If your number came out scary, you counted wrong.
THE BUILD
1. Write down every decision your front desk makes in a day. Phrase each one as a question with a fixed set of answers. (45 minutes)
2. Turn each into a typed question. Emergency: yes or no. Service type: pick from your six. Urgency: score 1 to 5. (30 minutes)
3. Wire your missed-call and inbound-text feed to send the message plus your questions in one call. (2 hours)
4. Set thresholds. Above 0.90 the system acts on its own. Below 0.90 it pings your phone instead. (20 minutes)
5. Run it in shadow mode for two weeks. Log every decision. Let it send nothing. (14 days, no effort)
6. Compare the log against what you would have done. Move the threshold until the disagreements are rare enough to live with. (2 hours)
Working system in a weekend. Trustworthy system in about three weeks.
WHAT IT COSTS
Jev decisions: pennies a month at this volume.
An LLM for the few messages that need a real written reply: $4 to $15 a month.
Phone system API access: $0 to $30 depending on your provider.
Your time: one weekend, plus the shadow period.
THE FAILURE NOBODY PLANS FOR
Read this part twice.
Jev cannot return a malformed answer. Define your categories as water heater, drain, HVAC, or other, and it cannot invent a fifth one. The schema will not allow it.
That is a different thing from being right.
It can hand you "drain" with a clean 0.91 attached when the correct answer was water heater. The schema protects you from garbage. It does nothing about wrong.
On TypeSafe's own four-workflow benchmark, accuracy lands around 68%. TypeSafe built that benchmark and has said the harness may favor them.
Which makes your threshold the only thing standing between a wrong answer and a customer.
So shadow mode is not optional and two weeks is not generous. Hand-label 200 of your own decisions, compare them to the log, and set your threshold off your numbers instead of somebody's launch post.
The shops that get burned in the next six months are the ones who read "cannot hallucinate" and stopped reading there.
Like this and I'll send you The Phone Rang While You Were Under a Sink. 60 pages, the full missed-call build, every message, every guardrail.
Not free.