The message comes in.The job opens itself.

A customer texts, emails or leaves a voicemail. Stet Desk reads it, opens the customer and the job, starts your process, and drafts the reply under your rules. You fix what you'd fix, and you press send. Nothing else does.

Built for service businesses of 1 to 50 people. Fence, HVAC, plumbing, landscaping, cleaning, consulting. It does not answer calls, schedule, or book appointments; it reads, opens, drafts and remembers.

Your desk0.0 sreplaying a real afternoon
Voicemail · 2:14 PM
Hey, this is Tom Alvarez over on Cedar Creek Road, 2210. About sixty feet of old chain link along the back we want replaced with six-foot cedar privacy and a double gate for the mower. Before Thanksgiving if possible. 512 555 0142.
Customer
Tom Alvarez
Job
Back fence, cedar privacy
Site
2210 Cedar Creek Road
Wants
Before Thanksgiving
Site visitDraft the quoteReviewApprove & send
Draft · from your template · given 2 rules
Told twice. It's a rule now.
Always say plus tax after the price and name the start day. The next draft is given this.
Only the owner can press this. The database refuses anyone else.
ledgerread 5.1 sdraft 2.4 s$0.016
5.1s
from voicemail to an open job
$0.004
to draft a quote email from your template
0
messages the software can send
1
ledger row per model call, with the cost

The loop

Four steps, every job. The fourth one is where it learns.

A message arrives

Paste it, forward it, or let the inbox take it. Text, email or a voicemail transcript.

The job opens

Customer, job, site, scope, the date they asked for, and the questions you'd want answered before quoting. The original message stays on the job.

The reply drafts itself

From your approved template, given your standing rules for that step. It lists which rules it followed so you can check.

You fix it, it remembers

Edit the draft and say one sentence about what to learn. Said again on another job, it is a rule. You press send.

What it does

The parts of running a service business that happen at a desk, done at the desk, while you are not at one.

Inbox

One place a customer's message lands. The assistant reads it and opens the job. If it cannot find a name, an address or a scope, it asks you rather than guessing.

Your process, not a chatbot

Site visit, draft, review, approve, send. Each step has an owner: you or the assistant. The assistant only works at its own steps.

Drafts from your template

The quote email you approved is the skeleton. The draft fills it from the job and your rules and comes back for your review, never to the customer.

Told twice, it's a rule

Every edit you make is remembered. The same lesson on a second job becomes a standing rule. One click makes it a rule sooner.

Teach it from your price list

Paste your rates, your terms, the notes you give a new hire. The rules come out as candidates and you say yes or no to each.

On duty

One switch puts the assistant on a two-minute schedule inside the database. Drafts happen while you are on a roof. Switch it off and the key it used is revoked.

Where the rules live

Not in the prompt. In the database, where a model cannot talk its way around them.

RuleWhat it meansEnforced by
Only you sendA message leaves as sent only when the owner presses send. Any other sender, the assistant included, is refused.Trigger
Templates are approved per versionEditing an approved template makes a new version that is not approved. Nothing goes out under it until you approve it again.Constraint
A hard cap on model spendEvery call is reserved against your monthly cap before it runs. A refused reservation never reaches the model.Ledger
Your customers are sealedNames, emails, phones, addresses, messages and drafts are ciphertext at rest under a key held in the database vault, one per workspace.Vault
Everything is on the recordEvery step, draft, edit, lesson, approval, send and model call is a row in an insert-only audit log you can export.Audit log
One workspace cannot see anotherRow-level security on every table. The assistant acts as you, inside your workspace, and nowhere else.RLS

Your data

The four questions a lawyer tells you to ask any software vendor that added AI, answered flatly.

No.

Is my data used to train a model?

Nothing you or your customers write is used to train anything. The model vendor today is Anthropic, called per task, and that is a configuration row you can read.

You.

Who owns the outputs?

Every draft is yours the moment it lands. It sits in your workspace, sealed under your key, and leaves only when you send it.

Yes.

Do I get notice before AI touches my data?

The assistant only works at the steps of your process you gave it. Adding a step is your action, in your audit log, with your name on it.

Yes.

Can I audit what it did?

The ledger holds every model call: model, tokens, cost, latency, request id, which job, which rules it was given. The audit log holds everything else. Both export.

Compared to

The big platforms are building the same loop for 500-person contractors. Read their pages, then ours.

ServiceTrade AI

  • "Embedded. Explainable. Built for commercial service execution."
  • Agents that draft quotes and schedule, for commercial contractors.
  • No price on the page. No cost per action. No word on who presses send.

Adminify

  • "AI Employees" that answer calls and book appointments for franchise systems.
  • Seven channels, seven languages, one to two weeks of setup.
  • No price on the page. Nothing on where your rules live or what trains on your calls.

Stet Desk

  • The cost of every action is a number you can read: $0.004 for a quote draft today.
  • Your rules are enforced in the database. Only you can send. Written down above.
  • Nothing trains on your data. You can check, because every call is on the ledger.

Pricing

Model calls are passed through at cost and shown on your ledger. The plan pays for the desk.

Solo

One owner, one phone, one inbox.
[price] /month
  • Inbox, jobs, drafts, rules
  • On duty schedule
  • Model calls at cost, capped at [cap]/month
Start with Solo

Crew

An owner and the people who answer the phone.
[price] /month
  • Everything in Solo
  • Staff seats that draft and never send
  • Per-person spend caps
Start with Crew

Office

Several processes, several templates, someone who audits.
[price] /month
  • Everything in Crew
  • Your own processes and record types
  • Audit export and a signed data agreement
Talk about Office

Questions

Does it send anything to my customers?

No. It drafts. The database refuses any sender but the owner, by a trigger, not a setting. When you press send, your name is on the audit row.

What happens when it gets something wrong?

You fix the draft and tell it one sentence. That sentence is a lesson; said again on another job it becomes a rule the next draft is given. Wrong drafts get rarer in your own words, not ours.

Where do my rules come from?

Three places: what you fix in drafts, what you paste in from a price list or terms document and approve, and what you type as a rule. All three end up in the same manual, which you can read and export.

Which model does it use, and can that change?

Today Claude Sonnet drafts and reads messages and Claude Opus extracts rules from documents. The model for each task is a configuration row, so it can change without changing your rules or your data.

How much does it cost to run?

Model calls are billed at cost and every one is a ledger row with its price. Measured today: $0.012 at most to read a message, $0.004 to draft a quote email. Your workspace has a monthly cap that is enforced before a call runs.

Can my bill run away?

No. A hard monthly cap per workspace and per person is checked before any call runs, and a refused reservation never reaches the model. The cap is yours to set.

Is my customer data encrypted?

Yes, at rest, field by field, under a key that belongs to your workspace and lives in the database vault. The database does not hold a name, an email, a phone number, a message or a draft in the clear.

What if I want it out of my business?

Remove the starter kit with one button, or close the workspace. Your rules, your ledger and your audit log export first. Nothing of yours stays in a model.

Does it work with Claude or ChatGPT?

Stet Desk runs on Stet, your business's institutional memory, which Claude and ChatGPT can use as a remote MCP server with the same rules. The desk is the version with an inbox and a process.

Forward one message. Read the draft that comes back.

Twenty minutes, your real voicemail. We paste it in together and you watch the job open, the draft land, and the ledger row appear. Then you decide.