You hire it the way you would hire a person.
There is no project and no setup call. You sign up, give Sanaf a job it already knows, put it in the Slack or Teams your team already uses, and connect the apps the job needs. It drafts before it sends, asks before anything it cannot undo, and writes down everything it did. You promote it as it earns it.
An illustration of how the conversation reads, not a transcript from a client account.
Sanaf starts free, in minutes, and it starts in shadow mode. You pick a job from the thirty five it knows, connect the apps for that job with sign-ins scoped to it, and add it to a channel. From then on it works the job: it drafts, you read, you promote it a level at a time per kind of action, and every action it takes lands in a work log your team reads in a thread. Anything consequential waits for a person until you decide it should not.
From sign-up to a promoted employee
- Step 1
Sign up, free, no card
You get $120 of credit that never expires. The console is where you manage the employee; your team never has to open it.
- Step 2
Give it its first job
Pick the job where the work piles up: the support inbox, lead follow-up, the invoices. It arrives with the job description written, and you can edit it in plain English.
- Step 3
Put it in your Slack or Teams
Add it to a channel with one click, through the same install review every app goes through. Your team mentions it like a colleague. If you have neither, it has its own chat in the console.
- Step 4
Connect the apps, one at a time
The inbox, the calendar, the books, the CRM: each is a sign-in you approve, scoped to the job it is for, held in a vault the model never sees, and revocable on its own.
- Step 5
Read its drafts in shadow mode
For the first stretch it writes what it would have sent and shows you. Nothing leaves. You correct it the way you would a new hire, and the corrections stick.
- Step 6
Promote it a level at a time
Per kind of action, you choose: do it, ask first, or never. Answering a question can be automatic while sending an invoice still waits for you. It runs the job on its schedule from then on.
The four things that make it a hire
A job description, not a prompt
Every job it holds is a written description you can read and edit: what it does, what it never does, and who it asks. It proposes changes to its own instructions with citations, and you approve them.
Rules per kind of action
Not one switch. Reading is one rule, replying is another, spending money is a third, and each can be stricter when nobody is around.
Everything written down
What it did, what it asked, what it used, and what is waiting on you, in a log you can read in the channel. Mistakes are found in the log, not in a customer’s reply.
Memory with receipts
It remembers how your business works, and every memory says where it came from. You can question it, correct it, and see the version history.
Consequential actions stop at a card, and wait for you
This is what a held action looks like in the channel. Sanaf says what it wants to do and why, and nothing happens until a linked manager answers.
An illustration of how the conversation reads, not a transcript from a client account.
Ask it the way you would ask a colleague
@Sanaf, what did you do this week and what is waiting on me?
From now on, ask me before replying to anyone who mentions a refund.
Take on the invoices job as well, starting in shadow mode.
Why did you answer the delivery question that way? Show me where you learned it.
What Sanaf comes with
Most AI employee products are one shared assistant your whole team talks to. Sanaf is a hire: a job description for every job, logins scoped to each, rules you set per action, a trial period, and a memory you can question. These are the mechanics, not promises.
A hire with a job description, not a chat box
Sanaf holds a written job description for every job you hand it: what it owns day to day, what it hands to a person, which apps it works in. Thirty five are written before you start, and the thirty sixth is yours to write. It is a colleague with a role, not an assistant waiting for a prompt.
Elsewhere: One assistant shared by the whole workspace, wearing every hat at once with no job to be measured against.
Logins scoped to the job
Every app connection is granted for a job and scoped to the narrowest permission that lets that work happen. Revoke a single connection without touching anything else, and see at a glance which job each one serves.
Elsewhere: Integrations connected once and shared across everything the assistant might ever do.
Approval rules per action, not one switch
For every kind of action, you choose: do it, ask me first, or never. Add conditions like an amount, after hours, or a new contact. Unattended work runs one notch stricter than work you asked for in the moment, and every decision carries a one-line reason. A safety check proves the rules hold before it goes near a customer.
Elsewhere: A single "ask before anything irreversible" setting, or an always-approve button per tool.
Starts in shadow mode and earns trust
Sanaf watches first and drafts what it would have done without sending anything. You review the drafts, then promote it a level at a time, with a report that says how often you agreed. Every promotion is your explicit call and is written down.
Elsewhere: Live on day one, learning by making mistakes in front of your customers.
Learns by proposing, never by rewriting itself
Overnight it reviews the day and proposes changes to its own instructions, each one citing the moments that prompted it. You accept or reject. Accepting creates a new version; rejecting is remembered so it stops asking. It cannot change itself without you.
Elsewhere: Notes the assistant writes to itself that nobody reviews.
Memory with receipts
Ask "why do you think that?" and it answers with when it learned the fact, where, and what it replaced. Correct a fact inline and the old one keeps its history. Nothing it acts on is untraceable.
Elsewhere: A memory you can reset but cannot inspect.
Everything it does is written down
Every turn is a replayable log: what it read, what it called, what it asked, what it sent. Your manager view shows it live while it works and keeps the timeline afterward, so "what did Sanaf do on Tuesday?" has an exact answer.
Elsewhere: A chat transcript and a credits total.
One bill, for the work
There is no seat fee and no per-job fee. You pay for the work Sanaf actually does, and the work log shows you what each piece of it used, so handing it another job costs nothing until that job produces something.
Elsewhere: A seat fee on top of credits whose price is never explained.
Working before the call you did not have to book
Sign up, give Sanaf its first job, and add it to your Slack. There is no demo gate, no waitlist and no onboarding queue between you and a colleague doing real work, and the $120 of credits you start with is enough to find out on your own jobs.
Elsewhere: A demo call, a trial that starts when sales says it does, and a setup checklist you complete yourself anyway.