Built so you can let it act
Anyone can connect an AI model to your systems. Whether you can trust it to act comes down to one question: are the limits enforced by code that runs every time, or written in a prompt that usually works? Ours are enforced.
It can only use the tools it was given
Every tool an employee uses passes through a single gate. A tool its skills do not declare is refused - even if the AI asks for it, and even if your workspace has that system connected. There is no side door, which is what turns “it may read the calendar but not send email” from a hope into a fact.
The record is written before the action
The entry describing what is about to happen is saved first, on its own. If anything fails midway, you are left with an entry saying so, ready to check - never with something that happened and was not written down. A log written only on success has its gaps in exactly the wrong places.
The AI never touches your systems directly
We do not let the model provider call your systems on the model's behalf. Tools run on our side, so the checks always stand in front of them. It costs us some convenience, and it is why the first two rules hold.
One workspace cannot see another
Separation is enforced by the database itself, not by remembering to add a filter. A query that does not say which workspace it belongs to returns nothing - so a missed check shows an empty screen, never another company's customers.
The level is calculated, and saved with its reason
What an employee may do is the stricter of your workspace policy and its own setting, within what the skill allows. Each factor is written onto the action, so the record shows the level and the reason in words - not a number someone has to decode.
Hired, not built
An employee arrives knowing its job. Answer a short interview about your business, and it is ready to work.
Ready-made employees
Six roles, each with the right skills already chosen and each skill set at a sensible, cautious level.
Hired through an interview
Answer questions about your business - what you sell, in which language, when you are open - and the setup is written for you.
A readiness checklist
It will not go live until it has what it needs, and the list tells you exactly what is missing.
A job description in plain words
Describe the job the way you would for a new hire. The AI's instructions are generated from it, and you can read those too.
Its voice, in each language
A tone, a dialect, and your own wording rules in Arabic and English - so it sounds like your business, not like a template.
Draft, publish, roll back
Changes are saved as a draft and reach customers only when you publish. Every version is kept, with who published it and what changed, and you can go back to any of them.
A team, not just one employee
Copy an employee that works well, give it another brand or branch, and hire as many as you need.
A role we do not have yet
Describe the job and the systems it needs; it becomes a numbered request you can follow.
Your own templates
Save an employee you have tuned as a template of your own, and hire the next one from it with the same skills, levels and know-how.
Limits you set
Four levels of autonomy, a budget, and a record of which level applied and why - enforced by the platform on every action.
Four levels, set per skill
Suggest only, draft for your approval, act and tell you, or act within a budget. You choose the level for every skill.
One place for your approvals
Everything that needs a person waits in one list, with what it will do, why it needs you, and what it will cost. Approve from your phone, or straight from WhatsApp.
A budget per skill, per day
How many times, and up to what amount. Your workspace sets the ceiling; an employee can only be held to a stricter one.
It steps back on its own
When it is unsure, or a rule applies, it drops a level. Over budget, it stops and waits for you - a budget that only slowed things down would be decoration.
One policy for the whole workspace
A ceiling by type of risk - reading, writing, money, anything that leaves the business - that no employee can be set above.
The level, and the reason for it
Set to L2, capped at L1 by your policy, held at L1 by a budget - whichever it was, it is written on the action in plain words.
Guardrails, in your own words
Never promise a delivery date, never quote a price that is not on the list, never discuss competitors - and what to say instead, in your words.
A moment to undo
Something it did on its own can be reversed within a short grace period.
Answers from your own information
Your documents, your website, your price list - with the source of every answer, and an honest “I don't know” where there is none.
Your documents, website and spreadsheets
Point it at your website and it reads the whole site - every page and the PDFs it links to - and reads it again every week. Upload files, or a manual a vendor hosts, and it cites the page they came from.
A map of your company, with a source for every fact
Your documents and website are read into facts - hours, policies, prices, places, people - each with the page it came from. Exact facts wait for your approval; your corrections hold. Every employee is introduced to the company from it.
The source of every answer
Each answer shows where it came from. When there is no answer, it says so and brings in your team - it never guesses from the nearest thing it found.
A price list it reads line by line
Products and services as a spreadsheet, so a question about one price gets one clear answer, not a paragraph.
Each employee reads its own section
Your refund policy is not what the receptionist needs. You decide which sections each employee may use.
What it could not answer
A report of the questions it had no answer for, so the next page you write is the one your customers needed.
Lists it searches and adds up
Products, prices, menus, listings, programmes and invoices, uploaded as a spreadsheet and found by their columns. Totals are counted by the database, never by the model's arithmetic.
Where your customers are
The same employee, within the same limits, on whichever channel a customer uses.
Your own business number, WhatsApp's 24-hour messaging rule respected, and voice notes answered with voice notes.
A chat window on your website
One line of code. It loads almost nothing until a visitor opens it, and it can never break the page it sits on.
Instagram and Messenger
Direct messages on both, answered by the same employee.
Telegram
A bot under your own name, answering with the same knowledge and the same limits.
A mailbox it works through - reading, summarising and drafting replies - with anything outgoing waiting for you.
A private preview link
Try an employee through a private link before anyone else can reach it. Shown once, and can be revoked.
Your own app
An API for each employee, so it can answer anywhere your developers can send a request.
Phone calls, incoming and outgoing
An employee that answers the phone and makes follow-up calls. It is built, and waiting on a telecom licence and a carrier in each country - we will not connect a number before that country's regulator allows it.
It remembers your customers
Who they are, what they bought, what they asked last time - and customers can see what is kept and ask for it to be forgotten.
One customer, across every channel
Someone who messaged you on WhatsApp and then on your website becomes one customer record, once there is good reason to believe it is the same person.
It knows who it is talking to
Name, language, past purchases, what they asked last time - ready when the conversation starts, not looked up mid-reply.
It checks before it shares
Anything kept about a customer is shared only after the person on the other end has been verified, to the level that detail requires.
See it, correct it, forget it
A customer's record can be shown, edited and deleted. Editing replaces a detail; forgetting erases it. They are different, and the difference is kept.
Conversations pick up where they left off
It remembers what was said earlier, and a visitor who refreshes your website finds the conversation just as they left it.
Work you can follow
A task has a status, an owner and a history - not a chat thread with a label on it.
Hand over a task
Give an employee a piece of work as you would give it to a colleague, and watch its progress.
A board of work in progress
Open, waiting, blocked, done - across every employee.
Work that repeats
The weekly report, the month-end follow-up - scheduled once and done without being asked.
Hand over to a person, or to another employee
With everything already said and collected, so the customer never has to repeat themselves.
The details it must collect
Name the fields - order number, branch, preferred time - and they arrive as a clean spreadsheet, checked and in the right format.
The steps a conversation follows
Greet, identify the customer, find the order, offer the options, confirm - as steps you write and reorder.
It writes first, within your rules
A reminder before an invoice or a renewal is due, an order confirmed before it ships, a follow-up it promised - only where there is a reason to write, in a template WhatsApp approved, never at night, and never again after someone says stop.
Bookings it makes itself
Books, moves and cancels within your hours and how many you can take at once, and offers a time that comes free to whoever is waiting for it.
Quotations priced line by line
Each line priced from your own list, the totals worked out by the system rather than the model, and one numbered quotation per conversation - amended, never duplicated.
Connected to the tools you use
Email, calendar, spreadsheets, your CRM, your store, your rental system and your accounts - with only the access each job needs.
Google Workspace
Gmail, Calendar, Drive and Sheets - connected once for your workspace, with only the access needed.
Microsoft 365
Outlook, Calendar, OneDrive and Excel, in the same way.
HubSpot
Knows who it is talking to, and keeps contacts, deals, tickets, tasks and notes up to date without anyone typing them in.
Your accounting system
Zoho Books, or the receivables list you keep: who owes what and since when, statements and reminders - and a payment recorded only once a person approves it.
Anything with an API
Give it an OpenAPI description or an MCP server, and your system's actions become tools you grant one at a time - and can try out before an employee uses them.
Access it cannot overstep
Each connection is limited to what the skills need, stored in a secure vault, and revocable from one place.
Your online store
Salla, connected through its own app: products found by what a customer asks for, a customer's own orders, a checkout link or a cash-on-delivery order, and each new order confirmed on WhatsApp.
Your rental system
Booqable: what is available for the dates asked, quotations priced from your own rates, and your own stock offered before anything you rent in.
See everything it does
Every conversation, every action, what it cost, and how well each employee is doing its job.
Every action, recorded before it runs
What it did, with which skill, at what level and why, and what it cost. The record comes first, so even a failure leaves an entry to check - never a silent side effect.
Every conversation, readable
Filter by channel, by employee, or by whether someone had to step in - and search all of it.
A scorecard, like any hire
How much it resolved, how often it escalated, what it cost - each compared with the previous period, in words rather than arrows.
Customers say how it went
When a conversation ends, the customer is asked once, on their own channel, to rate it from 1 to 5. The score and their words land on the conversation and the scorecard, and a low one tells the team.
A thumbs-up on WhatsApp counts
A customer's emoji on a reply - on WhatsApp, Telegram, Messenger or Instagram - is their rating of it, and lands where a thumbs-down goes: the queue, the scorecard, the team.
Told while the customer is still there
Each customer message is read for how they seem. An unhappy one flags the conversation and tells the team at once - by email or WhatsApp - with the reason in words. Nothing is taken from the customer automatically.
Costs, broken down
By employee, by skill, by week. An action without a known price is flagged, never counted as free.
Notified where you actually are
In the console, by email or on WhatsApp - or as one daily summary instead of a stream, if you prefer.
Tested before it goes live
Write down what a good conversation looks like, run it, and see exactly what the employee would have done.
A safe place to try it
Talk to the employee with its tools simulated: you see what it would have done, and nothing real happens.
Write down what good looks like
A scenario is what a customer says and what must be true afterwards. Run them before you publish, and see each check pass or fail with the conversation beside it.
Told when a change makes a test out of date
Edit the employee after a passing run, and a notice tells you the result no longer reflects what is live.
A thumbs-down becomes the next test
A real conversation that went wrong can be turned into a test, so it never goes wrong the same way twice.
Every role rehearsed before it reaches you
Before a role is offered, it is hired in a test business and run through its scenarios in Arabic and English. It is offered only once every scenario that must pass has passed.
Underneath it all
Your data stays yours, the record is complete, and the bill is explained line by line.
Your data cannot leak into another workspace
Separation is enforced by the database itself, not by remembering to add a filter - a query that does not name its workspace sees nothing at all.
You choose the region
Where your data is stored, and where the AI runs.
Take it all with you, or delete it
A full export of your workspace, and a deletion that really deletes.
Clear invoices in both languages
Priced in US dollars with no tax added, numbered invoices in Arabic and English - ready for your accountant.
A ceiling on your bill
Set a spending cap for the period. Reaching it pauses the work instead of surprising you.
Arabic from the start, not translated later
Right-to-left throughout, Arabic typography set on its own terms, and every screen checked in both directions before release.
One search box for the whole workspace
Ctrl K from any page finds employees, conversations, customers, leads, tasks and documents - a conversation by what was said in it, too.
Usable by keyboard and screen reader
Meets WCAG 2.1 AA, verified by an automated test on every screen rather than by opinion.
Check it against your own requirements
The security page answers what a technical review asks first. A demo answers the rest, on your own information.