What is an internal AI agent on WhatsApp?
It is an agent that works for your team, not for your customers. Your employees are on WhatsApp all day already, so the agent is there with them: a salesperson asks from the car how many are in stock and what the customer pays, a manager asks which customers owe money, and someone in service checks what is open for a customer. The agent reads the answer from Priority through MCP for Priority, not from a document it was fed in advance.
We built this pattern for two organizations: the sales and service team of a technology company, about 20 employees, and the staff and managers of a retail chain. Everything this guide describes is something one of those agents does.
How it works
- An employee's messageText, a Hebrew voice note or a file
- Identity at the doorA number not on the verified employee list gets no information
- with their permission levelThe lead agentPasses each part to the agent that specializes in it
- A specialist agentSales, Priority queries, finance or service
- when something is to be recordedA preview and an approvalOne approval allows one operation
- through MCP for PriorityRecorded in Priority
- A read-backReported as recorded only with evidence in hand
How does the agent split the work?
One WhatsApp request can touch stock, price and the customer's credit. So the agent is built as a team: a lead agent takes the message, works out what is being asked, passes each part to the agent that specializes in it, and joins the answers into one.
| Agent | What it handles | Its limit |
|---|---|---|
| The lead agent | Takes the message, splits it and returns one answer | Never writes to Priority itself |
| Sales | Quotes, a customer's purchase history, customer follow-ups and sales performance | A salesperson gets the figures of their own customers |
| Priority queries | Customers, orders, items and stock | Read only, no financial tables |
| Finance | Invoices, payments, customer debts and P&L | Only for people allowed to see finance |
| Service | Service calls, technician schedules and dispatch | Within the permissions of whoever asks |
| Writing with approval | Every create and update in Priority | The only agent that writes, and only after approval |
When only one agent can write, everything written is easy to check, and it is easy to make sure every write was approved.
Can you send it a voice note or a file?
Yes. A Hebrew voice note is transcribed to text before the agent starts working, and the agent works on the transcript, not on the recording. When the transcript is unclear, it asks for the message again instead of guessing. In the technology company's agent the transcript is written to the log, so you can check later exactly what was asked. The agent can also answer with a voice note.
A file sent to the agent on WhatsApp, such as a PDF, Word, Excel or PowerPoint file, it reads and answers from: what it says, a summary, or the numbers in it. It opens only the file it was sent.
How does the agent know who is writing, and what they may see?
Identity is checked at the door, before the agent reads a single row from Priority. A number that is not on the verified employee list gets no information at all. A WhatsApp display name, membership in a group or a line like "I'm the manager" grants no permission.
- Each employee has their own permission level: what they see, and which actions they can ask for.
- Access to financial data is set per person: P&L, revenue and procurement costs are three separate permissions.
- A salesperson gets the sales figures of their own customers, and a request for another salesperson's figures is declined.
- When a request goes beyond the permission, the agent says which role it needs, and does not try to get around it.
What happens when you ask it to record something in Priority?
A preview
The agent prepares the record and shows exactly what is about to be recorded. For a quote, for example: the customer, the items, the quantities and the total.
Approval
Nothing is recorded until a person approves, and one approval allows one operation.
A manager's approval, when needed
A salesperson's quote can go to a manager for approval first. When the manager does not answer, the request moves on to the next approver, and the salesperson gets an update once there is a decision.
A read-back
After the write the agent reads the record from Priority again, and reports it as recorded only with evidence in hand. When the result is unknown, it checks before trying again, so it never creates a duplicate.
Toward Priority the agent works through MCP for Priority: writing is off by default, in a conversation every operation is shown for approval first, and before an update the server checks nobody changed the record since the preview. You can also start with reading only: that is how the retail chain's agent started.
How do you know the answer is right?
- Every number comes with its source, the time it was checked and the report period.
- Zero, missing information and a partial result are three different things, and the answer says which one it is.
- A total over a long period is never reported from a partial scan: the agent splits the period into months or weeks until every part is complete, and only then adds them up.
- A document number or a code the agent did not get from Priority never enters the answer: an automatic check compares every such number with what the tools returned, and sends the agent back to correct it.
A result of hundreds of rows arrives as a short sample with a link to download the full file, and you can ask for it as Excel or CSV.
And in a WhatsApp group?
The agent takes part only in groups the managers approved, and answers in them only when it is mentioned. In a group all participants share one conversation, and the agent attributes each request to the person who wrote it and answers within that person's permissions. Group membership grants no permission, and what was said in a private chat with the agent never reaches the group.
Can it manage tasks and reminders?
Yes. When something needs a follow-up, the agent records a task with an owner, a next step, a due date and a status, and reports the open tasks when asked. You can ask it in the chat for a reminder or a recurring task, such as a morning brief at a set hour, and pause or cancel it. It confirms a reminder only after it is actually scheduled, and never promises one that was not set.
Can you connect Outlook too?
Yes, as an add-on. Through the Outlook module of MCP for Priority the agent reads mail, searches the mailbox, opens attachments, and sees the calendar and contacts, with personal OAuth for each user. In what we built, each person's mailbox answers only that person, and only in a private chat. A request for someone else's mail, or from a group, is declined.
Anything that leaves the mailbox or changes it, such as sending, replying, forwarding or a meeting invite, is shown first with the recipients, the subject and the full text, and waits for explicit approval. A changed draft needs a new approval, and the agent says "sent" only with the server's result in hand. A request written inside an email is information, not an instruction to the agent.
How do you start?
Start with reading only, and with the questions your team asks every day: open orders, overdue debts, items below minimum stock and today's deliveries. Then add writing, one kind of operation at a time and with an explicit permission, and connect channels one by one.
The full checklist, what to check in any agent before you let it write to Priority, is in the guide "How to build an AI agent Priority can trust".
Common questions
- What is the difference between an internal AI agent and a chatbot?
- A chatbot answers from information fed to it in advance. An AI agent gets a goal, reads current data from systems such as Priority, takes actions and reports what it did, and a person approves before anything is recorded. An internal agent does this for your employees, within each one's permissions.
- Can the agent record in Priority without approval?
- No. In a conversation every write is shown for approval first, and after it is recorded the agent reads it back from Priority. In MCP for Priority writing is off by default, and you can start with reading only.
- Does every employee see all the data?
- No. Each employee is identified at the door and gets only what their permissions allow. Financial data is open only to the people set up for it, and a salesperson gets the sales figures of their own customers.
- Does it work with Hebrew voice notes?
- Yes. A voice note is transcribed before the agent starts working, and it can answer with a voice note too. When the transcript is unclear, it asks for the message again.
- What does it cost?
- The agent is built for each organization around its teams and permissions. It works with Priority through MCP for Priority, whose plans and prices are on its product page, and the Outlook module is an optional paid add-on.