How an AI agent was added to the mobile CRM so users can search records and create leads, contacts, tickets, appointments and tasks in one conversation, in English or Hebrew
Mobile CRM users often need to search for a record or create a lead, contact, ticket, appointment or task while they are away from a desktop. Doing that through menus and forms takes time, and switching to a separate AI chat loses the CRM context.
The conversation stays on one session, so a follow-up such as a title or a time continues the same request. The reply language follows the user’s message, English or Hebrew. Dates, times and picklists follow the signed-in user.
Mobile AI Agents brings a configurable AI agent into the mobile CRM. An administrator defines what the agent is allowed to do. The mobile app sends the user’s message to the CRM. The agent searches or creates the record, then answers with what actually happened: found, not found, created or not created.
Field and mobile users need the CRM while they are speaking or typing a short request, not while they are filling a long form.
Typical requests include: Find a contact by organization, email, phone or picklist value, Find a ticket by ticket number, Create a lead or a contact from a conversation, Open a support ticket, Schedule an appointment, Create a task and add the title or time in a later message.
Doing this manually means opening the right module, searching the right field, and completing every mandatory value. Relative words such as “today” and “tomorrow”, a time such as “12:00 AM”, a Hebrew picklist label, or an organization name are easy for a person to say and easy for a plain text search to miss.
A separate AI chat does not know the CRM schema, the user’s date format, the user’s timezone, or which records that user is allowed to see.
The business requirement was to let the mobile user ask in a conversation, and to let the CRM perform the search or the create and report the result.
Mobile AI Agents was built as an administrator-configured agent for the mobile app.
Administrators open AI Settings and create an agent. They choose the provider and model, write the main instructions, and turn on only the actions the business wants. Each create action can be limited to selected CRM fields.
The mobile app calls runAiAgent with the user’s message. The CRM runs the enabled actions and returns one message plus the records that were found or created.
A follow-up uses the same ai_session_id, so the CRM keeps that conversation. This is separate from the mobile login session.
The objective was not to let the model change the CRM on its own without a result the user can see.
The objective was to turn a short spoken or typed request into a CRM search or a CRM record, and to tell the user what was found or created.
Search entity modules, except Users and Settings, using the database schema.
Create a lead, contact, ticket, appointment or task.
Carry out more than one action in the same turn and return one message.
Continue the same request when the user adds a title, a date or a time.
Answer in English when the user writes in English, and in Hebrew when the user writes in Hebrew.
Ask only when a value cannot be decided, such as a person or company name that was never mentioned.
| Capability | What it does |
|---|---|
| Search Records | Finds records across CRM modules by any stored field, including ticket number, email, picklist and related organization name. |
| Create Lead | Creates a lead from the conversation. |
| Create Contact | Creates a contact from the conversation. |
| Create Ticket | Creates a support ticket. |
| Create Appointment | Creates an event. “Today” and “tomorrow” become real dates. A missing end is 30 minutes after the start. |
| Create Task | Creates a task. A later title or time stays a task and is not turned into an appointment. |
| One Conversation | ai_session_id keeps history for that conversation. |
| Result Message | The message says found, not found, created or not created after the action finishes. |
An administrator creates the agent, selects OpenAI, Grok or Gemini, and enables the allowed actions and fields.
The app calls runAiAgent with the message. On the next turn it also sends ai_session_id.
The agent reads the message and the saved conversation. It decides which enabled action to run. “Tomorrow” and “today” are converted with the user’s timezone. Hebrew and English picklist labels are both accepted.
Search runs against the CRM schema. Creates run as the signed-in user. Appointment times are stored in the database timezone, so 12:00 AM stays 12:00 AM for that user.
The response includes the message, the state, ai_session_id, the action, the full actions list, and any search results or created records.
The user says:
Find a contact whose organization name is New.
The organization field stores the related account, not the name itself. The agent matches the account name and returns the contact.
A contact with a blank organization is not returned. The same request in Hebrew, such as a search for the organization name, uses the same field.
The message is the result, for example: “Found 1 in Contacts: …” or “No records found.”
The user says:
Create an appointment tomorrow at 10:00 AM. End time is 11:00 AM.
The agent does not ask the user to convert “tomorrow”. It uses the user’s date and timezone.
If the user does not give an end time, the end is 30 minutes after the start. If the user does not give a title, the agent supplies one.
12:00 AM is saved in the database timezone so the calendar still shows 12:00 AM.
The user then says:
Title is Call the customer and the time is 10:00 AM.
The user first says “Create a task”. The agent asks for the missing title or time and returns state: awaiting_input.
The user then says: Title is Call the customer and the time is 10:00 AM.
Because the same ai_session_id is sent, this stays a task. It is not changed into an appointment only because a time was given. The message then reports that the task was created, or that it was not created.
Meaning: The agent still needs a value it cannot decide.
What the App Does: Show the question and keep ai_session_id.
Meaning: The search or create finished successfully.
What the App Does: Show the result message and the records.
Meaning: The action could not be completed.
What the App Does: Show that it was not found or not created.
The response always includes: message — the result, not “I am searching” or “I am creating”, ai_session_id — this conversation, not the mobile login session, action — the first action, or empty while waiting, actions — every action from that turn, search_results — matching records, when the action was a search
In simple terms, the user talks to the mobile app, and the CRM remains the system that searches and creates the records.
Give every CRM mobile app user a personal AI assistant. Configure its instructions, conversation memory and enabled actions, and let users create, update and search CRM records, book appointments and manage tasks using simple voice or text requests.
Give every CRM mobile app user a personal AI assistant. Manage the assistant from one screen in CRM Settings, see its model, history and enabled actions at a glance, and switch it on or off in one click, with features such as:
Set up the mobile assistant without code. Write its instructions in plain English, choose which actions it may take, and pick the exact fields it may fill for each one, from new leads to ticket updates, with features such as:
Users tap the assistant, say what they need in plain words, and the AI does the CRM work for them. One sentence can create a contact, add a note and schedule a follow-up task. The results come back as cards the user can open straight away, with features such as:
The API key stays on the server. The mobile app does not receive the provider key. The same conversation behavior is used whichever provider is selected.
Navigating module menus and manually configuring field-by-field filters on mobile screens.
Finds a record from a sentence instead of a field-by-field filter.
Typing exact database strings and dealing with mismatched language picklist values.
Matches the picklist label the user says, and answers in that language.
Manual relational lookup where searching by organization name misses contacts with ID-only relations.
Searches an organization by its name, not by an empty id on the contact.
Calculating calendar dates, converting timezones, and entering mandatory start and end fields.
Accepts today, tomorrow and a spoken time, and keeps a task as a task.
Context is lost between screens or turns, forcing users to repeat details and start over.
One ai_session_id carries the conversation, so the app does not rebuild the history itself.
Silent saves or ambiguous errors leaving users uncertain if the record was actually created.
The user hears whether the record was found or created.
Users can search and create from one message instead of navigating several screens.
Search uses the real field, including related names, picklists and the user’s date and time.
The agent can ask for a missing name and then finish the same task or appointment.
An English question receives an English answer. A Hebrew question receives a Hebrew answer.
Only enabled actions run, and create actions use the fields the administrator selected.
state, action and actions tell the mobile app whether to ask another question or to show the record.
The agent runs only for a signed-in mobile user. Search results are limited to records that user may view. New records are assigned to that user.
Administrators control:
The API key is stored with the agent configuration and is not returned to the mobile app. ai_session_id belongs to the user who started that conversation.
The same pattern can grow without changing the way the app holds a conversation.
Update an existing record from a follow-up
More modules, using the same schema-based search
Voice: the spoken sentence is already a message, and ai_session_id keeps the next utterance on the same request
Usage history for administrators
Richer record summaries in the result message
If something necessary is still missing, the agent asks, the app sends the next message with the same ai_session_id, and the original request continues.
The agent does not silently invent a person or a company that was never mentioned. It asks.
It does decide the values a person would expect it to decide:
The mobile user still sees the result and the record. The administrator still decides which actions exist.