ZENZAP AI agents in Israel, connected to monday.com and Salesforce

What ZENZAP is and where AI agents fit

ZENZAP is an enterprise messaging platform with AI agents. Teams use it like any messaging app, but on company-managed accounts instead of personal phone numbers: admins control groups and access, and AI agents and bots take part in conversations as members.

An agent that chats but cannot read or write the business system is little more than a friendlier FAQ page. The value comes when the agent is connected to where the work is recorded, such as monday.com or Salesforce, and follows the same rules a good manager would. That connection is what CloudX builds.

How it works

A message in a ZENZAP group or direct chat reaches the agent service CloudX builds. The model chooses one of the registered tools, and the tool calls monday.com (GraphQL API) or Salesforce (REST API) with that role's credentials. The answer, or an approval request, goes back to the same chat.

Why chat is the right interface for frontline teams

Guards, technicians, drivers, volunteers and salespeople in the field rarely open a back-office system, but they do answer messages. When the process meets them in the chat they already use, the data actually arrives, and it arrives while it is still fresh.

The trade-off is that conversation is messy. People write in fragments, mix languages, send voice notes and change their minds. A good agent turns that into structured data, asks when something is missing, and never fills gaps by guessing.

Agents in production and complete demos

Two ZENZAP solutions built by CloudX run in production. The first is an agent for a sales operations team that turns quotes sent in chat into deals, accounts and contacts in monday.com. The second is a competitor price monitoring flow for a retail company that sends alerts to a ZENZAP bot.

Alongside them we built complete demos on real processes: shift scheduling for security companies, an agent pack for field technicians, volunteer operations, and an agent that works against a live system through a permissioned API. Each demo shows the full flow an organization gets, from the first message to the record in the system. A video of the shift scheduling demo is available on the CloudX website.

What we build with ZENZAP

AI shift scheduling agent with rest-rule checks

In our shift scheduling demo for security and guarding companies, guards submit availability by text or voice message, in their own words. A worker agent records it, and a manager agent shows who has submitted and who still needs a reminder.

AI drafts the schedule, then a separate rules checker enforces the hard constraints: 8 hours of rest between shifts, 36 hours of weekly rest, no more than 3 nights in a row, required certifications and no overlaps. The rules are configuration, not code changes: rest hours, weekly rest, consecutive nights and certifications are set per organization and collective agreement.

The manager approves, the schedule is published to each guard with read receipts, and swap requests are handled in the chat. monday.com holds the approved schedule, and every assignment records whether it came from AI or a manual change, and why.

Field technician bot: jobs, arrival and SLA

The field service demo is a Hebrew-first agent pack for technicians. A technician asks "what do I have today", accepts a job, reports "on the way" and gets Waze or Google Maps links, reports arrival, which is checked against the SLA, pauses and resumes, and sends photos and voice notes. The agent also suggests the order of the route.

Behind the chat, 57 tools are defined in a single registry, so every action the agent can take is explicit. Each tool has a schema, a required role and a permission check, so adding a capability means adding one reviewed tool, not changing a prompt.

From a quote in chat to a CRM deal

For a sales operations team, an agent that runs in production sits in a ZENZAP group. Salespeople drop a quote in the group the way they already share it with colleagues, as text, a voice note or a PDF. The agent replies in the thread, confirms what it created or updated in monday.com, and asks one question at a time when a field is missing. Each deal owner also receives a weekly digest of their open deals in chat.

Agents that read and write your systems

In the "Agent Ready" demo, an agent reads live data from a CloudX-built system through a permissioned API and writes updates back. The API, not the agent, defines what is allowed. This is the pattern we recommend for any agent that touches real records.

Bots that bring business data into chat

For a retail company we built a competitor price monitoring flow that runs in production. Prices are collected on a schedule from defined competitor sources, matched to the company's products and compared with its own prices, and an alert goes to a ZENZAP bot when a price changes. The pricing team sees the change in the chat it already uses, not in an email report that gets lost.

Agents for volunteer operations

The volunteer operations demo has registration, a volunteer agent and a staff agent, with monday.com as the system of record. A volunteer registers, updates availability and asks questions from the chat, and staff approve sensitive actions before they run.

How we keep agents safe

An agent in a company chat will receive unexpected messages, including attempts to make it do things it should not. Our design rules:

• Rules in code, not in the prompt. Hard constraints such as rest hours or certifications are checked by a deterministic rules checker, not left to the model.
• Human approval for consequential actions. The manager approves the schedule; staff approve sensitive volunteer actions. We do not build agents with unrestricted autonomous actions.
• Role permissions. A worker agent and a manager agent see different data and can do different things.
• A defined tool list. The agent can only call tools that are explicitly registered, through a permissioned API.
• Prompt-injection protection. Messages are treated as data, and the agent's permissions are enforced server-side by the API, so even a message that manipulates the model cannot make it act beyond its role or tools.
• An audit trail. The business system records who did what and whether it came from AI or a person.

What the system behind the agent needs

Most agent failures are not model failures. They happen because the business system has no clear place for the data, no status the agent can rely on, or permissions that were never designed for an automated user.

So before building an agent, we look at the system it will write to. Is there one record per person, job or deal? Is there a field that tells whether a change came from AI or from a person, and why? Is there an API that exposes only what this role should see? When the answer is no, we fix that first, in monday.com or Salesforce, because we build those systems too.

Hebrew, voice and messy messages

The agents we build are Hebrew-first, including slang, abbreviations and typos, and also handle English. A voice message is transcribed and handled like any written message.

We define where messages, voice notes and transcripts are stored, which fields are sent to the model, and how long they are kept. The system of record stores only the structured result plus a reference to the conversation.

In testing we do not stop at clean messages: we also test partial messages, last-minute changes, requests that break the rules and deliberate attempts to make the agent bypass permissions.

Typical business problems this solves

• Shift availability collected by phone, WhatsApp and spreadsheet, with rest rules checked by memory.
• Field technicians who update the office by calling, so SLAs and evidence are tracked after the fact.
• Salespeople who send quotes but do not update the CRM.
• Volunteers or frontline staff who will never log in to a back-office system.
• Alerts that arrive by email and get lost, when the team actually lives in chat.

How a project runs

1. Process mapping. Which conversations happen today, who decides what, and where the data should end up. Process first, then AI.
2. System of record. Decide where the truth lives, usually monday.com or Salesforce, and what the agent may read and write there.
3. Rules and permissions. Write down the hard rules, the roles and the actions that need human approval.
4. Build the agent and its tools. A small set of explicit tools, a rules checker where needed, and tests that include hostile and messy messages.
5. Go live with a first group. Real users on one process, so rules and wording are refined against real conversations.
6. Expand and support. Roll out to more teams and processes, and stay on for support and changes to rules and tools.

The timeline depends on scope. One agent on one process is usually measured in weeks, while several agents across processes take longer.

Before go-live we agree what to track: adoption through the agent, questions it asked to complete data, and manager edits, so every expansion decision rests on real usage.

ZENZAP agents vs other approaches

We choose the channel by the process, and because we also build inside monday.com and Salesforce, you get the right one either way.

Choosing a partner for chat agents

• Which of your agents run in production today, and can we see a demo of the rest?
• Which rules are enforced in code, and which are left to the model?
• Which actions need human approval?
• How does the agent connect to our system, and who controls its permissions?
• How do you test against prompt injection and messy input?
• Can you also build the system the agent writes to?

Why CloudX

CloudX is an Israeli consultancy that works on Salesforce, monday.com and AI, and ZENZAP is one of our partners. We build both sides of the conversation: the agent in ZENZAP and the business system behind it. Because we build the monday.com and Salesforce side too, the agent gets a clean record to write to, a status it can rely on and an API scoped to its role.

Our approach is "Process First. Then AI." Our ZENZAP agents run in production in sales operations and retail, and our demos show the full flow for shift scheduling, field service and volunteer operations. Bring one process your frontline team handles by phone or WhatsApp today. We will show you the shift scheduling or field technician demo and map what your agent would do. Book a demo.