Resources / Webinars / Replay
How to build an AI agent in monday.com: a live walkthrough
A support ticket comes in. Before a human ever sees it, an AI agent has already read it, matched it against a library of known issues, drafted a response, and marked it resolved. No queue. No wait time.
River Sol
Marketing Assistant · May 22, 2026
May 22, 2026 · 27 min · Free replay
At Workiflow's live session on AI agents, Head of Sales Ruan and the Workiflow team walked through exactly how to build a working agent from scratch, the prompting framework behind it, the setup steps inside monday.com, and two real agents already running in production. Here's what you need to know.
What is a monday.com AI agent?
A monday.com AI agent is a configurable assistant that takes real action, resolving tickets, drafting emails, updating boards, generating reports, without a developer, and without code.
This isn't a fixed if-this-then-that automation. An agent reasons. It reads unstructured input like a ticket description or a meeting transcript, makes a judgment call inside the guardrails you set, and decides what to do next.
/monday.com%20AI%20Agents_%20The%20Complete%20Guide%20for%20Business%20Teams%20(2026)%20image%201.png)
As Workiflow's CEO Danny Esquivel put it on the call:
This isn't AI replacing your team. It's AI plus your team.
— Danny Esquivel, CEO, Workiflow
The point is to take the rudimentary work off people's plates so they can spend their time on work that actually grows the business.
Why getting this right depends on how you prompt it
Workiflow starts with three prompting rules that decide whether the whole build goes smoothly or turns into a mess of trial and error.
1. Talk to it like a smart friend, not a search bar
A vague ask gets a vague result. Don't say "make me a website." Say what you actually want: a task tracker with a form to add tasks, a list to display them, and buttons to mark them complete or delete them.
2. Show, don't just tell
If you want a specific look or feel, point to an example. "Make a button that looks pretty" gets guesswork. "Make a button styled like Airbnb's" gets a result you can actually use.
3. Break big asks into small steps
Don't ask for "user authentication" in one shot. Ask for the login form. Then the validation. Then the connection to the authentication service. One piece at a time builds something stable. One giant ask builds something fragile.
The 4-part framework for defining any agent
Once the prompting mindset is right, every agent gets built around four questions:
What is this agent, specifically? Not "help with the pipeline." An SDR. A service rep. A project coordinator.
What's actually breaking? Where does the process stall, and what needs to happen to unstick it?
What skills does it need? Drafting campaigns, qualifying leads, triaging tickets.
What are absolutely off-limits? Brand voice, tone, actions that require a human sign-off before anything goes out the door.
Skip the guardrails step and you're the one explaining to a customer why an AI sent them something it shouldn't have.
Two agents already running in production
Before building one live, the session walked through two agents Workiflow already uses internally.
Jesse
Jesse pairs with monday.com's built-in AI notetaker. When a sales call ends, Jesse reads the transcript, pulls out action items, groups them by meeting, tags them with priority, and creates the tasks, with the source meeting and context attached to each one. Then it puts matching deadlines directly on the calendar. No one touches a keyboard.
Isabella
Isabella runs scheduled portfolio checks and turns them into finished reports, daily check-ins, weekly summaries, a health snapshot with what needs attention, who owns it, and what to do about it. What used to be a report someone built by hand is now a document that's just there when you need it.
Building an agent from scratch: the actual steps
Here's the exact process demonstrated live, building a support-ticket resolution agent from a blank slate.
Step 01
Create it and name it
In the AI dropdown, hit Start from Blank. Give it a name, a job title, an avatar. It sounds cosmetic. It isn't, your team treats it like a teammate with a defined job, not a black box.
Step 02
Add skills
Skills are pre-built guidelines, instructions, examples, output formats, for doing a specific task well. monday.com ships with certified skills for things like project risk insights, social post creation, and email building. Stack the ones relevant to the job.
Step 03
Connect tools
This is where an agent stops being a chatbot and starts taking action. Connect Gmail, Slack, the calendar; whatever the job requires. For each one, decide the decision boundary: should it always act automatically, or should it decide case by case?
Step 04
Grant knowledge and access
Upload the SOPs or reference docs it needs. Connect it to the boards it needs to read or edit, and be deliberate about which. A ticket-resolution agent might need full edit access to a knowledge base (so it can log new patterns) but only the access it actually needs on the live tickets board.
Step 05
Write the instructions
Type them out yourself, or just describe the job conversationally in chat. The agent will draft its own instructions and ask you what triggers it, a status change, a new item and what it should do in response.
Step 06
Set the triggers
Define what starts the work: an item created, a status changed, a schedule. You can adjust these anytime.
Step 07
Test before it touches anything real
Run a simulation first. A good test shows exactly what the agent would do, matches, drafts, resolutions, without making a single live change. Treat it like a new hire. It won't be perfect on the first pass. Run it through a few scenarios, review, adjust, then go live.
Step 08
Go live and watch the audit trail
Every action gets logged. What it read, what it decided, its confidence score, what it changed. Nothing happens in the dark.
A second build: cleaning up a messy CRM
The session also built a deduplication agent live, on request from a viewer. The build:
Generate the structure
The agent created a master contact list, a duplicate-pairs board, and an action log. From a single prompt.
Populate and connect
Boards linked together, a duplicate-finder skill added.
Scan safely first
The agent matched contacts on signals like shared email, phone, and company name. High-confidence pairs got flagged for auto-merge. Everything else went to a human for review. Nothing was deleted or merged automatically.
Humans make the call on merges
If a duplicate contact is already linked to other records, like an open deal, the agent doesn't guess. You decide: keep the newest record, merge and preserve history, or handle it case by case.
That sequence, scan safely, flag, human review, then act. Is the pattern worth copying for any agent that touches data you can't afford to get wrong.
What it costs to run
Scope each agent to one narrow job instead of one agent trying to do everything. After a month of runtime, you'll know exactly where its ceiling is, instead of guessing.
You can run agents from your phone
Beyond the monday.com interface, agents can be messaged and triggered directly through WhatsApp and Telegram. You don't need to open a browser to check in or give it a new instruction.
monday.com agents vs. Claude/ChatGPT + MCP: which should you use?
If you're already using Claude or ChatGPT with MCP connections, do you still need monday.com's native agents?
The platform that owns its own AI agents understands that platform best. monday.com's native agents are built with direct knowledge of its own infrastructure, they don't just plan the work, they execute it.
Claude and ChatGPT tend to be stronger for generalized, cross-platform work: scaffolding a new CRM structure, standing up a portfolio solution, getting a build 60% to 80% of the way there. But operating that build day-to-day, inside monday.com, is almost always faster natively unless you're using something like Claude Cowork to plug in externally.
MCP is what lets an AI agent talk to a platform's backend using plain language instead of code. It's an API for agents. It pairs well with native agents. It doesn't replace them.
Frequently asked questions about monday.com AI agents
A configurable, in-platform assistant that can read, reason over, and take action on your boards and connected tools, like Gmail and Slack, without requiring code.
No. Agents are configured and directed in plain language, by design, for people without a technical background.
Automation follows a fixed recipe. An agent reasons over unstructured input, a ticket, a transcript, and makes a judgment call inside the guardrails you set before choosing what to do.
Agents consume credits per action, priced roughly a cent apiece and sold in bundles. Cost scales with task complexity, so scoping each agent to a narrow job keeps usage predictable.
Use monday.com's native agents for day-to-day execution inside the platform. Use Claude or ChatGPT with MCP for cross-platform scaffolding or generalized builds that span multiple tools.
Yes. Agents can be messaged and triggered directly through WhatsApp and Telegram.
The real bottleneck isn't the technology
Building an AI agent in monday.com doesn't require a development team. It requires a clear description of the job, the problem, and the guardrails, and a willingness to test before you trust it with real data.
The teams moving first aren't the ones with the biggest engineering budget. They're the ones who scoped one real problem, built the agent to solve it, and let it start working while everyone else is still writing a requirements doc.
/ Start
Workiflow is a monday.com Platinum Partner and CRM Specialist that has served nearly 1,000 clients across SMB to Fortune 500. To learn how vibe coding can be applied to your operations, contact the Workiflow team.
30 minutes. No commitment.