Never Miss Another Lead In Your Shared Inbox (Claude + monday.com)

How Do You Stop Missing Leads In A Shared Inbox?

Connect an AI agent to the inbox. It reads each email, works out whether it's a new lead, a customer case or neither, and logs the leads and cases on their own boards in monday.com.

A case is an issue or question from someone at a business you already sell to. The same approach works for other kinds of email, such as marketing requests going to a marketing board or finance emails going to a finance board.

What Columns Does A Leads Board Need?

A standard leads board does the job: source, company name, email, enquiry summary, phone, status, job title, website, industry and company size. One email won't fill every column, but it's enough to start working the lead and gather the rest during the sales process.

Getting this right matters because missed leads are expensive. The faster you reach a lead, the more likely they are to convert, and leads sent to a shared inbox are often missed or answered slowly.

What Should A Cases Board Capture?

The issue description, a priority, and the contact's company, email, job title and phone number. Ours also had website and industry, most likely left over from copying the leads board, and a cases board probably doesn't need them. Add whatever else your cases need.

Missed cases cost money too, because a customer who gets a slow answer, or none, may churn. When someone emails a person directly, that person usually logs the case because it's their inbox. Nobody owns a shared inbox, which is why emails sent there get left.

What Instructions Does An Email Triage Agent Need?

It needs to know what to do with each type of email. We asked Claude for instructions telling the agent to triage every email that reaches the shared inbox, then create a lead on the leads board with every detail it can capture, or a case on the cases board with as much detail as the email gives.

We also gave the reason: answering leads and cases faster, so fewer customers churn and more leads convert. We spoke the whole brief into a new Claude chat using dictate rather than typing it, and Claude came back with the instructions quickly.

They go into the Create agent page under Agents in the Claude Console, which you can log in to as an admin on your Claude account. It asks for a name, model, description and system prompt, and can write most of that for you from a summary.

Why Should A System Prompt Be Shorter Than The Instructions?

Because an agent with fewer instructions tends to work more efficiently and effectively, the same way a person gets a job done faster with a two-page handbook than a thousand-page one.

Check the instructions over before using them. Our monday.com account is connected to Claude through an MCP, so Claude could look up the real boards, their IDs and their columns, and write the instructions around them. It's worth setting up, and we show how in a separate video on the channel.

Then paste them into the summary function at the top of the Create agent page and press Generate. It filled in the name, description and system prompt, and the system prompt came out a lot shorter than what we pasted in.

Which Tools & Permissions Does An Inbox Agent Need?

Gmail, because that's where the shared inbox is, and monday.com, where the leads and cases go, with every action allowed except the ones it must never take. Here are the settings in the order the Create agent page asks for them:

1. Model: keep the effort level on low. Reading an email and creating an item is a simple job, so a smaller model such as Sonnet, or even Haiku, would likely be enough, and every step up in model or effort adds to the cost of each run.
2. Tools: press the plus button, add Gmail as an MCP server and press Connect. You'll be asked to create a vault to hold the sign-in, so do that, then add monday.com the same way and reuse the vault.
3. Permissions: each tool action can be set to Always allow, Always ask, Always deny or Auto. Always ask stalls the agent until someone approves it inside the session and Auto has failed for us, so we allow everything and use Always deny on anything it must never do, such as replying to emails.
4. Skills and advisor: no skill is needed. We'd suggest an advisor on a more capable model, such as Opus, that the agent asks when it can't tell whether an email is a lead or a case, the way a person would check with their line manager.
5. Press Create agent. You can edit it and save new versions as you refine the instructions, which is where a lot of the work in building any agent goes.

How Do You Deploy The Agent & Decide When It Runs?

Give it an environment to run in and a deployment that decides when it runs. Under Environments, press Create environment and host it on the Anthropic cloud unless you have your own servers. Unrestricted networking is fine for a low-risk job like sorting emails, and you'd limit it where more could go wrong.

On the agent's Deployments tab, press Create deployment, choose that environment and the same vault, and add an initial message such as "Please start work". You can cap what each session spends, say $10, but a session that hits the cap stops mid-job, so set it after testing, once you know the rough cost.

We triggered it manually for testing, as setting up the webhook isn't covered in the video. Once live, a webhook that fires when a new email arrives suits this job best. A schedule can end up re-reading emails it has already handled, so its instructions would need a way of logging them to avoid duplicates.

How Fast Can An AI Agent Log A New Lead?

In our test, 25 seconds, at a cost of $0.44. We sent a test lead to the demo inbox, pressed Run now and watched the session search the inbox, read the email and create the item.

The run summary showed one lead created, no cases, nothing ignored and no failures. Since the test came from our own address, the agent noted it could read as internal, but treated it as a genuine enquiry because the content and signature clearly came from a prospect, and took the contact details from the signature.

The test lead was on the leads board with the details mapped as well as the email allowed, ready for the sales team to work. Compare that with a lead that sits in a shared inbox for, say, three days before anyone replies.

Can An AI Agent Turn A Customer Email Into A Case?

Yes. The second test was an email from a customer saying people in their team could no longer see the sales pipeline board in the left-hand menu, asking whether permissions had changed and hoping for a fix before a 2pm pipeline review. The agent finished within 36 seconds, at a cost of $0.48.

The run summary showed one case created, no leads, nothing ignored and no failures. On the cases board the new case had the contact's name, email, job title and company, an issue description and a priority the agent set itself, so the team could start on it well before the 2pm meeting.

It flagged again that our own mailbox showed as the sender, which only happened because we sent the tests ourselves.

Does The Agent Ignore Cold Sales Emails?

Yes. To try to break it, we sent a cold pitch from someone who had seen us listed as a monday.com partner, offering to generate leads through LinkedIn content. The agent marked it as ignore, a cold LinkedIn lead generation pitch, and added nothing to monday.com.

It also recognised that the two earlier test emails had already been logged. That filtering is a big part of the job, because most of what reaches a shared inbox is noise, with the occasional lead or important email mixed in.

Why Trigger The Agent On Each New Email?

Because in testing it seemed to scan the whole inbox each time it ran, and as the inbox grows that gets more expensive, slower and less efficient. Running it each time a new email arrives, rather than having it search everything for what's new, is the change we'd make once the agent goes live.

Across the three tests it did what we wanted: the lead and the case were added automatically and the cold pitch stayed out of monday.com. Like any agent, it would still need refining over time.