Start with one narrow job. Choose a single repetitive task, such as flagging deals that have gone quiet, and define the trigger, the data the agent reads, the decision it makes and the action it takes. Then connect your CRM, an automation tool and an AI model, test the result against deals whose outcome you already know, and turn it on for the team once the output holds up.
You can do this without an engineering team. You do need someone who knows your CRM and your sales process to build it and to keep it working as things change.
What is an AI sales agent?
An AI sales agent is an automated workflow that watches for a condition, reads the relevant context, makes a judgment and takes an action without anyone asking it to.
A deal risk agent is a good example. When an open deal goes a set period without contact (two weeks, for example), the agent reads the deal history and the last call transcript, then sends the rep a Slack message that names the risk and suggests a next step.
A dashboard or a report waits for someone to open it. An agent does the checking on a schedule and brings the result to the person who needs it.
Which agent should you build first?
Pick the most painful manual task your team repeats every week and automate that one task, with one trigger, one action and one output. These work well as a first build:
| Agent | Trigger | What it does | What it replaces |
|---|---|---|---|
| Deal risk alert | An open deal has had no activity for a set period | Pulls the deal record, summarizes the last interaction and messages the rep with a suggested next step | A manager checking for stale deals by hand |
| Contact enrichment | A new contact is created in the CRM | Looks up job title, company size, tech stack and funding stage, then writes them to the CRM fields | Reps researching each new lead before they act on it |
| Pre-call brief | A calendar event with an external attendee is about to start | Sends the rep the deal snapshot, last call summary, open risks and attendee profiles in Slack | Call prep that gets skipped when the day is full |
Build one and confirm people use it before you build the next.
What are the parts of an agent?
Every agent has four parts:
- Trigger: what starts it. A CRM field change, a calendar event, a time interval or a webhook.
- Context: what the agent reads before it acts. CRM deal data, call transcripts, email threads, LinkedIn profiles.
- Decision: what the AI model evaluates. Is this deal at risk? Is this lead a good fit? Which question matters most on this call? This is the judgment that used to need a person.
- Action: what happens next. A Slack message is sent, a CRM field is updated or a task is created.
An automation tool connects the four parts to your CRM and your messaging tools. The AI model sits in the decision step.
Which tools connect the pieces?
Start with the tools you already have. If you need to choose an automation tool, these are the common options:
- n8n is open source and can be self-hosted, and it has built-in connections for HubSpot and Slack. Its visual builder shows what happens at each step, which helps when something breaks.
- Make has a drag-and-drop interface and prebuilt templates, so it’s an easier starting point for a team with no technical background. Custom logic in the decision step is harder to build there.
- Zapier works well for simple if/then automation. Branching logic and multi-step context reads get expensive and awkward, so it makes most sense when your team already runs it for other things.
For the decision step, Claude and ChatGPT can both do the job. I compare them in Claude vs ChatGPT for sales workflows.
For data, you connect through each tool’s API (HubSpot, your call recorder and so on). If you use Claude, MCP connectors can give it access to your tools directly, which removes a layer from the setup.
How do you test it before it touches live deals?
Test the agent on static examples before it runs on live data.
Build a test version that reads from a fixed CRM snapshot. Run it against a batch of deals whose outcome you already know and check whether it makes the calls you’d make. Check that the Slack message, CRM update or task is formatted correctly and reaches the right place.
The problem to look for is missing context. You asked the agent to assess deal risk and didn’t give it the contact’s seniority. You asked for a pre-call brief and left out the deal stage. Read each output and ask what a good rep would want to know that the agent didn’t have.
Once it passes on historical data, run it in shadow mode for a few days. It fires in production and logs its output without sending anything. Compare its alerts with what you’d have flagged yourself, adjust, and then turn it on.
What makes a first agent fail?
Three things to watch for:
The trigger is too broad. If the agent fires on every deal update, the team gets a stream of alerts and learns to ignore them. Make the condition specific. “No activity in two weeks, close date this quarter, still marked active” is a condition worth an alert. “Any deal that changed” will fire all day.
The output goes somewhere nobody looks. An alert in a busy group channel or a neglected inbox goes unread. Send it where the person already pays attention. For a rep, that’s usually a direct message.
The context is incomplete. With too little deal information, the model gives generic advice. A brief that says “review the customer’s challenges and prepare relevant questions” is useless. A brief that says the prospect raised budget approval timing on the last call, and that the CTO on the buying committee hasn’t joined a call yet, is useful. Go through every field you pass to the model and ask whether a human analyst would have enough to work with.
How do you know it’s working?
Look for a change in behavior. The count of alerts fired tells you very little.
- For a deal risk agent, ask reps whether they hear about stale deals sooner, and look at how quickly flagged deals get a response.
- For an enrichment agent, compare the fill rate on key CRM fields before and after.
- For pre-call briefs, ask reps whether they walk into calls with better context.
If nothing changed, the agent is running and nobody is using it. Change where the output goes and what it says.
What does an agent need after launch?
An agent depends on the CRM underneath it, and the CRM keeps changing. Fields get renamed, deal stages get added, reps join and leave, and the sales process shifts. Any of those can make an agent fire on the wrong deals or stop firing, usually without an error anyone sees. Someone needs to own it, review its output from time to time and update it when the process changes.
An agent is also only as good as the records it reads, so CRM data quality is worth fixing early. Once the first agent is in use, a useful next build is filling in MEDDIC fields from call transcripts.