AI Sales agent

In partnership with

You can build an AI sales agent in 10 minutes.

But 99% of people will never attempt it.

I've built multiple AI agents for my businesses.

And I'm still shocked at how easy it is.

Here's the thing though...

It took me 3 months of trial and error to figure this out.

I’ve put this together so you don't have to.

If you can write a basic checklist, you can build an agent.

One that will genuinely save you (and your team!) 100s of hours.

Let me walk you through each step in detail 👇

Hiring in 8 countries shouldn't require 8 different processes

This guide from Deel breaks down how to build one global hiring system. You’ll learn about assessment frameworks that scale, how to do headcount planning across regions, and even intake processes that work everywhere. As HR pros know, hiring in one country is hard enough. So let this free global hiring guide give you the tools you need to avoid global hiring headaches.

Download the free guide today

1. PICK ONE BORING JOB

This is where most people go wrong.

They think "I want to use AI" instead of "I want to solve this specific problem."

Stop thinking "AI in general." Pick ONE painful workflow you repeat every single week.

Here are real examples that work brilliantly:

→ Qualifying leads: You get 50 inbound leads. An agent can score them High/Medium/Low based on your ICP criteria and tell you who to call first.

→ Summarising sales calls: Instead of listening to hour-long recordings, an agent can pull out the key pain points, objections, next steps, and buying signals.

→ Drafting proposals: Feed it your discovery notes and it spits out a first draft proposal tailored to that specific prospect.

→ Cleaning CRM data: It can spot duplicates, fill in missing fields from LinkedIn, and flag stale deals that need attention.

→ Pre-call research: Before every meeting, it pulls together a one-page brief on the company, recent news, and relevant talking points.

The key is to define success in one sentence:

"Given X, the agent should output Y so that Z happens."

For example: "Given a sales call transcript, the agent should output a structured summary so that I can update the CRM in 2 minutes instead of 20."

If you can't write that sentence, you haven't picked a specific enough job yet.

2. MAP THE STEPS LIKE A SOP

Now turn that job into 4 to 7 clear steps.

Think of it like writing instructions for a new employee.

Write it as: Input → Actions → Decision → Output

Let's use lead qualification as an example:

Step 1 (Input): Receive new lead with name, company, role, website
Step 2 (Action): Check website for company size, industry, funding status
Step 3 (Action): Compare against ICP criteria
Step 4 (Decision): Score as High, Medium, or Low fit
Step 5 (Output): Return the score plus one-sentence reasoning plus recommended first action

Now mark which steps are:

→ Pure rules: "If revenue > £1M, score higher" – this is easy for AI
→ Heavy reading/writing: Analysing a website or writing a summary – AI is great at this
→ Judgement calls: "Is this company a good cultural fit?" – might need human review

BUT! Remember this:

If your process is broken, AI will just automate the broken process faster. Fix the logic first, THEN automate it.

3. CHOOSE YOUR AGENT PLATFORM

You don't need to be a developer.

Pick one based on your skill level:

No/Low Code Options: 
→ OpenAI Agent Builder (easiest starting point)
→ Zapier (good if you already use it, and still pretty easy)
→ Make (more flexible than Zapier)
→ n8n ( most powerful, steeper learning curve)

Dev-Friendly Options: 
→ LangChain or LangGraph for Python
→ OpenAI Agents SDK
→ CrewAI (for multi-agent setups)

For the record, I’m not technical, so I do not go down this path…

You only need THREE things:

1. Access to a strong model (Claude, GPT-5.2, etc.)

2. Tool calling (the ability to connect to other apps)

3. Basic logs (so you can see what it's doing)

That's it. Don't overthink the infrastructure.

If you're new to this, start with OpenAI's Agent Builder or Zapier. You can have something working in literally 10 minutes.

4. DEFINE INPUTS, OUTPUTS, AND TOOLS

This is crucial. Most people skip this and end up with a vague chatbot instead of a reliable agent.

Treat your agent like an API, not a conversation.

Specify your inputs: What fields are required? Be precise.

For a lead qualifier, your inputs might be:
→ Lead name (text, required)
→ Company name (text, required)
→ Company website (URL, required)
→ Lead's role/title (text, required) → Source of the lead (text, optional)

Define your outputs: What exact format should it return? JSON fields or a fixed template that the rest of your workflow can use.

For example:

{
  "fit_score": "High",
  "reasoning": "Series B SaaS company in target vertical with 50+ employees",
  "first_action": "Send personalised video within 24 hours",
  "missing_info": "Budget timeline unclear"
}

Attach your tools:

→ Data tools: Search your CRM, query databases, look up LinkedIn
→ Action tools: Send an email, post to Slack, create a task in Asana
→ Orchestration tools: Schedulers, webhooks, queues

The more specific you are here, the more reliable your agent becomes.

5. GIVE THE AGENT A JOB DESCRIPTION

This is where the magic happens.

Write a clear system prompt that covers:

Role: "You are a senior sales development rep focused on qualifying inbound leads for a B2B SaaS company."

Boundaries: What it must NEVER do.
→ Never invent data that isn't provided
→ Never make up company information
→ Never send external communications without human approval

Style: How should it communicate?
→ Concise and structured
→ British English spelling
→ Always include reasoning for scores

Examples: Include 1-2 example conversations showing good inputs and outputs.

Here's a pro tip:

Use a "think-then-act" pattern. Tell the agent to reason through its logic BEFORE taking action. This catches errors before they happen.

Something like: "Before returning your assessment, first list the evidence for and against this lead being a good fit, then make your final determination."

6. ADD MEMORY AND CONTEXT

AI agents have no memory between runs unless you give them one.

There are three types of memory to consider:

Conversation state: Pass recent messages so it stays on topic. If you're building a multi-step workflow, the agent needs to know what happened in step 1 when it gets to step 3.

Task memory: Store key decisions or variables for the current run. "The prospect mentioned budget constraints in the discovery call" – this should carry forward to the proposal draft.

Knowledge memory: Connect a vector store or file search over your docs. This is how you give the agent access to your sales playbook, case studies, pricing guide, or competitor battle cards.

Ask yourself: "What does this agent need to remember for the next step to be smarter than the last one?"

If the answer is "nothing," you might not need memory. But for most sales workflows, context is everything.

7. ADD GUARDRAILS AND HUMAN CHECKS

This is how you make it trustworthy enough to actually use.

Mark high-risk actions that always need human approval: 
→ Sending emails to prospects
→ Changing data in your CRM
→ Spending money (booking ads, etc.)
→ Anything that can't be easily undone

Add simple rules like: 
→ Never invent login credentials or IDs
→ Ask for clarification when the brief is ambiguous
→ Flag uncertainty rather than guessing

Log everything.

Every tool call. Every decision. Every output.

This lets you audit behaviour, spot patterns, and continuously improve the agent. When something goes wrong (and it will), you want to know exactly what happened.

8. WRAP IT IN A SIMPLE INTERFACE

The best agent in the world is useless if nobody uses it.

Keep the interface boring and obvious:
→ One field for inputs
→ A clear "Run agent" button
→ A result panel

Options for deployment:
→ Internal chat interface
→ Button inside an existing app
→ Slack or Teams command
→ Simple web form using Streamlit, Gradio, or your React app

Don't over-engineer this. The fancier the interface, the less likely people will actually use it.

9. TEST ON 5 REAL TASKS

Before you declare victory, run your agent on 5 real tasks from your actual workflow.

Watch the trace: → Which tool did it call? → In what order? → Did it get stuck anywhere?

Score three things:

1. Correctness of the final result (did it get the answer right?)

2. Steps it took (did it take weird detours?)

3. Time saved vs you doing it manually (is this actually worth it?)

Create a simple table:

Tighten the prompt, tools, or rules wherever it failed.

Use your platform's built-in observability if available. Most of them have trace viewers that show you exactly what's happening under the hood.

Your takeaway…

Those that adopt AI correctly now will be too far ahead to catch later.

If you want to be one of them, you need to start automating tedious work.

(The stuff that you don't need a human for...)

And here's the thing about building sales systems:

Katie came from an ops background. Sales terrified her.

In her words: "It used to send shivers down my spine."

Then she joined the Snowballn' community and implemented the sales systems we teach

In her First 4 weeks: 64 leads + $50k revenue… then 4 months later: $508,000 revenue and then 9 months: $5.8m in booked revenue!!

That's the power of a repeatable system.

And they are available to you when you join Snowballn’

Become a Snowballn' member and get access to AI tool walkthroughs, sales training, biweekly calls, monthly masterclasses, and a community of people who are leading the way with what the new age of sales looks like:

Sign up here: Join Snowballn’

Power to you,

Mike