Most AI sales agents fail for a boring reason: they are asked to do everything, so they do nothing well and nobody trusts what they produce.
The ones that work own one job, use real data, and stop for a human before anything reaches a prospect or the CRM.
This guide shows you how to build an AI sales agent that way, with a worked example we tested on 10 real target accounts.
30-Second Summary
- An AI sales agent is a model that reasons plus tools that act: it decides the next step and calls the tools to do it, without a rule for every case.
- Give it one job first. Account research, lead qualification and follow-up are good starters because each one has a clear output a human can check.
- A working agent needs four parts: a model, tools (here the Evaboot MCP server), a written job description and a human approval step.
- Our example agent takes 10 target companies and returns the senior sales leaders, verified emails and a why-now brief per company, then pushes to the CRM after your yes.
- In our test it previewed 78 people for free, exported only the people it picked (25), and graded 5 accounts strong, 3 medium and 1 weak.
You can build this with no code at all: Claude and an MCP connection do the work, and an n8n version runs the same steps from a form.
We built and ran the full agent on a real Sales Navigator seat and a real HubSpot account. Copy it below, or follow the same method with your own tools.
We will walk you through the process step by step.
Everything is in one public GitHub repo: the skill, the agent instructions, a small helper script and the n8n workflow. Paste the prompt into your AI Agent, or pick the n8n tab to run it from a form.
Let your AI agent set it up
Paste this prompt into Claude Code, Codex or Cursor. Your agent clones the public repo, follows its instructions and asks before every change.
Clone https://github.com/evaboot-labs/ai-account-research-agent, read AGENTS.md and follow the skill. Then research these accounts for me: <your 10 companies>. Show me the picks and the cost before you export anything.
In the repo
- The ai-account-research skill: resolve, preview, pick, export, brief, approve, push
- research.py: the people search, the picks and the facts, the same on every run
- TARGET.md: who to look for at each account
- The n8n workflow, with its setup guide
Open the repo on GitHub
Set it up in n8n
Paste this prompt into your AI agent and it walks you through the import, one step at a time. Or open the workflow JSON on GitHub and import it yourself.
Set up the AI account research agent in my n8n. Clone https://github.com/evaboot-labs/ai-account-research-agent and follow docs/n8n.md. Walk me through importing n8n/ai-account-research-agent.json, adding the Evaboot, Anthropic and Google Sheets credentials, and filling in Your settings, one step at a time.
In the repo
- The workflow JSON: a form in, an approval sheet and a CRM push out
- docs/n8n.md: the three credentials and the settings, step by step
- The same picks and signals as the script
We used Claude, Sales Navigator and HubSpot in our example, but with a few tweaks, the same method works with another model, data source or CRM.
In this guide
- What Is an AI Sales Agent?
- Which Job Should Your First AI Sales Agent Do?
- How to Build an AI Sales Agent: Step by Step
- How to Build the Same Agent in n8n
- AI Sales Agent Example: 10 Target Accounts
- Common Mistakes When Building an AI Sales Agent
What Is an AI Sales Agent?
An AI sales agent is software that uses a language model to do sales work: it reads the situation, decides the next step and calls tools to carry it out.
The tools are what make it an agent. Without them, a model can only write; with them, it can search, export, update and send.
Two questions come up first. How is it different from a chatbot, and how much should it do alone?
- How Is an AI Sales Agent Different From a Chatbot?
- Autonomous or Assistive?
1. How Is an AI Sales Agent Different From a Chatbot?
A chatbot answers what it is asked, often from a fixed script. It waits for a person and stops when the conversation ends.
An agent works toward an outcome. You give it a goal such as "find the buyers at these 10 companies", and it chooses and chains the tool calls to get there.
Chatbot vs AI sales agent
Chatbot: "Who is the VP of Sales at Pexip?" It answers from what it knows, which may be out of date.
AI sales agent: "Research Pexip." It searches Sales Navigator, picks the senior sales leaders, exports their verified emails and writes why the account might buy now.
2. Autonomous or Assistive?
An autonomous agent runs a whole workflow on its own. An assistive agent does the heavy lifting and hands the result to a person to approve.
Start assistive. An agent that spends credits, writes to your CRM or emails prospects should stop for a human yes until you trust its output.
Which Job Should Your First AI Sales Agent Do?
One job, with an output you can check. An agent asked to research, qualify, write and follow up at once is hard to test, and one weak step spoils the rest.
Rank the sales work that eats the most hours on your team, then pick the one with the clearest result.
These four jobs suit a first agent because a person can check each output in minutes:
| Job | What the agent returns | How you check it |
|---|---|---|
| Account research | The buyers at each target company, verified emails, a why-now brief | Read the brief against the account |
| Lead qualification | A fit score and a reason per lead | Spot-check the scores against your ICP |
| Follow-up | A drafted next email for leads who went quiet | Read the drafts before they send |
| CRM hygiene | Updated titles, companies and emails | Compare a sample with LinkedIn |
We have tested workflows for two of these already: AI fit scoring for an ICP lead list and no-code lead generation workflows.
This guide builds the first one, account research. It is the job reps put off most, and the one where invented facts do the most damage.
How to Build an AI Sales Agent: Step by Step
Our agent takes about 10 target companies and returns the 2 to 3 most senior sales leaders at each one, their verified emails and a why-now brief, then waits for your approval before the CRM.
It runs in seven steps. The model decides only where judgement is needed; the rest is the same on every run.
- Connect a Model to Your Sales Tools
- Write the Job Down as a Skill
- Resolve Each Company to Its LinkedIn Page
- Preview the Buyers for Free, Then Pick
- Export Only the People You Picked
- Write a Why-Now Brief From Real Data Only
- Approve, Then Push to Your CRM
1. Connect a Model to Your Sales Tools
An agent needs a model that reasons and tools that act. We used Claude for the reasoning and the Evaboot MCP server for the prospecting tools.
MCP (Model Context Protocol) is an open standard that lets an AI Agent call outside tools. You connect a server once, and the model picks the tool from your plain-English request.
In Claude, open Customize > Connectors, click Add, then Add custom connector:
Name it Evaboot, paste https://mcp.evaboot.com/mcp and sign in to your Evaboot account in the window that opens.
Claude Code, Codex and Cursor use the same URL with your API key as a Bearer header. The repo's .mcp.json does this for Claude Code, and this comparison covers the ChatGPT side.
2. Write the Job Down as a Skill
A skill is the agent's job description: one file that says what to do, in which order, and where to stop and ask.
Ours has two rules that matter more than the rest:
- Confirm before anything that writes or spends. Creating a list, exporting, and pushing to the CRM all wait for a yes, with the credit cost shown.
- Write briefs only from the export. No facts from the model's memory or the web, and "no timing trigger" when there is nothing to say.
Who to look for goes in a separate TARGET.md file: CRO, CSO, SVP or VP of Sales, Head of Sales and Sales Director, the 3 most senior per company.
Claude also lets you set write tools to "Needs approval", so the connector itself asks before it changes anything:
3. Resolve Each Company to Its LinkedIn Page
Company names are ambiguous. "Pexip" and "PEXIP LIMITED" are two LinkedIn pages, and a wrong match sends the agent to the wrong people.
The agent searches Sales Navigator for each name and keeps the exact company id. When two pages match, it shows you both and lets you pick.
Then it saves the resolved companies as one Sales Navigator account list:
An account list is also reusable: you can open it in Sales Navigator, add companies, or manage it by hand.
4. Preview the Buyers for Free, Then Pick
The agent runs one people search on the account list, filtered on sales leadership titles. Reading the results is free: nothing is exported yet.
Our 10 companies returned 78 people with a sales leadership title. Exporting all of them would have cost 78 credits for people we did not want.
So the agent picks first, with the same rules on every run:
- Most senior first: CRO and CSO, then SVP, VP, Head of Sales, Director.
- Global before regional: "VP of Sales" before "VP Sales, Asia".
- Not buyers: pre-sales, sales engineering, enablement and partnerships roles are dropped.
- Clean duplicates: the same person listed twice, or under a page name such as "Quentic (an AMCS company)", counts once.
That left 27 people at 9 companies. The 10th, a 27-person fintech, had no sales title at all; the skill then offers the founder or CEO instead.
5. Export Only the People You Picked
The agent saves the 27 picks as a lead list and exports only that list with Evaboot. The export returns verified emails plus the company and role data the brief needs.
It cost 1 credit per lead. The preview in step 4 is what kept the bill at 27 credits instead of 78.
Two surprises came back, and both are why you check before you push:
- One person had changed company. He was listed at BlueOptima in the search but works at Vanta now. The agent moves people like him out of the briefs.
- One title was out of date. A CRO now shows as "Co-Founder, Board Member", so the brief says the role moved on.
6. Write a Why-Now Brief From Real Data Only
A why-now brief says why this account might buy this quarter, in two sentences. Ours uses only fields from the export, so every claim can be traced to data.
These are the signals it reads, strongest first:
| Signal | Export field | Reads as |
|---|---|---|
| New sales leader | Has New Position, months in role under 6 | A new leader resets the plan and the tools |
| Fast growth | Headcount growth over 30% in 2 years | More people and more territory to cover |
| Team build-out | Two or more people under 12 months in role | The team is being rebuilt |
| Team shape | Sales Team Size, Business Development Size | A big prospecting team to keep fed |
| Shrinking headcount | Headcount down over 10% in 2 years | Same targets, fewer people |
Each account gets a grade: strong for a new sales leader or fast growth with a new hire, medium for one weaker signal, weak for none.
One trap: Evaboot splits time in role into years and months in two fields. Read only the months field and a CRO who joined 4 years and 7 months ago looks brand new.
The agent writes everything to a sheet for your review, one row per contact, strong accounts first:
Each strong account also gets a one-line opener for the most senior contact, built on the same signal and nothing else.
7. Approve, Then Push to Your CRM
Nothing reaches the CRM until you say yes in the sheet. Read the briefs, mark the rows, and tell the agent to push.
The push goes through Evaboot's CRM integration, so HubSpot, Pipedrive, Attio or any connected CRM works. Our 25 leads landed in HubSpot in one call:
The push sends the whole export, not row by row. If you reject a lead, delete it in the CRM afterwards, or skip the push and import the approved rows of the sheet.
To keep it running, put the account list on a recurring export: each run returns only new people at your accounts. That is what we call IBAO (Intent Based Automated Outreach): fresh, signal-led leads flowing into your CRM without a manual pull.
How to Build the Same Agent in n8n
n8n runs the same agent from a form, without a coding agent. You type your companies, and the workflow fills the approval sheet.
The AI Agent node only does the one step that needs judgement: resolving company names. Every other step is a fixed node, so the picks match the script.
- Resolve the Companies With an AI Agent Node
- Preview, Pick and Export With MCP Client Nodes
- Approve From a Form Page
1. Resolve the Companies With an AI Agent Node
A form trigger asks for your companies and how many leaders per company. An AI Agent node with Claude and the Evaboot MCP tools resolves each name and saves the account list.
Only two Evaboot tools are switched on for this node, so it cannot export or push by mistake.
2. Preview, Pick and Export With MCP Client Nodes
MCP Client nodes call one Evaboot tool each: the free preview, the lead list, the export. Code nodes in between apply the same picking rules as the script.
One timing detail matters: the workflow waits 30 seconds after saving the lead list. In our test, an export started at once returned no leads.
3. Approve From a Form Page
When the sheet is filled, the form shows a second page: push these leads to your CRM? Yes pushes the export through Evaboot; no stops the run.
Our n8n test took 3 companies, picked 9 sales leaders, wrote the briefs, and pushed the 9 leads to HubSpot after the yes. The whole run took 3 minutes 21 seconds, approval included.
AI Sales Agent Example: 10 Target Accounts
We ran the agent on 10 B2B software companies from 27 to 378 employees. Here is what each step returned:
| Step | Result |
|---|---|
| Companies resolved | 10 of 10, one name had two LinkedIn pages |
| People previewed (free) | 78 |
| Picked and exported | 27 picked, 25 returned, 1 had changed company |
| Emails | 23 of 25 found, 13 rated safe |
| Briefs | 5 strong, 3 medium, 1 weak, 1 company with no sales title |
| Pushed to HubSpot | 25 leads, one call |
The strong accounts all had something you could open an email with: a VP Sales two months into the role, two new sales leaders in six months, or headcount up 50% in two years.
The weak one is useful too. "No timing trigger" is an honest answer, and it tells you to put that account in a slower sequence instead of inventing a reason.
Common Mistakes When Building an AI Sales Agent
Four mistakes cause most broken sales agents, and our test hit two of them before the fixes:
- Exporting Before You Pick
- Letting the Model Fill Gaps From Memory
- Trusting the Search Results as Current
- Pushing Without a Human Yes
They all come down to scope, data and guardrails:
1. Exporting Before You Pick
Exporting a whole search and filtering afterwards pays for people you throw away. Preview for free, pick, then export only the picks.
2. Letting the Model Fill Gaps From Memory
Asked to "personalise", a model will happily mention a funding round that never happened. Give it only the export fields, and let it say "no timing trigger".
3. Trusting the Search Results as Current
Search results lag: in our run, one picked leader had already moved to another company. Check the export against your target list before anything leaves the sheet.
4. Pushing Without a Human Yes
An agent that writes to the CRM unchecked spreads its mistakes to every rep who opens a record. Keep an approval step until the output earns your trust.
Frequently asked questions
How do I create an AI sales agent?
Pick one job, connect a model to tools, and add an approval step.
Connect Claude or another model to your sales tools through an MCP server, write the job and its rules in one file, and test on 3 to 5 accounts.
The example in this guide does account research with Claude and the Evaboot MCP server, and the full setup is in a public GitHub repo.
Can I build an AI sales agent on my own, without code?
Yes. Claude with an MCP connector needs no code: you write instructions in plain English and approve its actions.
For a fixed workflow, n8n runs the same agent from a form, with an AI Agent node for the judgement step and ready-made nodes for the rest.
How much does an AI sales agent cost?
Mostly the data. In our test, the export cost 1 Evaboot credit per lead: 25 credits for 10 accounts.
Add your Claude plan, and n8n if you use it. Previewing before you export is the biggest saving: it cut our export from 78 people to 25.
Can AI agents do sales?
They can do the research and admin around selling. Finding buyers, checking emails, writing briefs and updating the CRM are all good agent jobs.
The conversation itself still works best with a person: our agent stops before any email and hands the opener to a rep.
What is the difference between an AI sales agent and a chatbot?
A chatbot answers; an agent acts. A chatbot replies to questions, often from a script.
An agent takes a goal, such as "research these 10 accounts", and chains tool calls to reach it: search, pick, export, write and push.
Rate this article