AI sales agent software › How to build an AI sales agent

How to build an AI sales agent: the six parts, and when to buy instead

An outbound AI sales agent is six parts wired together: a source of people, research on each one, a model that writes, an approval step, a sender, and a loop that reads replies and results. You can build a working test from a language model's API, a contact data API and a mailbox. The hard parts aren't the writing; they're deliverability, handling replies and knowing which messages work.

Prices read November 13, 2026 from Enrow's pricing page, Moab's Prospeo docs and the vendors linked below. Sending rules from Google's email sender guidelines.

The six parts

People. A search over a contact database by title, company size, industry and location, then an email finder. Through Prospeo's API, as Moab's docs price it, a search costs $0.0245 per page of up to 25 people and each email found $0.0245; Enrow finds emails from $17 a month for 1,000. Start with a list you'd be happy to email by hand.
Research. For each person, fetch what the message will lean on: the company's site, a recent announcement, what the role usually owns. Store the source with each fact. A model given no facts invents them, and an invented compliment is worse than none.
Writing. A prompt that holds your offer, who it's for, the facts from step 2 and the rules: length, one ask, no claims you can't back, how to sign off. Change one thing at a time (the audience, the channel or the angle) so you can tell what made the difference.
Approval. A queue where a person reads, edits and approves each draft before it goes. Most products on this site have one: HubSpot's "Review emails", Artisan's approval mode, Agent Frank's co-pilot and Moab's approve step. Build this before you build automatic sending.
Sending. The Gmail or Outlook API from a mailbox on a separate domain, or a sending tool through its API. Google requires every sender to Gmail to set up SPF or DKIM and keep the spam rate reported in Postmaster Tools below 0.1%, never reaching 0.3%; above 5,000 messages a day it requires SPF, DKIM and DMARC, and one-click unsubscribe for marketing messages. LinkedIn is harder: its User Agreement bans using bots to send messages.
Replies and results. Stop the sequence the moment someone replies, sort replies (interested, not now, wrong person, unsubscribe), honor every opt-out, log replies and meetings in your CRM, and count per experiment: sent, clicked, replied, meetings, and what it cost.

The smallest version worth running

Before any automation, run the loop by hand for 50 people: a script searches and researches, the model drafts, you read every draft and send the good ones yourself. In outline:

people = search(title="Head of Ops", industry="logistics", size="20-200", limit=50)
for person in people:
    facts = research(person.company_domain)        # keep the source of each fact
    draft = write(offer, person, facts, rules)     # one angle for the whole batch
    queue.add(draft)                               # a person approves before anything sends

Then wait. Replies take days, and AiSDR's own rule of thumb is 1 to 3 meetings per 100 people contacted, so 50 people tells you whether the drafts are good enough to read, not whether the campaign works. If you need new mailboxes, warm them first: Salesforge tells Agent Frank customers to allow two weeks.

Where homemade agents go wrong

Too much volume from one domain. Spam complaints above Google's 0.1% target hurt delivery for everything else from that domain, including your ordinary mail.
Personalization that isn't true. Wrong job titles, stale news, a product the company dropped. Keep sources and let the approver see them.
Follow-ups after a reply. The most common embarrassment. Check for replies before every send, not once a day.
Lost opt-outs. Someone who says stop on LinkedIn must also stop getting email. Keep one list.
No measurement per angle. If every message is different, you learn nothing. Group drafts into experiments and compare them.
LinkedIn automation. Scripts that send connection requests or messages break LinkedIn's terms and put the sending account at risk, whoever wrote the script.

Build or buy: what each costs

500 new people a month, list prices read November 13, 2026. All the sums are on AI SDR pricing.
Route	Monthly cost	You still do
Build it yourself	About $18 for data (Enrow's $17 plan plus Prospeo searches), plus model use and a mailbox	Everything: code, sending, replies, measurement, upkeep
Moab	$49, plus the same data	Connect your data and sender accounts; approve drafts
Artisan	$250 to $280	Set up campaigns; approve if you choose
AiSDR	$750	Set up campaigns; co-pilot for replies if you choose

No-code workflow tools such as Gumloop, n8n, Zapier, Lindy and Relevance AI can wire these parts together without writing code; the top Google result for "ai sales agent software" is Gumloop's guide to doing that. They save the plumbing, not the deliverability and reply work. The build route's real cost is your time, now and every week after: data APIs change, mailboxes need watching, and reply handling has edge cases. Building makes sense if outbound is your product's core, you need something the tools don't do, or you're learning how it works. Otherwise buy, and spend the time on the list and the offer.

The middle path: drive a product from your own code or assistant

Several products now expose their agent through an API or an MCP server, so you can run them from your own scripts or from Claude or ChatGPT without building the plumbing. Amplemarket sells an MCP add-on; Salesforge has an MCP server and a command-line tool; Qualified lists "MCP and API access" on its Enterprise plan; 11x lists an API; Artisan's Employee plan takes webhooks as a data source.

Moab exposes everything through a REST API and an MCP server: "campaign agents" (one audience, channel and angle each) write drafts, a run "never sends", and approved drafts go out through your Gmail, Resend, lemlist, HeyReach or Instantly account, with clicks, replies and cost per agent by day. It brings no contact data and has no track record yet.

For what these agents are and how they differ from voice and chat agents, see what is an AI SDR.

Questions

What do I need to build an AI sales agent?

A source of contacts and emails, a way to research each company, a language model with a prompt for your offer, a step where a person approves drafts, a mailbox on a separate domain that meets Google's sender rules, and a way to stop on replies and record opt-outs.

Can I build an AI sales agent with ChatGPT?

A chat assistant can research a company and write a draft. An agent also needs a list of people, a mailbox it can send from on a schedule and a record of who replied, which you connect through an API or an MCP server, or get in a product.

How long does it take to build an AI sales agent?

A test where the model drafts and you send by hand can run within days. Results take longer: new mailboxes need about two weeks of warm-up, and at 1 to 3 meetings per 100 people you need a few hundred sends to judge an angle.

About this page

Published by Winding AI, Inc., the company that makes Moab, which is named above as one option. Winding AI earns from Moab subscriptions, not from links to other products. Back to the AI sales agent software comparison.