AI Sales Agent: What It Does, Where It Fails and How to Test One

An AI sales agent is software in which a language model calls tools, such as your CRM, inbox, calendar or a lead database, to finish a sales task with several steps instead of only suggesting text. Most products sold under the name cover one slice of the job: answering inbound leads, outbound prospecting (often called AI SDR or AI BDR), account research, or keeping the CRM current. Today they reliably take over research, first drafts, routing and record-keeping. Choosing whom to target, judging a deal and standing behind an offer stay with people.

Customermates, an AI-native CRM, publishes this guide and describes its own setup in a clearly labelled section. Facts about other vendors come from their own pages, checked on 8 October 2026.

Four types of AI sales agent

The label covers products that do very different work. The examples are described from each vendor's own page.

AI SDR versus AI BDR. Vendors use both terms loosely. An AI SDR (sales development representative) usually qualifies leads and hands a meeting-ready opportunity to a salesperson, inbound or outbound. An AI BDR (business development representative) usually opens new outbound conversations. Compare the steps a product actually performs, not its label.

Which steps an agent can take over

Sales development runs through seven steps. The middle column is what works reliably in small-team selling today; the right column is what to check in a demo.

Steps 3, 4 and 7 pay off first for most teams, because they involve a lot of typing and little judgment. Step 1 is the one to keep: an agent that picks its own targets produces volume, not pipeline.

Can an AI sales agent replace your SDR or BDR team?

For inbound work, partly. An inbound agent can answer a visitor at night, ask the qualifying questions and offer a meeting slot, and that is where the gain is most visible. For outbound work in relationship sales, no. An agent can research and draft much faster than a person, but it cannot judge which of fifty plausible accounts is worth a partner's time, or notice that a contact's last post makes the timing wrong.

Agentforce is a good test case. Salesforce describes its SDR agent as working inbound leads within guardrails you set, answering questions from your business data and offering calendar slots. That replaces a share of first-response work. It does not replace the person who owns targeting, discovery calls and the handover.

What does an AI sales agent cost?

It depends on the pricing model. Agentforce prices, checked on 8 October 2026 from Salesforce's pricing page:

Outbound agents add costs that never show up on the agent's price list: the lead data, sending domains and mailboxes, and the hours someone spends reviewing drafts and fixing records. Count those before comparing list prices.

Cold outreach by agents in Germany

An outbound agent that writes to strangers is a legal question before it is a software question. Under § 7 (2) no. 2 UWG, advertising by email without the recipient's prior express consent is an unreasonable nuisance, and the rule makes no exception for business recipients. The narrow exception in § 7 (3) covers existing customers whose address you received in a sale. For phone calls, businesses need at least presumed consent and consumers express consent.

So in Germany an autonomous email agent that writes cold to purchased lists is usually not an option, however good its texts are. Agents fit better where the contact started the conversation: inbound requests, replies, existing clients and LinkedIn conversations you began yourself. GDPR also applies to every contact fact the agent collects. This is general information, not legal advice.

A four-week pilot

1. Pick one workflow

   Choose a task you repeat weekly, for example answering new inbound requests or following up on sent proposals. Write down how often it happens and how long it takes now.

2. Write the rules on one page

   Your qualification rule, tone of voice, what the agent may never promise (prices, dates, discounts) and when it must stop: unclear recipient, opt-out, conflicting records.

3. Week one: read-only

   Let the agent produce briefs and lists only. Compare them with what you know and note every invented fact.

4. Week two: one reversible write

   Allow it to create a task or update one field. Check the records for duplicates and wrong links afterwards.

5. Weeks three and four: drafts

   The agent saves drafts in the real threads, you edit and send. Count how many go out with only small changes.

6. Decide on sending

   Allow sending only for the message type where your edits have become small, and test it with a recipient you control first.

Before you sign with a vendor, ask five questions: which steps it performs itself and which need another product; where its contact data comes from; whether reading, drafting, writing and sending are separate permissions; how an opt-out stops every channel; and what a failed run leaves behind in your CRM.

Example: a small-team sales agent in Customermates

This section describes Customermates, the CRM that publishes this guide. Mate, its assistant, runs in Customermates Cloud and uses hosted AI credits. Routines are saved instructions that Mate runs on a schedule or when a record changes.

Take a three-person consultancy on the Pro plan. Each consultant connects 1 account in the unified inbox, usually the work mailbox in Gmail or Outlook; LinkedIn, WhatsApp, Instagram and Telegram are the other channels. Business includes 3 accounts per user, enough for mailbox plus LinkedIn. Deals have a Single choice stage field: First call, Scoping, Proposal sent, Won, Lost. The firm's services and writing rules sit in one Knowledge Base page that Mate can read in chats and routine runs.

Inbound triage, every morning. A routine set to On a schedule, Every day at 08:00:

Look at the threads in my inbox that received a message from outside
our company since yesterday at 08:00. For each sender, find the
contact and organization or create them. If the message asks about
our services, create a deal in the stage "First call" assigned to me
and save a reply draft in the thread that suggests a short call. List what you did and what you
were unsure about. Do not send anything.

Proposal follow-up, on a record change. A second routine set to When something changes, event Record updated for Deals, changed field "Stage", filtered to deals assigned to the owner: when a deal moves to "Proposal sent", Mate creates a task linked to the deal and its contact, and saves a short follow-up draft in the latest thread.

Both instructions end with "Do not send anything" on purpose. When asked to send an email or chat message, Mate sends it without a second confirmation. It asks for approval before deletions and a few admin actions, and a routine refuses those because nobody is there to approve. A routine runs as its owner, with the owner's channels and permissions. Event triggers ignore further events for a short window after a run starts, so a daily sweep is the safer design for inbound mail. Each run also has a credit cap, so on a busy inbox a run can stop part-way and is then marked Partial; the limits are listed in the routines documentation.

For outbound, Mate works one contact at a time: with a connected LinkedIn account that has its own Sales Navigator subscription, it can search Sales Navigator, save leads to lists, create the contacts in the CRM and write a first message for you to review in the chat. It sends only when you tell it to, one message at a time.

What it costs: Pro for three people is €87 per month, billed monthly, with the inbox, up to 5 routines per person and monthly hosted AI credits per active user; see pricing. Starter, at €36 per month for three people, includes Mate and 1 routine per person but no inbox, so the agent would only see CRM records. The trial runs 7 days.

With your own AI client: Claude, ChatGPT and other clients can do the same CRM work through the Customermates MCP server, on every plan and self-hosted, using your own AI subscription. Messaging tools need Pro or higher in the cloud. The server tells clients to confirm before sending or deleting, but whether you see that confirmation depends on the client.

Frequently asked questions

Vendor and legal details checked on 8 October 2026 against Salesforce Agentforce SDR, Agentforce pricing, HubSpot AI, Salesloft AI agents, Artisan Ava, 11x and § 7 UWG.