AI sales agent: what the software does on its own, where a person still signs off, and the rules it has to follow.

Definition

An AI sales agent is software that uses AI to perform sales tasks with some autonomy: researching accounts, building lead lists, drafting and sending outreach, answering inbound leads, qualifying interest and booking meetings on a rep's calendar.

The word that matters is autonomy. A tool that suggests a line for a rep to type is an assistant. A tool that decides who to contact, writes the message, sends it and handles the reply is an agent, even when a person approves some of those steps.

NIST's AI Risk Management Framework describes an AI system as an engineered or machine-based system that, for a given set of objectives, generates outputs such as predictions, recommendations or decisions, and notes that AI systems operate with varying levels of autonomy. An AI sales agent sits toward the autonomous end of that range.

In sales, the goal is a qualified meeting or a qualified lead, and the tools the agent uses are the CRM, the inbox, the calendar and data providers. The agent takes a series of actions across those tools, not a single answer in a chat window.

AI sales agents are sold under several names: AI SDR, AI BDR, autonomous sales agent, digital worker, AI sales assistant, plus product names that change between releases. The names overlap, so compare products by asking what each one does without a person in the loop.

What an AI sales agent does, and what it does not do

The table lists the tasks AI sales agents are built for, how much of each the software can do alone, and where a person should still review the work before it reaches a prospect. The review column is a recommendation written for this page, not a vendor setting.

What AI sales agents do not do well is also a list: understand a buying group's politics, notice that a prospect's tone changed, judge whether a deal is worth a discount, or know that the person they are writing to already declined last quarter unless that is in the data.

What vendor documentation says AI sales agents do

Product pages describe AI sales agents in outcomes. Help documentation describes them in settings, which is more useful to a buyer. The examples below come from vendor documentation read on Oct 1, 2026. They are examples of documented features, not a ranking, and the names and settings change between releases.

Microsoft Dynamics 365 Sales

Microsoft Learn lists several AI agents in Dynamics 365 Sales. Its Sales Qualification Agent page describes two modes. In Research-only mode, the agent researches assigned leads, checks them against a target customer profile and generates an outreach email that a seller decides whether to send.

In Research and engage mode, the same page says the agent also sends outreach, follows up, detects purchase intent, checks BANT criteria and hands promising leads to sellers. The page states that the agent does not replace the seller's judgment or decision-making process.

Microsoft's responsible AI FAQ for the engage mode adds two details worth copying into any evaluation. Admins can configure an AI disclaimer for the emails the agent sends. And to research a lead, the agent passes the company name, website and any fields you choose to Bing Search.

The pace of change shows on the same site. Microsoft Learn says that from Sep 30, 2026, new instances of its Sales Close Agent cannot be created, that existing ones are removed on Oct 30, 2026, and that customers should create a Sales Development agent instead.

HubSpot prospecting agent

HubSpot's Knowledge Base describes a prospecting agent that monitors companies for intent signals, such as funding news or research on topics you choose, finds up to three contacts per enrolled company that match your personas, and generates personalized outreach based on CRM data and those signals.

The agent runs on plays: written instructions grouped into audience, selling context, outreach, guardrails and automation. The automation tab offers two options. With Review before sending, a person approves each message. With Send automatically, the agent sends without review. A daily limit caps new contact suggestions per play.

Salesforce Agentforce

Salesforce's Trailhead describes its sales development agent as handling personalized outreach to leads, responding to questions and scheduling introduction meetings for reps, grounded in the company's sales, product and customer data. It describes a separate Sales Coach agent that runs role plays tailored to a rep's deal.

A Trailhead setup unit adds two practical points. The agent works as its own Salesforce user record with permissions, connected to a mailbox and calendar. And without a knowledge source, the unit says, the agent relies on general AI training instead of your company's own answers.

Read together, the documentation agrees on a shape: a target profile, written instructions, a knowledge source, a choice between review and automatic sending, and a handoff to a person. Those are the settings to ask about, whatever the product is called this quarter.

Types of AI sales agents

Vendors split the category in slightly different ways. Salesforce's Trailhead, for example, separates autonomous and assistive agents. Two splits are useful when you compare products: how much the software decides on its own, and which direction the leads come from.

Autonomous sales agent

An autonomous sales agent runs a task from start to finish inside rules a person set: it builds the list, writes, sends, follows up and books. People review samples and exceptions, not every message. This is the model AI SDR products describe, and the Send automatically setting in HubSpot's documentation is one example of it.

AI sales assistant

An AI sales assistant, sometimes called a supportive or assistive agent, does the work but leaves the final action to a rep: it researches, drafts, summarizes and suggests, and the rep clicks send. Some CRM and sales engagement platforms include assistant features next to their agents.

Inbound agents and outbound agents

Inbound agents answer people who already raised a hand: website chat, demo requests, form fills, replies to marketing email. Outbound agents contact people who did not ask. In this page's view, the legal, reputational and deliverability risks are much higher for outbound, because the recipient never agreed to anything.

AI sales agent vs chatbot

A rule-based chatbot follows a script or decision tree written in advance and waits for a visitor to start. An AI sales agent reads context, decides the next step inside its rules, and takes actions in other systems: it sends a follow-up, books a meeting, updates the CRM or hands a lead to a rep.

This comparison was written for this page, and the line blurs: some chat tools now use language models, and some AI sales agents work mainly in website chat. The useful question stays the same. Does the software only answer, or does it also act in other systems without a person approving each step?

AI sales agent use cases across the sales cycle

The AI SDR gets the attention, but vendor documentation describes agents at other points of the sales cycle too. The table maps documented use cases to what the documentation says the agent does. Each row names one documented example; it is not a list of every product that does it.

Assistive use cases, such as coaching and questions on sales data, carry less outside risk, because nothing reaches a buyer unless a rep sends it. Customer-facing use cases, such as outreach and nurturing, need the rules, approvals and legal checks described further down this page.

Where each agent sits next to your CRM, data providers and sending tools is a question of the sales tech stack. Two agents that both write to the CRM, or both email the same contact, are a stack problem before they are an AI problem.

Benefits of AI sales agents, and what they cost sales teams

The benefits sales teams expect from AI sales agents come from time, not from better selling. Agents work across time zones, respond in real time, and take the repetitive sales tasks that fill a rep's day: research, data entry, first drafts and follow-ups. The list below is this page's view.

• Speed to lead. AI sales agents can answer inbound leads in real time, day and night, while the customer is still on the website, rather than the next working morning.
• Coverage. Agents can work the leads human sales teams skip: small accounts, old leads and follow-ups after the third touch, based on rules instead of on who has time. Microsoft's FAQ describes its engage mode as focused on lower priority leads sellers cannot manage due to volume.
• Consistency. Every lead gets the same qualifying questions and the same data captured in the CRM, which makes sales data comparable across the team.
• Research at scale. Account research a rep would skip for a small account is done for every prospect on the list, so personalized outreach is based on sources a reviewer can check.
• Testing without hiring. Sales teams can test a new segment, region or offer with an agent before they hire people for it.

The costs are also time. Someone on the sales team owns the rules, reviews samples and fixes the data, every week in the first month and every month after that. Without that owner, AI sales agents send the same mistake to every customer and prospect on the list.

What is an AI SDR?

An AI SDR is an AI sales agent built to do the job of a sales development representative: find and research prospects, send the first touches, follow up, handle replies, qualify interest and book meetings for account executives. It is the form of AI sales agent this entry spends the most time on.

A human SDR does the same job with judgment, a phone and a LinkedIn account, at a much lower volume. The AI SDR trades judgment for volume and speed, which is why the rules it runs under matter more than the model it uses.

Some teams use an AI SDR to replace outbound headcount; others use one to cover the work SDRs skip, such as inbound leads that arrive at night, old leads nobody followed up, or accounts too small for a person's time.

What an AI SDR does with one lead

A lead arrives or matches a signal. The AI SDR researches the company, checks fit against the profile, writes a first email and sends it or queues it for approval. If the lead replies, it classifies the reply, answers questions from the knowledge source, offers meeting times and hands the conversation to a rep.

In this page's view, qualification is where AI SDRs and human SDRs differ most. An AI SDR can ask the same qualification questions every time and record the answers. A human SDR hears hesitation, follows a tangent, and knows when a lead needs a call. In this page's view, many teams will want both.

AI SDR vs human SDR vs AI guided selling vs sequencing tools

Four things are easy to confuse in sales software pages. The difference is who decides and who acts. This comparison was written for this page.

The line between the first and third columns is recommend versus act. AI guided selling tells a rep what to do next; an AI sales agent does it. Some platforms include several of these, and the settings decide which one you are really running.

A sequencing tool is the older layer underneath, the sending engine inside sales engagement software. It sends the steps of a sales cadence on a schedule, and an AI SDR may use one to send.

AI sales agent tools: the main categories

AI sales agent tools come in a few categories. They are listed here as types, not ranked, and a team may already pay for one of them inside a platform it uses. The category decides where the agent gets its data and who on the team owns it. The table was written for this page.

Sales teams with a busy CRM have a reason to start with agents that live inside it, because customer and deal data are already there. Teams that mainly need new pipeline look at standalone AI sales agents, and should check where their contact data comes from.

How an AI sales agent works

AI sales agents, whatever they are called, run the same loop. Data goes in, rules limit what the agent may do, a language model decides and writes, an approval step catches errors, and the action is logged so the next run is better informed.

Data

The agent reads your ideal customer profile, CRM records, past emails, product documents and outside data such as lead enrichment and buying signals. Microsoft's FAQ puts it plainly: the more information available about a lead, the better the outputs.

Rules

Rules are the guardrails: which companies and titles are in scope, daily sending limits, topics it may not discuss, claims it may not make, when it must hand off to a person, and who is excluded, such as customers, open deals and anyone who opted out.

Model

A large language model does the reading, the deciding and the writing, sometimes with separate scoring models for fit and intent. The model is good at language and weak at facts it was not given, which is why the knowledge source matters as much as the model.

Approvals

Approval can be set per message, per batch, or on a sample after sending. Start strict and loosen only what the review shows is safe. Inbound answers on pricing, security, contracts and legal terms should always route to a person.

Where AI sales agents fit in the sales process

AI sales agents work best at the top of the B2B sales process, where tasks repeat and the cost of one mistake is a lost prospect rather than a lost deal. In this page's view, they get less useful as deals get larger and more personal.

Two jobs make a safer start: answering inbound leads fast, where speed matters and the lead already asked, or working old and low-priority leads that no rep has time for. Cold outbound lead generation at scale is the riskiest place to start.

Risks and rules for AI sales agents

An AI sales agent sends messages in your company's name, so the rules that apply to your reps apply to it, at a higher volume. None of this section is legal advice; check the laws that apply to your market with counsel.

Deliverability

Volume is a fast way to damage a domain. Gmail's sender guidelines require SPF or DKIM for all senders and ask that the spam rate reported in Postmaster Tools stay below 0.3%. Senders of more than 5,000 messages a day to Gmail also need DMARC and one-click unsubscribe for marketing messages.

Send outreach from secondary domains with warmed mailboxes, set daily limits per mailbox, verify addresses before sending, and stop a campaign when replies turn negative. The cold email hub covers the setup in more detail.

CAN-SPAM for commercial email

In the United States, the FTC's CAN-SPAM guide states that the law makes no exception for business-to-business email. Commercial messages need honest header information and subject lines, identification as an ad, a valid physical postal address and a clear way to opt out.

Opt-out requests must be honored within 10 business days, and the opt-out mechanism must keep working for at least 30 days after the message is sent. The FTC also says you cannot contract away your legal responsibility when someone else sends email for you, which includes the vendor of your AI sales agent.

AI voice calls and the TCPA

Some AI sales agents place calls with a generated voice. In a February 2024 declaratory ruling, FCC 24-17, the FCC confirmed that the TCPA's restrictions on "artificial or prerecorded voice" cover current AI technologies that generate human voices, because a person is not speaking them.

As a result, the ruling says, such calls need the prior express consent of the called party unless an emergency purpose or exemption applies. The FCC's rules at 47 CFR 64.1200 go further for telemarketing: calls with an artificial or prerecorded voice to cell phones or residential lines need prior express written consent.

Section 64.1200(b) also requires every artificial or prerecorded voice message to state clearly, at the beginning, the identity of the business responsible for the call. The ruling adds that telemarketing messages must offer the specified opt-out methods. Treat AI voice outreach as a legal decision, not a settings toggle.

LinkedIn and automation

In section 8.2 of the LinkedIn User Agreement, members agree not to use bots or other unauthorized automated methods to access the Services, add or download contacts, send or redirect messages, create, comment on, like, share or re-share posts, or otherwise drive inauthentic engagement.

The same section prohibits software, scripts, robots and browser plugins used to scrape or copy the Services, and the agreement says LinkedIn may restrict, suspend or terminate an account that breaches it. An agent that acts through a member's account risks a restricted LinkedIn account. The LinkedIn automation guide covers the details.

Claims about the agent, and claims the agent makes

The FTC treats claims about AI like other advertising claims. Its 2025 final order against Workado bars the company from claiming its AI detection product is effective unless it has competent and reliable evidence at the time the claim is made. That is a fair standard for anything a vendor says its agent delivers.

In August 2025 the FTC sued Air AI, alleging deceptive earnings and refund claims for services whose flagship feature was advertised as conversational AI that could replace human customer service representatives.

In March 2026 the FTC announced a settlement: under the proposed order, Air AI and its owners will be banned from marketing business opportunities. For a buyer, the lesson is to ask for evidence and read refund terms before signing.

The same caution applies to what the agent writes. The FTC's Endorsement Guides say an endorsement must reflect the honest opinion of the endorser. Its rule at 16 CFR 465.2 makes it a violation for a business to create a testimonial that misrepresents that the testimonialist exists or used the product.

Whether that rule reaches a given B2B message is a question for counsel. Forbidding the agent to quote customers or invent testimonials is simpler, and it belongs in the brief. Disclosure is a product setting in some tools: Microsoft documents an AI disclaimer admins can add to emails its qualification agent sends.

Data protection

AI sales agents process personal data: names, work emails, job histories and the content of replies. Privacy laws such as the GDPR in the EU and UK and state privacy laws in the US may apply, depending on where the prospects are.

Know where the data came from, how the vendor stores it, and whether it trains models on it. Microsoft Learn, for example, notes that data used by the agents in Dynamics 365 Sales might be processed and stored in regions outside the user's primary region. Ask every vendor the same question.

Accuracy

NIST's Generative AI Profile, NIST AI 600-1, names this risk confabulation: the production of confidently stated but erroneous or false content. An agent can repeat one such error across thousands of messages. Failures to watch for are an invented detail about the prospect, a wrong job title, or a feature the product does not have.

The same NIST profile names automation bias, an excessive deference to automated systems. That is the reviewer's risk: approving the agent's work because it usually looks right. The control is the same as for a new rep: approved claims, forbidden claims, and a person reading samples every week.

How to evaluate an AI sales agent

This checklist was written for this page. It is vendor-neutral on purpose: run it against any product, including the agent features already inside your CRM or sales engagement platform.

• Approvals. Can you require approval per message, per batch or by sample, and change that per campaign?
• Rules and exclusions. Can you block customers, open deals, competitors, regions and people who opted out, and does the block hold across every channel?
• Sources. Does each researched claim show where it came from, so a reviewer can check it in seconds?
• Knowledge base. Can you control which product documents and answers the agent uses, and forbid topics?
• Sending setup. Does it respect mailbox limits, support your own domains and handle unsubscribes and bounces automatically?
• Handoff. How does it pass a lead to a person, with what context, and how fast?
• CRM integration. Does it read and write your CRM without creating duplicates, and can you audit what it changed?
• LinkedIn. Does it automate actions on LinkedIn accounts? If yes, weigh that against the User Agreement.
• Voice. If it places calls, how does it handle consent, caller identification and opt-outs under the TCPA rules?
• Disclosure. Can you add an AI disclaimer or signature to what the agent sends, if your team decides to?
• Data terms. Where is data stored and processed, who can access it, is it used to train models, and how do you delete it?
• Documentation. Is there a public help page for each setting the demo shows, with a last-updated date?
• Pilot terms. Can you run a limited pilot on your own data before a long contract?

How to roll out an AI sales agent, step by step

NIST's AI Risk Management Framework organizes AI risk work into four functions: govern, map, measure and manage. The steps below were written for this page and follow a similar order in sales terms: decide who owns it, define the job, measure it against a control, and adjust.

1. Pick one job

   Choose a single, narrow task, such as answering inbound demo requests or following up on leads older than 90 days, where the result is easy to measure.

2. Write the rules before the prompts

   Define the ICP, exclusions, approved claims, forbidden topics, sending limits and the moments when a person takes over. Get sales, marketing and legal to agree.

3. Prepare the data and the domains

   Clean the CRM records the agent will read, verify contact data, and set up authenticated, warmed sending domains separate from your main domain.

4. Run with full approval

   For the first weeks, a person approves every message. Log each edit and each rejection, and why, and turn them into rules.

5. Compare with a control group

   Run the same job for a similar segment without the agent, and compare meetings held, pipeline created and complaints over the same period.

6. Loosen approval, keep sampling

   Move to batch or sample approval only for message types with no errors in review, and keep a weekly sample for everything.

Metrics for AI sales agents

Will AI sales agents replace SDRs?

In this page's view, AI sales agents replace tasks, not the whole role. They are strong at research, first drafts, fast inbound replies and follow-up that people skip. They are weak at phone conversations that need judgment, multi-threading into a buying group, and reading when a prospect needs a person.

The vendor documentation points the same way. Microsoft's qualification agent hands promising leads to sellers, HubSpot's offers a review step before sending, and Salesforce's passes leads to a rep. None of the pages read for this entry describes a setup with no person in it.

The SDR job shifts toward owning the agent's rules, reviewing its work, and spending more time on the accounts and conversations that deserve a human. A team that cuts every SDR also loses the people who knew which messages were wrong.

Mistakes with AI sales agents

• Starting with cold outbound at full volume instead of one narrow, measurable job.
• Sending from the main company domain.
• Letting the agent send before anyone read a sample.
• Calling a message personalized because it contains a first name and a company name.
• No exclusion list, so customers and open deals get cold emails.
• Automating LinkedIn accounts and losing them to a restriction.
• Turning on AI voice calls without checking consent and caller identification rules.
• Letting the agent quote customers or write testimonials nobody gave.
• Judging the agent on messages sent or meetings booked instead of meetings held and pipeline.
• No human handoff for pricing, security and legal questions.
• Assuming the vendor is responsible for CAN-SPAM compliance.
• Trusting a product name or feature list from last quarter instead of reading the current help documentation.
• Ignoring the edits reviewers make, when those edits are the rules the agent is missing.

In a sequence

Before an AI sales agent writes the first touch of a cadence, it needs a brief. The brief below was written for this page: it tells the agent who to contact, what it may say, and when to stop and hand over to a person.

Goal: book a first meeting with {{targetRole}} at {{companyType}} companies that {{signal}}.

In scope: {{industries}}, {{companySize}} employees, {{regions}}.
Never contact: current customers, open opportunities, {{competitors}}, anyone who opted out.

You may say: {{approvedClaims}}.
You may not: quote prices, promise terms, name customers, or state anything about the prospect you did not find in a source you can link.

Hand over to {{ownerName}} when the prospect asks about pricing, security, contracts or legal terms, or replies with an objection.

Limits: {{dailyLimit}} new contacts per mailbox per day. Stop the campaign if unsubscribes or complaints rise.

Every message includes our postal address and an unsubscribe line.

Frequently asked questions

Cite this definition

Jeluvi, "AI sales agent", B2B sales glossary, https://jeluvi.com/glossary/ai-sales-agent/, last checked Oct 1, 2026.