AI SDR: What It Can Do, What It Can't, and How to Start

Most people who search for an AI SDR want one thing: more meetings without hiring another rep. That's a fair goal. But the tools are usually sold as "an autonomous rep that sends thousands of emails," and that is the version that quietly damages your pipeline.

Across the 35+ agents I've put into real businesses, the sales agents that hold up do less than the demos promise, and they do it on better leads.

What is an AI SDR?

An AI SDR is an AI agent that does the repetitive part of a sales development rep's job: researching leads, checking them against your ideal customer profile, writing a personal first message and booking or routing the ones that fit. It reads and decides; your CRM, enrichment data and inbox supply the facts.

A human SDR spends a surprising share of the week not selling. Salesforce's State of Sales report found reps spend only about 28% of their week actually selling. The rest is research, data entry, admin and internal work. That gap is where an AI SDR earns its place.

Here is how the main options compare:

The useful version is a helper that hands your sales team better leads and better drafts. Not a replacement that talks to prospects on its own from day one.

What can an AI SDR actually do well?

An AI SDR does well on tasks with lots of reading, a clear rule and a cheap mistake: enrichment, qualification, first drafts and routing. These are the hours a rep loses before the real conversation starts.

The tasks I'd hand over, roughly in order of risk:

• Enrichment. Pull company size, role, industry and recent news for a new lead. Nobody sees it but your team.
• Qualification. Score the lead against your written ideal customer profile and explain why in one line.
• Routing. Send good fits to the right rep, park the rest, flag anything odd for a person.
• First drafts. Write a short, specific first message based on the research, for a person to approve.
• Follow-up reminders. Notice that a lead went quiet and draft a follow-up, instead of letting it slip.
• Meeting booking. Offer times and book the call when someone says yes.

Qualification is where I'd start. When I ran an AI check on my own LinkedIn network, it went through 25,033 connections in 71 seconds for under a dollar and found 435 warm prospects. That's the job done well: reading at a scale no person would, against one clear question. Here's how that ICP check works.

If you'd rather start from a working build, the free n8n AI SDR setup researches leads and drafts outreach, and you can adapt it to your own profile.

What should an AI SDR never do on its own?

An AI SDR should never send messages at volume without review, make promises on price or terms, or argue with a prospect who pushes back. Those moments decide whether someone trusts your company, and a wrong word there costs more than a slow reply.

The risk isn't only reputation. It's deliverability. Since February 2024, Google and Yahoo require bulk senders to authenticate their domain, offer one-click unsubscribe and keep spam complaints below 0.3%. An agent that blasts generic messages to a cold list can push you over that line, and then your normal sales email starts landing in spam too.

A simple rule I use: if the message names a price, makes a commitment, answers an objection or goes to more than a handful of people at once, I approve it myself. The rest the agent may draft.

Should you buy an AI SDR tool or build your own?

Buy a tool if you want to start this week and your process is standard; build your own if your profile, data or tone is specific to you. Both work. The choice depends on who maintains it, not on the demo.

Either way, the work that makes it succeed is the same: a written profile, clean data and a person reviewing the output in the first weeks.

How do you set up an AI SDR, step by step?

Setting one up starts with writing your ideal customer profile first, starting on leads that already know you, and only widening what the agent may send when the numbers say it's ready. Here's the order I'd follow:

1. Write the profile down. Who fits, who doesn't, and the three signals that matter most. If two people on your team would score the same lead differently, the agent will too.
2. Pick one lead source. Inbound form leads or your own network. Not a bought list.
3. Connect read-only. CRM, enrichment, LinkedIn export or form data. Lookups only, no sending.
4. Turn on qualification. Every lead gets a score and a one-line reason. Check 50 by hand against what your best rep would decide.
5. Add drafts. The agent writes the first message; a person edits and sends.
6. Track edits. Count how many drafts go out unchanged. That number tells you when the agent is ready for more.
7. Widen carefully. Let it send follow-ups or book meetings on its own only for lead types where drafts rarely need changes.

Step 1 is where most projects fail. "Mid-size B2B companies" is not a profile. "Service businesses with 10 to 50 people, a founder still doing sales, and a team spending hours on lead admin" is. The technical side follows my general agent build process.

How do you measure whether an AI SDR is working?

You measure an AI SDR by qualified meetings booked and the quality of replies, not by how many messages it sends. Sending more is the easiest number to improve and the one that tells you least.

Track these per lead source:

• Qualification accuracy: how often your team agrees with the agent's score.
• Draft edit rate: how many drafts go out without changes.
• Positive reply rate: replies that move the conversation forward, not "unsubscribe".
• Meetings booked and held: the number your sales team actually cares about.
• Complaints and bounces: your early warning for deliverability trouble.

Gartner predicted in 2023 that by 2028, 60% of B2B seller work will be done through conversational interfaces powered by generative AI. Maybe. Even so, the companies that benefit will be the ones that measured what the agent did on real leads before they scaled it.

What breaks AI SDRs in production?

AI SDRs break for the same reasons other agents do: no written process, no record of what was sent, access that's too wide and no person on the risky steps. The model is rarely the problem.

The sales version looks like this:

• A vague profile. The agent qualifies fast and wrong, and your reps stop trusting the scores.
• Stale data. The prospect changed jobs six months ago; the message congratulates them on the old role.
• No log. A prospect says "your bot said we'd get a discount." Without a record of every message, you can't check.
• Too much access. An agent that can email your whole CRM will one day email the wrong segment.

The broader checklist of production failures covers the rest. The sales-specific rule on top of it: when the agent isn't sure a lead fits, it says so and hands it to a person.

Start with qualification on the leads you already have. It touches no prospect, it shows you where your pipeline really comes from, and it tells you whether outreach is worth automating at all.