What is an AI SDR? How AI sales agents actually work in 2026
An AI SDR finds prospects, researches them, writes outreach and handles replies. Here's how the pipeline works, the three kinds of AI SDR tools, where they go wrong, and 10 questions to ask before you buy one.
An AI SDR (AI sales development representative) is software that does the top-of-funnel work a junior sales rep would do. It finds companies that fit your ideal customer, picks the right people, finds their email addresses, writes personal outreach, sends it, follows up and sorts the replies.
The pitch is simple: outbound without hiring. In practice the category ranges from genuinely useful to quietly harmful, and the difference is mostly in details buyers don't check. This guide explains how an AI SDR works step by step, the three main kinds on the market, and how to tell a good one from a risky one.
What an AI SDR does, step by step
Under the marketing, almost every AI SDR runs the same pipeline.
1. Understand what you sell
You give it your website, and it drafts a product profile: what you sell, who buys it, the problems it solves, your competitors and proof points. The good ones make you approve this. Every email it writes later is only as accurate as this profile.
2. Define the ideal customer
It turns your description ("Series A SaaS companies in the UK with a small sales team") into searchable criteria: industry, size, location, job titles, plus softer signals like "hiring SDRs" or "recently raised".
3. Find companies and people
This is where tools differ most. Some search a proprietary database of hundreds of millions of profiles. Others plug into licensed data providers. A few scrape, which carries legal and reputational risk for you, not just for them.
The better tools also score each lead against your criteria and explain why it fits, so you can sanity-check the list before anything is sent.
4. Find and verify email addresses
Finding an address is easy. Finding one that won't bounce is not. Email finders guess patterns; verifiers check whether the mailbox exists. Many B2B addresses are on "catch-all" domains that accept everything, so they can't be fully verified.
This step quietly decides your bounce rate, and with it your deliverability. Expect a large share of found people to end up without a usable address. That's normal. Treating unverified addresses as verified is not.
5. Research and write
For each person, the AI pulls context (their company website, role, recent news) and writes an email. The best output reads like a thoughtful human wrote it in two minutes. The worst invents things: a podcast they never appeared on, a "recent post" that doesn't exist, a customer you don't have.
6. Send
Emails go out from a set of mailboxes at a controlled pace, in the recipient's working hours. Whose mailboxes is one of the most important questions in this whole guide. More on that below.
7. Handle replies
Replies are classified: interested, not now, out of office, wrong person, moved to a new address, unsubscribe. Each needs a different action. Some tools also draft or send replies automatically.
The three kinds of AI SDR tools
Done-for-you on a shared sending network
You paste your website, set a daily budget and the vendor does everything, including sending from its own pool of domains and mailboxes. Setup takes minutes and pricing is often per email.
The trade-off: your emails go out from domains you don't own, sometimes under sender names that aren't yours. Prospects reply to "Chris" when your signature says something else. Your brand's first impression rides on someone else's reputation. You also can't take that sending history with you if you leave.
Platforms that send from your own mailboxes
You connect your own sending mailboxes (usually on secondary domains, so your main domain stays safe). The AI does the finding, writing and reply handling, but every email comes from you.
Setup takes longer, since you need domains, DNS records and a warm-up period. In exchange your name is on every email, replies come to you, and the sending reputation you build belongs to you.
Do-it-yourself stacks
A data tool for leads, an enrichment tool for research, a sending tool for sequences, and an LLM in the middle glued together with automations. Maximum control and the most moving parts. Great if you enjoy operations; a time sink if you don't.
Where AI SDRs go wrong
Most bad outcomes come from a short list of failure modes:
- Invented personalisation. The model fills gaps with plausible-sounding facts. One fabricated compliment and the prospect knows the email was a guess.
- Unapproved claims. "We've helped 500 companies" when you've helped twelve. A good system only uses facts you've approved, with a source.
- Identity mismatches. Display name says one person, signature says another.
- Ignoring reply context. Sending step 3 to someone who replied "we've moved to a new email" or "talk to my colleague". This is the fastest way to earn a spam complaint.
- No unsubscribe. Many tools send plain-text first emails with no opt-out at all. That's a legal problem in several countries and a reputation problem everywhere. See Is cold email legal?
- Opaque metrics. Counting "accepted by the mail server" as "delivered", or treating tracked opens as engagement when privacy features inflate them.
- Opaque costs. If you can't see what each lead and each reply cost, you can't tell whether it works.
AI SDR vs a human SDR
An AI SDR is good at volume, consistency and the dull parts: research, list building, sending on time, never forgetting a follow-up. A good human SDR is better at judgement: spotting a real buying signal, handling a nuanced reply, building a relationship over months.
For founders and small teams, the realistic comparison isn't "AI vs a full-time SDR". It's "AI vs the founder doing outbound at 11pm, or not at all". In that comparison a well-run AI SDR wins easily, as long as it doesn't damage your reputation along the way.
10 questions to ask before you buy an AI SDR
- Whose domains and mailboxes send the emails? Can I use my own?
- What name appears in the From line? Is it mine?
- Where does the lead data come from? Is it licensed for this use? Can I see the source of each record?
- How are email addresses verified? Are catch-all and unverified addresses labelled?
- Can I see exactly what each prospect will receive before it sends?
- What stops the AI from inventing facts? Does it only use claims I've approved?
- What happens on a reply, an out-of-office, a "moved to a new address" and an opt-out?
- Is there a working unsubscribe link and a postal address in every email?
- How are metrics defined? What does "delivered" mean here?
- Can I export my prospects and history if I leave?
If a vendor is vague on questions 1, 3 or 7, be careful.
How Imperagent approaches it
We built Imperagent because we wanted an AI SDR for our own portfolio of products and didn't like the trade-offs above. It sends from your mailboxes under your name, shows the source of every lead, writes only from facts you approve, and stops the moment someone replies. It also includes a one-click unsubscribe and your postal address in every email.
If that's what you're after, claim one of the founding seats.
Sources
Outbound that runs itself, under your own name.
- Finds and verifies the right people
- Writes from facts you've approved
- Sends from your mailboxes and stops on reply