# Do AI SDRs Actually Work? What I Automate and What I Don't

> AI SDRs work for research, scoring and drafting, but fail as fully autonomous senders. A builder explains the human-approved setup, its cost and its risks.

By [El Mehdi Bouychou](https://elmehdibouychou.pages.dev/), independent AI marketing consultant · 2026-11-06

Every message drafted by the lead engines I build stops in a queue until a person approves it. That one rule is most of my answer to "do AI SDRs actually work?" They do, as long as they don't send on their own.

## Do AI SDRs actually work? The short answer

AI SDRs work for the legwork of outbound sales: collecting prospect data, scoring accounts against your ideal customer profile and drafting a first message. They break when they send on their own, because volume outruns relevance, deliverability suffers and every mistake goes out under your name. The setup that works is AI drafting and a person approving each message.

An AI SDR is software that does a sales development rep's prospecting, qualification and outreach tasks. Some products stop at research and drafts. The "autonomous" ones also send, follow up and book meetings with no person in between.

I'm [El Mehdi Bouychou](/), an independent AI marketing consultant. I build this kind of lead engine for B2B teams: it collects prospect data, scores leads with Claude, drafts LinkedIn and email messages, and waits for a person to approve each one. So this is a builder's verdict, not a vendor's.

One scope note. Answering inbound leads around the clock is a different job, because the buyer came to you. This article is about cold outbound, where you are the one interrupting.

*Updated November 2026.*

## What AI SDRs are good at: the legwork

AI SDRs are good at the repeatable work that happens before a message exists. That's a large share of an SDR's week, and almost none of it is talking to buyers.

- **Collecting and enriching prospect data.** Company, role, team size, recent changes, from sources you choose and are allowed to use.
- **Scoring fit against your ICP.** A model reads the record and says whether the account fits, and why. It can only score against criteria you've written down. A vague description of a "good lead" gets you vague scores.
- **Summarizing why now.** A short note on what changed at the account: a new hire, a new product, a new market.
- **Drafting the first message.** A starting point someone can fix in a minute instead of writing from a blank page.

These tasks have two things in common. They follow rules you can write down, and a wrong output costs nothing as long as someone reads it before it leaves. A bad score sends a draft to the bin. A bad draft gets rewritten.

That's also my answer to whether AI will take SDR jobs. It takes tasks. The research hours shrink, and judgment, review and conversations with buyers stay with a person, because those are the parts where a mistake can't be taken back.

## Where fully autonomous AI SDRs break

Autonomous AI SDRs break at the send, because that's where volume, accuracy and your name meet. Research and drafts can be redone. A sent message can't be unsent.

### Volume beats relevance

When sending costs almost nothing, tools send more, and relevance is the first thing to give. A prospect who gets an irrelevant message has two options: delete it or report it as spam. The second one feeds the number Google measures your domain by.

### Deliverability

Google requires senders to keep the spam rate reported in its Postmaster Tools under 0.3%, and recommends staying under 0.1%. That's three complaints per thousand messages. Sending at volume to stale or unverified lists eats that margin quickly, and once your domain's reputation drops, your good emails land in spam along with the bad ones.

### Confident mistakes

Models get personalization wrong with total confidence: the wrong job title, a funding round that never happened, a tone your company would never use. A human SDR makes mistakes too, but one at a time. An autonomous agent can repeat the same mistake across a whole list before anyone looks. In a review queue, those mistakes are edits. In an autonomous sequence, they're sent, and every copy carries your company's name.

### A trust problem in the category

In March 2025, TechCrunch reported that 11x, an AI SDR startup backed by a16z and Benchmark, had listed ZoomInfo and Airtable as customers. Both told TechCrunch they weren't customers and hadn't approved the use of their logos. ZoomInfo added that in a one-month trial the product performed "significantly worse than our SDR employees," and a former 11x engineer was quoted saying "the products barely work." 11x responded that its retention rate was 79%.

I'm not judging one company from one story. I am saying that if you're evaluating a vendor, ask for references you can call and a trial on your own list before you trust a logo wall.

## How I build a lead engine (and why a person approves every message)

The engine does everything up to the send, and a person owns the send. It's one of [the AI systems I build for clients](/#what), and the pipeline runs in six steps:

1. **Collect** prospect data on companies and roles that match the ICP.
2. **Score** each lead with Claude-powered logic against that ICP.
3. **Draft** a LinkedIn message or an email written for that specific person.
4. **Review.** A person approves the draft, edits it or rejects it.
5. **Send** only what was approved.
6. **Track** the pipeline in HubSpot, from first contact to sales opportunity.

If you build your own, the review at step 4 should answer four questions:

- Is this the right person at the right company?
- Is every fact in the message true?
- Does it sound like us?
- Would I send this myself, under my own name?

The line sits at sending because of irreversibility. Everything before step 5 can be redone cheaply: rescore the list, rewrite the prompt, throw away a batch of drafts. A sent message is permanent, so a person makes that call. My rule: automate what you can undo, approve what you can't.

Voice is part of the review. I also build brand voice frameworks for clients' content, and an outreach message that sounds like a different company undoes that work in a single send.

The audience raises the stakes. One outbound program I ran went to CTOs and engineering leaders in Morocco, senior, technical buyers. With an audience like that, one careless message costs more than a few hundred extra sends could earn. If you sell to a narrow list of US accounts, the same math applies.

The approach has costs of its own. Approval is a bottleneck: an engine can draft faster than anyone can read carefully. And it only works if the reviewer actually reads. What worries me most is approval sliding into clicking "yes" all afternoon, which is autonomous sending with extra steps, and nothing in the tooling will flag it for you.

## Autonomous AI SDR vs. human-approved AI SDR vs. human SDR

Human-approved AI fits small, high-value account lists. Full autonomy only makes sense for very high-volume, low-ACV motions where you can afford to burn domains.

| | Fully autonomous AI SDR | AI drafts, person approves | Human-only SDR |
|---|---|---|---|
| Who sends | The software | A person, after review | A person |
| Volume | Very high | Medium, capped by review time | Low |
| Personalization accuracy | Unchecked; errors repeat at scale | Checked before every send | High, but limited by research time |
| Brand risk | High | Low | Low |
| Deliverability risk | High with stale lists | Lower; volume and targeting stay controlled | Low |
| LinkedIn ToS risk | High if it automates LinkedIn activity | Low if a person sends | None from tooling |
| Compliance responsibility | You | You | You |
| Setup effort | Low to start, high to clean up | Medium: scoring rules, prompts, review workflow | Hiring, training, ramp time |
| Best for | Very high volume, low ACV, disposable sending infrastructure | Small, high-value account lists | Complex deals and motions still being figured out |

Look at the compliance row. It says "you" in every column, and the next sections explain why.

## What the split means for reply rates, brand risk and cost

Human approval doesn't promise a higher reply rate. It caps the downside and moves your cost from writing to reviewing.

**Reply rates.** I haven't found a neutral benchmark. AiSDR publishes one built on 75 teams running its product: teams that had a defined ICP, real sending infrastructure, a named owner for replies and a baseline to measure against moved from a median 2.4% reply rate to 6.8%. Treat it as vendor data. The conditions are the part I'd believe: a tool amplifies the motion you already have. I'd also stop counting replies. An "unsubscribe me" is a reply. In HubSpot I track the pipeline from first contact to sales opportunity, because that's the number that pays for the tool.

**Brand risk.** With approval, the worst case is a slow week. With autonomous sending, the worst case is a burned domain, a restricted LinkedIn account or a screenshot of your message in someone's post.

**Cost.** Approval adds reviewer time to every message, but it removes most of the research and writing time that came before it. Whether that trade works for you depends on volume. At a few dozen high-value accounts a week, review is cheap compared to one lost deal. At tens of thousands of sends, review becomes the expensive part, and that's where full autonomy has a real argument.

## The US rules an AI SDR still has to follow

AI-written cold email is legal in the US if it follows CAN-SPAM, and the law doesn't care who or what wrote the message. LinkedIn automation is a separate problem, governed by LinkedIn's own rules.

**CAN-SPAM covers B2B.** The FTC is direct: "The law makes no exception for business-to-business email." Your emails need accurate header information, a subject line that reflects the content, a clear disclosure that the message is an ad, a valid physical postal address and a clear way to opt out. You must honor opt-outs within 10 business days. Each separate email in violation can cost up to $53,088.

**You can't outsource liability.** The FTC also says that even if you hire another company to handle your email, "you can't contract away your legal responsibility to comply with the law." If an AI SDR vendor sends on your behalf and breaks the rules, your company is still on the hook.

**Gmail's sender rules.** Every sender must keep the spam rate under 0.3% and authenticate with SPF or DKIM. If you send 5,000 or more messages a day to Gmail accounts, Google also requires SPF, DKIM and DMARC together, plus one-click unsubscribe on marketing messages. An autonomous tool sending at scale can cross that line, and then the stricter rules apply.

**LinkedIn bans automation.** LinkedIn prohibits bots and third-party software that scrape data or automate activity on the site, including sending messages, and it can restrict or shut down accounts that use them. This is the strongest single argument for the human-approved model: AI can research and draft, and a person sends the LinkedIn message from their own account.

This is a practical summary, not legal advice. Check with a lawyer before you scale outbound.

## When an AI SDR is worth it (and when it isn't)

An AI SDR is worth buying or building when it speeds up an outbound motion that already works. It isn't worth it as a substitute for one you don't have.

**Worth it if:**

- Your ICP is written down clearly enough that a model can score against it.
- Outbound already produces meetings when you do it by hand.
- Your data is clean, verified and from sources you're allowed to use.
- A named person owns review, with time blocked for it.

**Not worth it if:**

- You're hoping the tool will invent a sales motion you haven't proven.
- You expect it to run without anyone checking the output.
- Your plan depends on automating LinkedIn activity.

The fastest test I know: write down the name of the person who will read every draft next Monday. If you can't, you aren't ready to buy the tool yet. If you can, the build is mostly a matter of scoring rules, prompts and a review queue.

If you want help with that part, you can [design a human-in-the-loop lead engine with me](/#now).

Questions about your own setup? [Get in touch](/#contact).

## Questions

### Can an AI SDR replace a human SDR?

Not for cold outbound, in my view as someone who built one. It can take over most of the research, scoring and drafting, which lets one person cover more accounts in the same hours. When ZoomInfo trialed 11x for a month, it told TechCrunch the product performed "significantly worse than our SDR employees." The human job moves toward judgment, review and real conversations.

### Do AI SDRs hurt email deliverability?

They can, mostly when they send high volumes to unverified or stale lists. Gmail requires senders to keep their spam rate under 0.3%, recommends under 0.1%, and requires SPF, DKIM, DMARC and one-click unsubscribe for anyone sending 5,000 or more messages a day. A person approving each message keeps volume and targeting under control.

### Is it legal to send AI-written cold emails in the US?

Yes, if the emails comply with CAN-SPAM, which applies to business-to-business email too. You need accurate headers, a subject line that matches the content, a clear disclosure that the message is an ad, a valid postal address and an opt-out you honor within 10 business days. Who or what wrote the email doesn't matter, and you can't hand your liability to the vendor. Not legal advice.

### Can I automate LinkedIn outreach with an AI SDR?

LinkedIn prohibits bots and third-party software that automate activity on the site, including sending messages, and it can restrict or shut down accounts that use them. AI can still help with the research and the draft. The safe setup is a person reading the message, then sending it themselves from their own account.

### What reply rate should I expect from an AI SDR?

There's no neutral benchmark. AiSDR's own report, built on 75 teams, says teams with a defined ICP, real sending infrastructure, a named reply owner and a baseline moved from a median 2.4% reply rate to 6.8%. That's vendor data. I'd track sales opportunities rather than replies, since an unsubscribe request also counts as a reply.

### How much does an AI SDR cost compared with a human SDR?

Compare total cost, not the subscription. On the AI side: the tool, data, sending infrastructure such as extra domains, and reviewer time. On the human side: a fully loaded SDR seat. Vendor prices change often, so get quotes for your volume. Then compare cost per sales opportunity, not cost per email sent.


## Sources

- [TechCrunch: a16z- and Benchmark-backed 11x has been claiming customers it doesn't have (Mar 24, 2025)](https://techcrunch.com/2025/03/24/a16z-and-benchmark-backed-11x-has-been-claiming-customers-it-doesnt-have)
- [FTC: CAN-SPAM Act Compliance Guide for Business](https://www.ftc.gov/business-guidance/resources/can-spam-act-compliance-guide-business)
- [Google: Email sender guidelines](https://support.google.com/a/answer/81126)
- [LinkedIn Help: Prohibited software and extensions](https://www.linkedin.com/help/linkedin/answer/a1341387)
- [AiSDR: AI SDR Reply Rates & ROI in 2026 (vendor benchmark)](https://aisdr.com/blog/ai-sdr-benchmarks/)