What an AI SDR actually does in 2026 (and what it can't)

Thirty days into the trial the dashboard looks excellent. 4,812 emails sent, 214 contacts tagged as interested, 19 replies, a follow-up drafted for every one of them. Then you open the inbox. Four are out of office, six are a version of not right now, three ask something the sequence cannot answer, two are annoyed, and one is real. Nobody has read any of them since Tuesday, because the entire reason for buying the thing was that nobody would have to.

That gap between the dashboard and the inbox is where almost every AI SDR disappointment lives, and it is not a product defect. The category sells a job title and ships a set of tasks. The tasks are genuinely good. The job title covers four or five things the tasks do not, and the one you most need covered is the one you quietly stopped staffing the day the invoice went through.

What does an AI SDR actually do, job by job?

An AI SDR does four jobs, and every one of them is reading, writing or remembering at a volume a person cannot hold.

The first is research. Give it a company and it reads the website, the careers page, recent LinkedIn activity and whatever contact database it is wired into, then produces a few usable lines about that account. A person doing this properly spends five to fifteen minutes per company and gets noticeably worse by account 40. Software does it in seconds and is exactly as careful on account 400 as on account 4. That consistency is the whole of its advantage, and it is a real one.

The second is list building and enrichment. It takes a definition of who you want, pulls matching companies and people, finds and verifies email addresses, dedupes against what you already have, and fills in roles, headcount and tech. This is the step that decides everything downstream, and it is also the step most buyers never inspect.

Third is drafting and follow-up discipline. It writes the first message per contact, writes the follow-ups, holds the cadence across a few weeks, and stops when someone replies. Follow-up discipline is where it beats people by the widest margin, because a founder running outreach manually stops at touch two the week a project goes sideways.

Fourth is reply classification. Every reply gets sorted: out of office, unsubscribe, wrong person, clear yes, clear no. Apollo's own breakdown of what an AI SDR can run autonomously lists six tasks, and the first four are these same jobs under different names. The last two on their list, inbound response and meeting scheduling, are where vendor language and buyer experience start to part company, because "replies to inbound inquiries, qualifies leads, routes to the right rep" is a sentence about judgement wearing the clothes of a task.

The job	What it genuinely does well	What breaks when nobody checks
Account research	Reads site, postings and activity for 500 accounts without tiring	Treats a stale job post or an old rebrand as a live signal
List building and enrichment	Finds and verifies contacts, dedupes, fills role and firmographics	Builds a plausible list of people who cannot sign anything
First drafts and follow-ups	Writes per-contact variants, holds cadence, stops on reply	Averages toward the same email everyone else is sending
Reply classification	Sorts out of office, unsubscribes, clear yes, clear no	Files everything ambiguous as a soft no, which is where deals hide

Why does "AI SDR" mean three different products?

Three different products are sold under those two letters, and the gap between the cheapest and the most expensive is roughly fortyfold.

The first is a stack you assemble and run. Apify at $19 a month for the Starter tier handles scraping, Clay at $167 a month on Launch handles waterfall enrichment, Instantly at $47 a month or Smartlead at $39 handles mailboxes and sending, and HeyReach at $79 a month covers LinkedIn with unlimited senders. The AI part is a model doing research and drafting inside that chain. You own the domains, the lists and the reply history.

The second is a platform that bundles all of it behind one login. Jeeva publishes Growth at $95 a month and Scale at $239 for 7,200 emails and 900 cell phone lookups. AiSDR moved to usage pricing at $1.50 per contact prospected with campaigns starting at $7.50 and a warmed mailbox at $33 a month. 11x lists Alice Growth from $3,750 a month billed annually, which is $45,000 a year for up to 2,000 new prospects a month, managed Gmail mailboxes and two-way CRM sync, and it charges per lead rather than per send.

The third is a managed service wearing the same label. AiSDR sells a fully managed version at $2,500 a month or $149 per campaign, and most platforms now have a done-for-you tier. Here you are buying an operator, and the AI is the operator's tooling rather than the thing you bought.

Those are three different purchases with three different failure modes. The stack fails when nobody owns it. The platform fails when its database is the same database everyone else is mailing. The managed service fails the way every outsourced channel fails, which is that the asset is not yours when the contract ends.

What can an AI SDR not do?

It cannot decide who is worth writing to, read a reply that is neither a yes nor a no, or run the call.

Start with the first, because it is the expensive one. Who belongs on your list is a judgement about your offer, your delivery capacity and which buyer closes without a six-month cycle. No enrichment waterfall answers that. Give the software a vague definition and it will return thousands of contacts that match the definition perfectly, which is the failure that looks most like success. If your definition is still "B2B companies in Europe", fix that before you buy anything, and the method for turning a definition into a list is in our guide on building a lead list from buying signals.

The second gap is the lukewarm reply. "Interesting, we looked at this last year" is not a yes, a no, or any bucket a classifier has. It is a conversation, and it is where a large share of real pipeline is sitting. Software routes it; a person has to answer it, and answer it the same day.

Third is the conversation itself, and the buyer research now says this plainly. Gartner's survey of 645 B2B buyers found that 69 percent turn to a sales rep to validate AI-generated insights, even while 67 percent say they prefer a rep-free buying experience. They will do the research without a seller in the room and then ask a seller whether the research was right. An AI SDR handles the first part of that well and cannot do the second part at all.

Anthropic's engineering guidance on agent design lands in the same place from the technical side: agents should "pause for human feedback at checkpoints or when encountering blockers", and the advice is to maintain simplicity and add complexity only when it demonstrably improves outcomes. An outbound system with no checkpoint between a drafted reply and a prospect reading it is not a more advanced system. It is the same system with the review step deleted.

What does an AI SDR cost in 2026?

Between about $90 a month and $45,000 a year, and the spread is explained by how much of the work you are handing over rather than by capability.

Route	Published price, October 2026	What it covers	Pick it when
Stack you run	$90 to $810 a month	Scraping, enrichment, sending, LinkedIn, one inbox	Someone on the team will own it daily
Platform, small tier	Jeeva $95 to $239 a month	Lists, email, LinkedIn, CRM sync, one seat	Low volume and you want one login
Platform, usage based	AiSDR $1.50 per contact prospected	Research, multichannel sends, unified inbox	Volume swings month to month
Platform, full tier	11x Alice from $3,750 a month, annual	2,000 new prospects, managed mailboxes, CRM sync	There is a sales team and a CRM to feed
Managed service	AiSDR $2,500 a month, or $149 a campaign	Someone else operates the platform for you	Nobody internally will own it

Three stack shapes cover almost every founder-led business, and all of these are list prices you can check yourself.

Lean and email only runs about $90 to $130 a month: Instantly Growth at $47, Apify Starter at $19, and three mailboxes on Google Workspace Business Starter at EUR 6.80 per user per month after the introductory period. Standard, with real enrichment, runs $350 to $730: Smartlead Pro at $94 or Instantly Hypergrowth at $358, Apify Scale at $199, Clay Launch at $167. Multichannel adds HeyReach Growth at $79 for unlimited LinkedIn senders, landing between $540 and $810.

Set that against 11x's $45,000 a year and the comparison stops being about software. That figure is the band a junior hire occupies, and the hire-versus-build arithmetic, including loaded salary and ramp, is worked through in our guide on hiring an SDR or building the system first.

Why did the first wave of AI SDR companies struggle?

They sold autonomy, shipped volume, and the numbers underneath the growth did not hold.

In March 2025 TechCrunch reported that several companies listed as 11x customers disputed it. ZoomInfo said it ran a one-month trial and that the product "performed significantly worse than our SDR employees". Airtable said its trial "was never used in production and never rolled out to our sales team". One employee told TechCrunch the company was "losing 70-80% of customers that came through the door", and that contracts surviving the three-month break clause came to roughly $3 million against a claimed $14 million; 11x said its retention rate was 79 percent. The founder stepped down as CEO that May.

Investors had flagged the shape of the problem earlier. In December 2024 TechCrunch quoted Tomasz Tunguz of Theory Ventures describing a public company's CRO whose AI SDR generated a large number of leads across nine months and produced no sales, and Chris Farmer of SignalFire warning that without proprietary data these products get absorbed as features by the incumbents who already hold the customer data. By April 2025, a founder raising against the category told TechCrunch plainly that classic AI sales reps had "failed" because they optimized for pure volume rather than for accounts worth contacting.

None of that means the tooling does not work. It means the autonomy claim was the product, and the autonomy claim was wrong. Gartner now predicts AI agents will outnumber sellers 10 to 1 by 2028 while fewer than 40 percent of sellers will say agents improved their productivity. Ten times the agents, under four in ten sellers who feel better off. That spread is about who owns the output, and it will not close by itself.

How do you tell a real AI SDR from a sequencer with a wrapper?

Ask where the contacts come from, because the answer separates the two faster than any feature list.

1. Where does the list originate? A closed database that every competitor also queries produces contacts who are already being mailed by nine other vendors. Custom scrapes and intent signals produce contacts nobody else has this week.
2. Does it verify addresses before sending, and show you the bounce rate per campaign? If verification is not a visible step, your domain pays for it.
3. Can you read every draft before it sends, in bulk, and edit the ones that are wrong? A tool that cannot show you a queue is asking you to trust a black box with your reputation.
4. Where do replies land, and does LinkedIn land in the same place as email? Two inboxes means somebody is checking one of them.
5. Does a reply on either channel stop the sequence on both? This is basic and it is often missing.
6. What happens to the domains, the lists and the reply history when you cancel? If the answer is that they stay with the vendor, you are renting a pipeline.
7. Is the pricing per lead, per send, or per seat? 11x charges per lead; Instantly and Smartlead charge by sending volume. Those bill very differently once volume moves.

A tool that fails question 1 and question 6 is a sequencer with a research layer on top, which is a reasonable thing to buy for $47 a month and an unreasonable thing to buy for $3,750.

What does a setup that holds up at volume look like?

Five mechanical steps run on software and two decisions stay with a person, and the two decisions bracket the five steps.

The setup that holds up at volume builds lists from custom scrapes and intent signals rather than a shared database, runs waterfall enrichment with several email finders in sequence and then verification, sends from secondary domains on a specialist sender, works LinkedIn alongside email with every reply arriving in one inbox, and writes the outcome into whichever CRM the team actually opens. Alternatives are real: LinkedIn Sales Navigator or Apollo for plain firmographic filtering, Clay when you want enrichment in one tool rather than several, lemlist or La Growth Machine on LinkedIn, Make or n8n for the wiring. The full build order for that chain is in our guide on building an AI outbound system.

If you would rather not assemble it, Ourix builds the system inside the business and runs it: verified lists of the buyers you want, outreach on email and LinkedIn, every reply in one inbox, and a pipeline view of what got booked, on your own domains and accounts. Research, list building, first drafts and follow-up run on AI, and a person owns the judgement and every reply. Campaigns go live in about 30 days, everything is yours to take over at any time, and prices are not published: the build is a one-time fee scoped to your channels and your market, and you get the exact number on the call.

Before any of that, the offer has to survive a cold read, because software sends a weak offer faster. The diagnosis for that is in our guide on why an agency offer does not convert on cold outreach.

How do you know whether it is working?

Count positive replies and meetings held, and ignore the sends entirely.

The arithmetic is unforgiving and it is published. lemlist's 2026 benchmark compilation puts the average B2B cold email reply rate at 3.43 percent against a 27.7 percent open rate, with meetings booked landing at 0.5 to 2.5 percent of contacts depending on deal type and follow-up quality. The same data has 58 percent of replies arriving from the first email and 42 percent from follow-ups, which is why cadence discipline is worth paying for. It also shows campaigns under 50 recipients averaging 5.8 percent replies against 2.1 percent for large sends, which is the clearest argument against letting a tool run at its maximum.

Two numbers decide whether to keep paying. Positive replies as a share of all replies tells you whether the list and the offer are right; under about 15 percent and the problem is upstream of the tool. Meetings held, not booked, tells you whether anyone is working the inbox.

Watch deliverability on the same cadence, because an AI layer reaches volume faster than a person ever would. Google's sender guidelines tell anyone sending more than 5,000 messages a day to keep spam rates in Postmaster Tools "below 0.30%", with a recommendation to stay under 0.10 percent, and require SPF, DKIM and DMARC on the sending domain. Sizing sends against mailbox count so you never approach that line is covered in our guide on how many domains and inboxes cold email needs.

Where is an AI SDR the wrong answer?

Four situations, and in all four the money is better spent elsewhere.

Enterprise buyers with long committee cycles are the first. A sequence cannot work seven stakeholders over eleven months, and the channel there is a named person's relationships. Low-ticket offers are the second: at a $500 to $1,500 monthly deal the cost per meeting rarely clears, whatever the tool costs. Third is a business still finding product-market fit, where every month changes who the buyer is and outbound just industrializes the wrong guess. Fourth, and most common, is having nobody to take the calls. If the founder is in delivery nine hours a day, meetings will land on a calendar that cannot host them, and the system will look broken when the constraint is capacity.

There is also a timing answer. If meetings are needed inside two weeks, warmup alone rules this out, and paid traffic or the phone is the honest recommendation.

What to do next

1. Write the four jobs on one page: research, list and enrichment, drafting and follow-up, reply handling. Name the owner of each. If reply handling has no name next to it, stop here and fix that first.
2. Define the buyer tightly enough that a scrape can find them, using one or two live signals rather than a firmographic filter.
3. Decide the response-time rule before you buy anything. Same day for anything positive, and a named person behind it.
4. Buy the smallest version that tests the chain. A sender at $39 to $47 a month, a scraper at $19, and 500 contacts is enough to learn whether the offer lands.
5. Run 30 days, then read positive replies as a share of replies and meetings held. Never the send count.
6. If replies are healthy and meetings are not, the gap is reply handling, and no upgrade fixes it.
7. If you would rather have the chain built and run for you than assemble it, take the free growth audit and go through your own numbers first, whoever ends up building it.