Most explanations of AI SDRs describe the output and skip the machine. You are told it "researches your prospects and writes personalised emails at scale", which is true and tells you nothing about why one deployment produces meetings and another produces a burned domain.
So here is the machine. Seven stages, in order, with what happens at each and what breaks.
The pipeline at a glance
| # | Stage | What it produces | Who does the work |
|---|---|---|---|
| 1 | Sourcing | A list of businesses matching your criteria | Data, not the model |
| 2 | Enrichment | Website, email, phone, socials, context | Data, not the model |
| 3 | Verification | Contactable or not, permanently marked | Data, not the model |
| 4 | Scoring | Priority order for finite sending capacity | Rules, sometimes the model |
| 5 | Drafting | The message a human will actually read | The model |
| 6 | Sending | Delivery from a mailbox with a reputation | Infrastructure |
| 7 | Reply handling | Detection, drafting, and stopping the sequence | The model, with a human |
Read the fourth column. The language model appears twice out of seven. This is the single most useful thing to understand about the category: an AI SDR is mostly not an AI product. It is a data pipeline with a writing stage in the middle, and evaluating one on its writing is like evaluating a restaurant on the plates.
Stage 1: sourcing
The pipeline starts with a description of who you sell to, and it has to turn that into real companies.
In practice this means querying live sources: map and directory data for local businesses, company databases for larger B2B, professional networks for people rather than firms. A good sourcing stage returns businesses that exist right now, with an address you could drive to. A weak one returns rows from a database that was scraped eighteen months ago, which is where a meaningful share of "this contact left the company" replies come from.
Two questions settle the quality of any sourcing stage:
- Is the data live or is it a snapshot? Snapshots decay at roughly the rate businesses change, which is faster than anyone expects.
- Can you constrain it the way you actually sell? "Roofing companies in Denver" is a real constraint. "Construction, North America" is a wish.
If a tool's story begins at "upload your list", stage one is still your job, and everything downstream inherits whatever you uploaded.
Stage 2: enrichment
Now each business needs to become contactable and understandable.
Contactable means an email address, a phone number, sometimes a WhatsApp-capable number. Understandable means the raw material for personalisation: what the company's site says it does, which services it lists, where it operates, whether it has social profiles, how big it looks.
This stage is where personalisation is actually won or lost, which surprises people who assume it is a writing problem. A model given three facts about a company will write three-facts-worth of specificity, no matter how good it is. A model given the company's own service list, city and recent site copy will write something that could not be sent to anyone else. Same model. The difference was upstream.
Stage 3: verification
Then the pipeline decides which of those contact details are real.
Email verification checks that the mailbox exists and will accept mail, without sending anything. Phone validation checks the number is in service and, for WhatsApp, that the number has an account attached. Both produce a verdict that has to be respected permanently: a contact marked invalid should never be attempted again by anything downstream.
This stage exists for one reason. Bounces are charged to your sender reputation, not to the list. A campaign with a high bounce rate does not just fail; it makes the next campaign fail too, from a mailbox that used to work. Skipping verification to save a step is the most expensive shortcut in the whole pipeline, and its cost arrives weeks later, disguised as a messaging problem.
We wrote about the mechanics of this in the email bounce rate guide, because it is the failure that most often gets misdiagnosed.
Stage 4: scoring
You cannot contact everyone today. Sending capacity is finite and deliberately limited, so something has to decide the order.
Scoring assigns each record a priority from what is known about it: how complete the contact data is, whether the email verified, whether the business matches the sharper end of your criteria, whether there are signals that it needs what you sell. A business with a verified email, a live website and no obvious online presence beyond it is a different prospect from a national chain with a marketing department, and they should not sit at the same place in the queue.
This is the stage most often skipped entirely, and skipping it is defensible early on. It becomes indefensible once your list is larger than your capacity, which happens quickly. A lead scoring framework does not need to be clever to beat alphabetical order.
Stage 5: drafting
Only now does the model write anything.
It gets three inputs: what it knows about the prospect from stage two, what it knows about your business, and the shape of the message it is being asked to produce. The second of those is the one teams underinvest in. An agent that has been told your value proposition, your tone, your pricing posture and the objections you actually get will write mail that sounds like your company. An agent given a product name and a link will write mail that sounds like every other agent, because that is genuinely all it has.
The output is a first-touch message and usually a set of follow-ups with different angles, so that touch three is not touch one with "just bumping this" on top.
Two rules that hold across every deployment worth copying:
- Specificity beats fluency. A slightly clumsy sentence that proves you looked at their website outperforms a polished one that could have been sent to anybody.
- The ask should be proportionate to the relationship, which at first touch is none. Asking for a reply is proportionate. Asking for a thirty-minute call is not.
Stage 6: sending
The message now has to arrive.
This stage has nothing to do with AI and decides more outcomes than the writing does. It covers which mailbox sends, how many messages that mailbox has sent today, whether it has been warmed, whether the domain has SPF, DKIM and DMARC configured correctly, and what happens when a mailbox starts failing.
The parts that matter:
- Mailbox rotation. Volume spread across several mailboxes looks like several people working. The same volume from one mailbox looks like software.
- Daily caps that respect mailbox age. A new mailbox sending at full volume on day one is a spam signal, not an efficiency gain. The honest numbers are lower than most people expect.
- Warm-up. Real conversations, at low volume, before real campaigns, so the mailbox has a history of being replied to.
- Failure handling. When a mailbox starts bouncing or gets suspended, the pipeline should reroute to a healthy one and tell you, not keep sending into a wall.
An AI SDR without this layer underneath is a machine for reaching spam folders faster than you could manually.
Stage 7: reply handling
Finally, the pipeline watches for what came back.
Detection is the mechanical part: reading the mailbox, matching a reply to the thread that produced it, and above all stopping the sequence. Nothing damages a promising conversation like follow-up four arriving after the prospect already answered. This is the single highest-value automation in the whole stage and it involves no intelligence at all.
Drafting is the part people mean when they say the AI "handles replies". It reads the reply, understands what was asked, and writes an answer for you to approve.
The question is whether it sends that answer without you. Our position, and it is a position rather than a fact, is that draft-and-approve is the right default and full autonomy is a decision to make deliberately per sequence. The reasoning is asymmetric risk: an autonomous agent handles the ninety easy replies perfectly and answers the tenth, the ambiguous one from the prospect who is genuinely interested but has a specific objection, with total confidence and no idea it has just cost you the deal.
Which stage fails first
When a deployment disappoints, the diagnosis is almost always upstream of where people look. Symptoms map to stages fairly reliably:
| Symptom | Where people look | Where it usually is |
|---|---|---|
| Low reply rate | Stage 5, the copy | Stage 1, the list |
| Replies say "wrong person" | Stage 5, the copy | Stage 2, enrichment |
| High bounce rate | The data provider | Stage 3, skipped verification |
| Nothing arrives at all | Stage 5, spam words | Stage 6, authentication or warm-up |
| Good conversations go cold | The prospect | Stage 7, sequence did not stop |
The pattern is consistent enough to be a rule: if replies are not coming, look at who you are writing to before you look at what you wrote. Better writing aimed at the wrong list changes nothing, and it is the more satisfying thing to fix, which is exactly why it absorbs the time.
How this maps to a real product
Leads Ranger runs these seven stages as one system, which is the reason it exists: the stages are individually unremarkable and collectively hard to keep in step. Discovery sources live businesses from map and directory data, enrichment fills in the contact and context, verification marks what is reachable, heat scoring orders the queue, AI agents draft from a knowledge base you train once, mailbox rotation and warm-up handle sending, and replies land in one inbox with the sequence stopped automatically.
You do not need our product to run this pipeline. You do need every stage of it, from somewhere, and the ones that are not the model are the ones that decide the result.
Where to go next
- What is an AI SDR? for the definition and the boundaries of the job.
- Do AI SDRs actually work? for the results question, with the caveats attached.
- AI sales agents vs human SDRs for how the two divide the work in practice.