"AI SDR" is a term doing a lot of work in a lot of marketing, so it is worth pinning down what is actually underneath it.
An AI SDR is software that performs the repetitive portion of sales development: researching prospects, writing first-touch messages tailored to each one, scheduling follow-ups, and drafting replies when someone responds. The name borrows from Sales Development Representative, a real job, and that borrowing is where most of the confusion comes from, because the software does perhaps sixty percent of that job and none of the sixty percent that people find hard.
The job, split into parts
Here is a real SDR's week, and which parts of it software can genuinely take.
| Task | Share of the week | Can AI do it? |
|---|---|---|
| Building and cleaning lists | High | Yes, better and faster |
| Researching each prospect | High | Yes, within the limits of public context |
| Writing first-touch messages | High | Yes, if trained properly |
| Sending and scheduling follow-ups | Medium | Yes, and more reliably than a person |
| Drafting replies | Medium | Yes, with review |
| Qualification calls | Medium | No |
| Handling objections live | Low | No |
| Deciding who is worth chasing | Low | Partly, with a scoring model |
| Building rapport over months | Low | No |
Read the right-hand column and the honest summary appears: AI takes the hours, humans keep the judgement. That is a genuinely large win, because the hours are most of the cost and none of the value.
What "personalised" means in practice
This is where the category earns its scepticism, so let us be specific.
Weak personalisation inserts variables into a template. Hi {{first_name}}, I saw {{company}} is doing great things in {{industry}}. Everyone has received this. Everyone deletes it. Adding an AI model to generate the same sentence with more adjectives does not fix it.
Real personalisation starts from something true about that specific business that nobody else would mention: the booking system they do not have, the review that complains about response times, the second location they opened last month, the role they are hiring for. That requires context, and context is why an AI SDR is only as good as what you feed it.
The two inputs that matter:
- What you sell, in your own words: your services, your pricing logic, how you talk, the objections you get, the proof you use. Without this the agent writes competent, generic marketing copy.
- What you know about them, which comes from your discovery data rather than from the model. An agent connected to live business data has something to work with. An agent handed a name and an email address is guessing.
This is why training an agent on a business brain once, properly, does more for output quality than any amount of prompt tinkering afterwards.
The autonomy question
Every AI SDR sits somewhere on a spectrum, and where it sits matters more than any feature list.
Full autopilot researches, writes, sends and replies with no human in the loop. It is the version that demos best and the version that occasionally answers a pricing question wrong, agrees to something you would not have agreed to, or replies with confident nonsense to a question about a product edge case.
Draft and approve does everything up to the send, then waits. You see the draft in your inbox, you change a word or you do not, you click send. It costs a few seconds per reply and it keeps human judgement exactly where the stakes are highest.
For first-touch messages at volume, autopilot is fine: the downside of a mediocre cold email is a mediocre cold email. For replies to real humans in a live conversation, draft-and-approve is the correct default, and any vendor telling you otherwise is selling the demo rather than the outcome.
Where AI SDRs quietly fail
Context poverty. The model can only work with what it is given. If your data is a name and a company, the output will read like it was written from a name and a company.
Volume without infrastructure. An AI SDR removes the effort ceiling on sending, which is exactly the ceiling that was protecting your domain. Without warm-up, verification and sending limits underneath, a good agent will help you reach spam faster than you ever could manually.
Sequences that do not listen. The classic failure: a prospect replies with genuine interest and receives follow-up number three the next morning because the sequence never stopped. This is not an AI problem, it is a plumbing problem, and it is embarrassing in direct proportion to how good the first message was.
Uniform tone across every segment. One trained voice is a strength until you sell to two very different audiences. Then you need two agents, not one with a longer prompt.
How to evaluate one in twenty minutes
Ignore the demo. Do this instead.
- Give it five of your real leads. Not sample data, yours, including the awkward ones.
- Read the five messages it produces. Ask one question of each: could this have been sent to any other business in this industry? If yes, the personalisation is cosmetic.
- Check what happens on reply. Does the sequence stop? Does the draft appear where you actually work, or in a separate tool you will forget to open?
- Ask what protects your domain. If the answer does not include warm-up, verification and volume limits, the agent is a liability rather than an asset.
- Ask what it costs at your real volume, including everything it depends on. Standalone AI SDR pricing frequently exceeds the platform it sits on top of.
The realistic verdict
An AI SDR will not close your deals and it will not replace a good salesperson. What it will do, reliably, is make sure that the eleven prospects you meant to follow up with last Thursday actually got followed up with, in your voice, with something specific to say.
That sounds modest. It is not. Most outbound pipeline is lost to inconsistency rather than incompetence, and inconsistency is precisely the thing software is good at eliminating.