AI SDRs: what they are good at, what they are not, and where the person still wins.

An AI SDR is software that finds prospects, researches them, writes and sends outreach and, in some products, handles the first reply, at a volume no person could match. It is good at the part of the job that was always a spreadsheet, and poor at the part that was always a conversation. The mistake buyers make in 2026 is not choosing the wrong tool. It is believing the two parts are the same job. This is an honest account of both, written by a company that runs the AI in its platform and the people on its phones.

What the software is good at

1. Research, at a scale nobody could do by hand.

   Reading a company’s website, news, hiring pages and posts and turning them into a reason to call, for every company on the board, before anyone starts work. A rep doing this by hand manages fifteen accounts a morning. The software does the board overnight.

2. Enrichment.

   Finding the right person, verifying the email, finding and checking the mobile number, and doing it again when the data goes stale. Waterfall enrichment across several providers is a solved problem, and a person should never do it.

3. Timing.

   Ordering the day by who is likely to pick up, sending the follow-up on the day it is due, moving an account up the queue when a signal fires. The fourth and fifth touches that a busy person forgets go out on time.

4. First drafts.

   An opening line built from the signal, for a person to edit. The draft is often close. It is rarely sendable as it is, and the tools that send it as it is are the ones whose email the market has learned to recognise.

5. Reporting.

   Dials, conversations, replies, meetings booked and held, per week, without anyone filling in a sheet on a Friday.

What it is not good at

1. The phone.

   A synthetic voice on a cold call is a call that ends. Prospects can tell, and a meeting is agreed in a conversation between two people, not in a thread.

2. The reply.

   "Not now, we just hired someone" is a conversation to a person and a closed ticket to a script. The reply is where the meeting is won, and it is the part the software most often gets wrong while sounding fluent.

3. Qualification.

   Deciding that this person, at this company, with this problem, is worth your closer’s hour is a judgement that needs to have been wrong before. Rules do not learn from a wasted meeting.

4. Deliverability.

   Automated email at volume is how a domain ends up in spam, and the damage is to your name, not the tool’s. The reply rates in the marketing material are earned early, before volume has burned the domain, and fall as it does.

5. Being believed.

   Buyers have learned what machine-written outreach reads like: the compliment on a recent post, the question that could be asked of anyone, the sign-off with a calendar link. Many delete it unread, and some hold it against the sender.

The test that settles it

Ask any AI SDR vendor, and any outsourced provider including us, for meetings held per month with a person who fits criteria agreed in writing. Not replies, not "positive responses", not meetings booked. Held. A tool that sends ten thousand emails and books twelve meetings of which six happen is a tool with a 0.06 per cent conversion, and ten thousand people have had your name in their inbox to get there. That is the arithmetic the reply-rate slide leaves out.

Where the line sits

The engagements that work in 2026 use both and are clear about which does which. Our own split is simple to state. Everything that happens before a conversation - the signals, the enrichment, the overnight research, the order of the queue, the timing of the follow-ups, the numbers on Friday - belongs to the software, Setta. Everything that happens in a conversation - the call, the reply, the judgement about whether this is a meeting, the handover to your closer - belongs to a rep, with a manager listening and correcting daily. The rep never opens a spreadsheet in the morning, and the prospect never talks to a machine.

The line moves, and it is worth being honest about the direction. Two years ago the draft was not close; now it often is. The reply is next, and for low-value, high-volume propositions with a self-serve product behind them, an AI handling the first reply is already a fair trade. For a considered sale to a senior buyer, it is not, and the deal size decides which you are.

When an AI SDR tool is the right answer

• A high-volume, low-value proposition with a self-serve or product-led sale behind it, where a human meeting is not the goal.
• Responding to inbound at speed, where the prospect has already raised a hand and the job is to reply in minutes rather than hours.
• A market so large that the board cannot be worked by people at all, and the trade of quality for reach is deliberate.
• As the platform underneath a person, which is the version we run.

When it is not

• A considered sale to a senior buyer, where the first conversation has to be specific to their systems and their problem, and the cost of a wrong first impression is the account. For LGN.AI, selling edge AI into mining and manufacturing, sequences ran well into double digits per account before a meeting was asked for; the value of each conversation ruled volume out from the start.
• A regulated or specialised market, where a rep who cannot hold the conversation does not get a second one.
• Anywhere the meeting is agreed in a conversation, which in UK B2B is most places we have worked.
• Anywhere your domain reputation is worth more than the reach.

How to evaluate one

1. Ask for held meetings, not replies.

   Per month, per client, with the qualifying criteria. If the answer is a reply rate, the product is a mail merge.

2. Ask whose domain sends.

   Yours, warmed, with SPF, DKIM and DMARC in place, at a volume that will not burn it; or theirs, which is a shared reputation you do not control.

3. Read ten real emails it sent.

   Not the examples in the deck. If you could tell they were machine-written, so could the buyer.

4. Ask what happens to a reply.

   Who reads it, how fast, and what the software does when the answer is "not now, try Sarah".

5. Ask who improves it.

   A manager listening to the output daily, or nobody. The tool does not coach itself.

The two side by side, with the shape of the cost, are on AI SDR or human SDR. An account-based programme where the deal size ruled out volume from the start is the LGN.AI case study.