What an AI SDR cannot do

An AI SDR cannot fix a weak offer or a poor list, close a complex deal, know a fact you never gave it, promise that LinkedIn will not restrict an account, confirm that a booked meeting took place, replace the person who takes over at the key moment, or judge strategy. In our own data the agents booked 351 showcase meetings from 10,978 conversations, and only 33 were confirmed held. It does the repetitive work of prospecting well. The rest is yours.

We make an AI SDR, so this is an odd post to write, and that is why it is worth writing. Buyers are tired of claims. The list below is what we tell customers on a call, with the evidence from our own account and from the case studies on this site.

1. It cannot fix a weak offer

An agent amplifies what you give it. If the offer does not interest the audience, the agent finds that out faster and at a larger scale. In our own LinkedIn data, reply rates held near 40% for twelve months while meetings per person messaged peaked at 2.4% in November 2025 and drifted below 1% from June 2026 (case study). The agents kept replying in minutes. The list mix and the message style changed. Speed was never the problem.

The same point is made in replies are easy, conversion is audience and offer. An agent can get you conversations. It cannot make people want what you sell.

2. It cannot close a complex deal

It can book a call, answer the early questions and hand over with context. A negotiation, a custom scope or a contract needs a person with authority and judgement.

In the Hotel Pirin case study, the reservations agent answered guests in a median of 41 seconds and quoted from live availability. It created no reservation. Staff, the owner or the reply desk entered every booking, and the owner negotiated and closed every group deal. That is a good result and an accurate description of the division of labour.

3. It cannot know a fact you did not give it

A model knows the world in general and your business only through what you tell it. If your price list, terms, availability or exceptions are not in its knowledge base or its rules, it will either ask, stop, or, in the worst case, guess. Sellinger's answer is to keep facts in a knowledge base, run campaigns in approval mode where every message waits for a person, and let the agent stop on a question it cannot answer, such as a price it has no answer for.

The reverse also holds. When we told the hotel's agent to quote as soon as it had dates and party size, the share of replies carrying a price went from 50% to 92%. Rules, not model power, changed the behaviour.

4. It cannot promise your account is safe

LinkedIn's rules do not permit third party software that automates activity on the site, and LinkedIn may restrict accounts it suspects of it (LinkedIn Help). Every automation tool carries that risk, ours included. Sellinger sets per account daily limits (20 invitations a day and 150 a week by default), a quiet window and a pause when LinkedIn refuses. Those reduce risk. They do not remove it. Read is LinkedIn automation safe before you decide.

5. It cannot tell you a meeting happened

This is the one that surprises people. Of 351 showcase meetings booked through our agents, 33 are confirmed held, meaning a stored recording exists or our team marked them. The rest may well have taken place, but the calendar marks a meeting "completed" when its time slot ends whether or not anyone joined, so we do not count those.

The lesson applies to anyone buying a tool: ask how "meetings" are counted. Count meetings honestly gives a ladder from booked to held to qualified to won.

6. It cannot replace the person at the key moment

An agent is fast, and a human is accountable. When a prospect asks for a price outside the rules, a custom scope, a legal question, or simply a call with a person, the right move is a handoff. The risk is a handoff that nobody sees. In the hotel case, after a guest was handed to staff the wait ran from three hours to 14 days, and staff answered group requests in a median of almost four hours. A handoff needs an alert, an owner and a clock. See speed is table stakes, the handoff is the bottleneck.

7. It cannot judge strategy

Which market, which offer, which price, which month to push. An agent can test an angle you propose and report on it, and an optimizer can propose changes to a campaign. Choosing the direction of the company is not on that list. Used well, the agent gives you faster feedback on your strategy, so you can change it sooner.

8. It cannot be graded on its word

Every vendor, including us, says their agent works. The only honest basis is a method you can check: a defined population, a counted outcome and a period. The post on whether AI SDRs actually work sets out what to ask for. For our part, we publish what did not work in our case study: a falling meeting rate, a signal source that produced almost nothing, and openers that never mentioned the post that put the person on the list.

Which objections come up most often?

"Then why buy one?" Because the repetitive work is most of the hours. A person can write ten thoughtful messages a day. The agent writes hundreds, each with a reference to the person, and answers within minutes at any hour. What you buy is coverage and speed, and the judgement stays with you.

"Will prospects know it is an AI?" Do not build a plan that depends on them not knowing. If a prospect asks, the honest answer is the right one, and the conversations that work in our data are the ones that are useful, not the ones that pretend.

"What if it says something wrong?" It can, as any model can. That is why approval mode, a knowledge base of facts and a stop on unanswerable questions exist. Start with approval on, read the first batch, and loosen only when the pattern holds.

"What if my market is tiny?" Then volume is not your constraint. Use the agent for research and drafting, keep approval on, and send fewer, better messages.

What does it do well, then?

Plainly: the repetitive work that has to happen every day. Finding people from a signal or a list, researching each one, writing a first message that refers to something true about them, following up on a schedule, answering the first replies in minutes, stopping on an opt out, and booking a call. In our data that produced 27,737 conversations opened and 10,978 replies in a year, from a small team. The what is an AI SDR post covers the work in order.

How should you use this list?

Use it as a buying checklist. For any AI SDR, ask:

• How are meetings counted, and how many were confirmed held?
• Where do facts live, and what happens when the agent does not know?
• Can a person approve every message, and take over a conversation?
• What does the vendor say about platform restrictions, and do they promise anything absolute?
• Who owns the handoff?

A vendor who answers all five without drama is worth a trial.