AI sales agent: what an AI SDR can and cannot do

An AI sales agent can research, write first drafts and qualify inbound leads. See which SDR tasks to automate, which to review, and the EU rules that apply.

Published
Reading
8 min
By
Authoricy

An AI sales agent can do the repetitive half of a sales rep’s job: research, follow-up, inbound replies, qualifying and booking, plus outreach drafts for a person to approve. It needs a person for the judgement half, such as pricing, objections and which accounts matter, and should not own a customer relationship. An AI SDR is the same software aimed at the sales development role.

The useful question is which of your sales tasks to hand over, and under which rules. If most of your leads arrive by phone, the same thinking applies to AI voice agents that answer, qualify and book calls.

AI sales agent, AI SDR or copilot

Vendors use these words loosely. They describe three different set-ups, and each fails in its own way.

Sales copilot AI SDR Voice agent
What it does Suggests replies, summarises calls and fills in the CRM for a human rep Finds prospects, writes outreach, follows up and books meetings with little human input Answers or places phone calls, qualifies the caller and books a time
Who is accountable The rep, for every message Whoever approved the list and the copy Whoever wrote the script and the hand-off rules
Typical failure Ignored by the team Sends confident nonsense to hundreds of people Mishears a name or a number

Most companies asking for an “AI SDR” need one narrow job done fast, such as replying to inbound leads within minutes. That is a smaller and safer project than an agent that prospects on its own.

The SDR job, task by task

This split is our judgement from building agents, not a vendor benchmark.

Task Let the agent do it Agent drafts, a person approves Keep with a person
Account research and enrichment Yes
Building a prospect list Yes: someone must check fit and lawful basis
First-touch email Yes, at least until the copy is proven
Follow-up on a sequence Yes, inside fixed rules
Answering inbound questions Yes, from approved facts only
Qualifying a lead Yes, on criteria you write down
Booking the meeting Yes
CRM updates and call notes Yes
Replying to an objection or a complaint Yes
Pricing, discounts and terms Yes
Deciding a large account is worth pursuing Yes

The pattern is simple. Tasks with a right answer suit the agent. Tasks where a wrong answer costs a relationship stay with a person.

What an AI SDR cannot do, and why

Judge fit from thin evidence. A human rep notices that a company just lost its sales director and is a bad bet this quarter. An agent reads the same page and sees a hiring opportunity.

Hold a relationship. Buyers reply to people they trust. An agent cannot be the person the buyer calls when something goes wrong.

Negotiate. Price, scope and terms depend on what your company will accept on a given day. Give an agent room there and it will promise things you cannot deliver.

Stay accurate at volume. A small error rate becomes many errors at scale. If the agent gets one fact wrong in every fifty messages, a run of 1,000 emails carries about twenty mistakes.

Carry accountability. When a message breaks a rule, the sender is responsible, not the software.

Why AI SDR roll-outs fail

Most failures come from what surrounds the model, not the model itself.

  • A bad list. The agent writes well to the wrong people.
  • Generic copy at scale. Personalisation that names the prospect’s city and job title fools nobody.
  • A burnt domain. Sending high volume from your main domain can push your ordinary email into spam folders. Use a separate, properly authenticated sending domain and keep volumes modest.
  • No hand-off rule. Nobody decided when a person takes over, so hot leads wait in the agent’s queue.
  • No reviewer. Output nobody reads drifts. Name one person who reads a sample every week.

The rules in Europe and Sweden

A planning summary, not legal advice.

GDPR. Prospecting named people means processing personal data. Direct marketing can rest on legitimate interest, but people may object at any time, and once they do, the data can no longer be used for marketing (GDPR Article 21). If the contact came from a list provider rather than the person, you must tell them where you got it, at the latest in your first message (GDPR Article 14). An AI SDR must therefore log the source of every contact and stop for good on a single “no”.

Email in Sweden. The Marketing Act implements the EU ePrivacy rule on unsolicited email. Marketing email to a natural person needs prior consent, with a narrow exception for your own customers buying similar products. Every marketing email, including one sent to a company, must carry a valid address for opting out (Marknadsföringslag 2008:486, 19 and 20 §§). A sole trader is a natural person, so check before you send cold email to one.

AI disclosure. The EU AI Act requires AI systems that talk to people to make clear they are an AI. The exception is when that is obvious (AI Act Article 50). The EU has amended parts of the Act’s timetable, so check the current date of application with your lawyer. An agent that replies to a prospect in a chat or by phone should say so.

Calls and LinkedIn. The same 19 § requires prior consent before marketing to a natural person through an automated calling system. An AI agent that places calls on its own may count as one. Phone outreach also has NIX rules, set out in our guide to AI outbound calling and where it is allowed. For LinkedIn, read the platform’s terms before you automate anything there.

A worked example: one inbound lead

A company that installs heat pumps for businesses gets a web form at 21:40. It reads: “Quote for our warehouse, about 2,000 m².” The agent works from a fact sheet the sales lead approved.

  1. 21:41, email. “Thanks, Anna. I’m the AI assistant at the company and can get you to the right engineer. Is the warehouse heated today, and roughly when would you want the work done?”
  2. 21:58, Anna replies. Oil heating, wants it done before next winter, asks for a ballpark price.
  3. 21:59, agent. It does not quote. “Price depends on a site visit. I can book one with our engineer. Does Tuesday at 10 or Thursday at 14 work?”
  4. 22:03, Anna picks Thursday. The agent books it, sends a calendar invite, and writes the lead into the CRM with size, current heating and deadline.
  5. Next morning, hand-off. The engineer sees a qualified lead with notes. A person owns it from here.

The agent answered at once, qualified the lead and booked the meeting. It refused the one question that needed judgement. The same flow by phone is what an AI receptionist for a small business handles.

What drives the cost, and a checklist

Cost comes from four places: the tool or build, contact data, sending infrastructure, and review time.

Do the arithmetic in hours first. Say your SDR spends 15 hours a week on research, sequences and CRM notes, and 10 on calls and replies. An agent can take most of the first block. It cannot take the second. If a reviewer then spends 3 hours a week checking output, you have freed about 12 hours. You have not replaced a person.

Off-the-shelf AI SDR tools suit a standard sequence on a standard CRM. A build makes sense when the agent must use your own systems, rules or data, as with most AI agents for business. If your CRM is the bottleneck, weigh HubSpot’s free CRM against a custom CRM first. Our projects open with a fixed-price discovery, and it ends in a fixed price for the build.

Before you switch one on, answer yes to each:

  1. Do we know where every contact came from, and our lawful basis for using it?
  2. Is sending set up on a separate, authenticated domain?
  3. Does every message offer a working opt-out, and does one “no” remove the person everywhere?
  4. Does the agent say it is an AI when it talks to people?
  5. Is there a written rule for when a person takes over?
  6. Is a named person reviewing output every week?
  7. Have we picked one metric for the first 30 days, such as meetings booked from inbound leads?

If the narrow job is answering and qualifying calls, we build custom AI voice agents on your calendar, your CRM and your rules, and you own the code.

Frequently asked questions

Can an AI SDR replace a human SDR?

Not fully. It can take over research, follow-ups, inbound replies, booking and CRM notes. Objections, pricing, account judgement and the relationship still need a person, and someone must review what the agent sends.

Do AI SDRs actually work?

They work on narrow, rule-bound jobs with a clean list and a reviewer, such as replying to inbound leads at once. They fail when they are given a poor list and told to prospect at volume with no hand-off rule.

Who wrote the email does not change the rules. In Sweden, marketing email to a natural person needs prior consent, emails to companies must offer an opt-out, and GDPR gives every named recipient the right to object. Where the agent interacts directly with a person, the EU AI Act requires that the person knows it is an AI.

Can an AI sales agent make phone calls?

Yes. A voice agent can call back a web lead, qualify them and book a meeting. Automated marketing calls to private persons can need their prior consent under the Swedish Marketing Act. The agent should say it is an AI at the start of the call.

Talk to the people who build it

Hear it answer your calls.

Book a call and we build a demo agent on your own questions before you decide.