How to Set Up an AI Sales Agent: Skills, Limits and What to Approve

An AI sales agent is easy to describe and easy to misjudge. The pitch is usually a tireless rep that finds leads, writes to them, handles replies and books meetings while you sleep. The reality, for any team that wants to keep its reputation, is more modest and more useful: a rep that does the repetitive parts of outbound in small, reviewable steps, and hands over to a person the moment a conversation needs judgment.

This guide covers how to set one up sensibly: the job description, the sources it may work from, which skills to allow and approve, guardrails, a rollout plan and how to judge results. It does not promise results, because those depend on your market, offer and list.

Key takeaways

What an AI SDR agent is, and what it is not

In a human sales team, a sales development rep sits at the front of the funnel. They take a list of possible customers, research each one, decide whether it is a fit, write something relevant, follow up a sensible number of times, read the replies, and pass genuinely interested people to a salesperson who runs the meeting. An AI SDR agent is that same job, split into steps a piece of software can perform and you can inspect.

The useful mental model is a junior colleague with a very narrow role and a very short attention span. It works through one prospect at a time. For each one it can research, score fit, draft a message, send it, wait, classify whatever comes back, and either follow up or stop. It does not carry a grand plan in its head. Each step is small, and each step can be looked at afterwards.

A language model can write a fluent sentence about a company that is slightly wrong, so the brief should push it towards safe claims: what you do, why you are writing, and facts from a source you control.

Before you configure anything: decide the job

Most disappointing agent setups fail before a single message is sent, because the job was never defined. Answer four questions on paper first.

• Who is the target? A role, in a type of company, with a problem you actually solve. “Small businesses” is not a target; “owner-run clinics with two to ten staff who still book by phone” is.
• Why would they care now? What is the specific reason for contacting this person today rather than at random? A form they filled in, a download, a referral, a signal from your lead finder.
• What is the single next step? A short call, a reply to a question, a demo. One thing, stated plainly.
• Who takes over? A named human, with a time by which they will respond. An agent that raises its hand to nobody is just noise.

If you cannot answer these, an agent will not answer them for you; it will invent plausible answers.

Write the job description: the system prompt

The system prompt is the standing instruction the agent works from. Treat it as a job description plus a code of conduct. Vague prompts produce vague, generic messages; specific prompts produce messages that sound like your company.

The rules worth writing down

Put each of these in the prompt in plain words, and back the important ones with guardrails so they are enforced rather than merely requested.

• One prospect at a time. Each message is written for a single person, using what is known about them. No templated blasts with a name swapped in.
• A specific reason for contacting. Every message states why this person is hearing from you. If the agent cannot name a reason, it should not send.
• Short, human messages. A few sentences, plain language, no jargon, no walls of text, no stacked links.
• No claims you cannot verify. No invented customer names, no made-up numbers, no pretending to have read something it has not. If it does not know, it says nothing.
• Stop after unanswered follow-ups. A fixed, small number of follow-ups, then the prospect is left alone.
• Never contact opt-outs. Anyone who has asked to stop, unsubscribed or been marked do-not-contact is permanently off limits.
• Hand over, do not improvise. Pricing questions, complaints, legal questions and anything emotional go to a human.

Describe the voice, not just the rules

Give the agent a tone and a few real messages your best salesperson has written. Scalry agents have a tone: professional, friendly, direct, consultative or creative. Pick the one a careful human on your team would use with your customers.

Choose who the agent may contact

The most consequential decision is not the wording of a message but the list the agent works from. The same message is welcome to someone who asked about your product and unwelcome to a stranger scraped from a directory.

Narrow the ideal customer profile

Start with one segment. A narrow profile makes scoring meaningful and results readable, and you can widen it later with evidence. You cannot easily undo a broad launch that damaged your sender reputation.

Separate sources that imply interest from sources that do not

Some sources suggest the person has shown interest. Others only suggest they exist.

The default sales agent in a new Scalry workspace follows this logic: it can auto-enrol contacts only from sources that imply interest, such as the lead finder, funnels, your website, referrals and ads. Keep that restriction unless you have a specific reason to change it, one source at a time.

Skills: the menu of allowed actions

Skills are the menu of things the agent may do. An agent that can send but must wait for approval cannot send until you say so, so the smallest useful menu is the simplest way to limit what can go wrong.

Scalry skills include signal hunting, ICP scoring, prospect research, writing in your voice, building sequences, classifying replies, qualifying meetings, sending email, SMS or WhatsApp, creating notes and tasks, moving deal stages, booking meetings and escalating to a human.

How approvals work in practice

In Scalry, each skill can require human approval. When it does, the agent does not act; it proposes the action, and the proposal waits in an approval queue. Items there carry a status: pending approval, approved, completed, failed, skipped or rejected. You read what the agent wants to do, approve it, or reject it. Rejections are valuable data, because the pattern in what you reject tells you what to change in the prompt.

Approve anything that reaches a customer, changes a commitment or is hard to reverse. Let the agent act freely on anything that only affects your own records.

Guardrails: limits that do not depend on the prompt

A prompt is a request. A guardrail is a rule the system enforces, and you want both. Scalry offers these guardrail types:

• Working hours and a cooldown period between actions on the same contact, so follow-ups are spaced rather than stacked.
• Max follow-ups, a daily limit and max leads per day, which keep a mistake small.
• Exclude tag and never contact, for customers, partners and do-not-contact people.
• Topic restriction and require approval, for subjects to avoid and actions that need a human check.

The default sales agent in a new workspace has a daily action cap, and an agent takes one action at a time per contact in cycles with a cooldown, rather than firing everything at once. That pacing is slower than a blast, which is exactly why it is safer and easier to review.

A three-week rollout plan

Earn trust in stages.

Week 1: draft only

Require approval on every action that touches a customer, including sending, stage changes and booking. The agent researches, scores and drafts; you read everything. The aim of the first week is not results. It is to find out how the agent misreads your brief.

Fix the prompt, not just the draft: if the same problem appears three times, the brief is the cause. Reject anything with an unverifiable claim, and check which prospects were enrolled and why.

Week 2: approve and edit

Keep approval on for sending and edit before you approve. Fold recurring edits back into the prompt, and check that the reply classifier handles opt-outs correctly.

Week 3 and beyond: relax only the low-risk approvals

If drafts have been consistently acceptable, stop requiring approval for notes, tasks and escalations. Keep approval on first contact for longer; many teams never remove it, and that is reasonable.

Review the action log

Check the log daily at first, weekly later. Read five sent messages as the recipient would, look at what you rejected and why, read every reply labelled negative or unsubscribe, and chase down anyone who got more messages than your follow-up limit.

Handing over to a human

An agent that spots interest and then drops the ball wastes the prospect’s goodwill. A good handover is fast, complete, and does not make the prospect repeat themselves.

Escalate when the prospect asks about price or terms, asks for a call, sounds frustrated, raises a complaint or legal concern, or replies ambiguously.

What a good escalation contains

It should carry who they are and where they came from, the reason used in the first message, the full conversation in order, what the agent thinks the last reply means (labelled as its reading, not fact), and one suggested next step.

Agree a response time with the person who takes over; the same working day is a reasonable start.

Deliverability and compliance

An agent can send more than a human can, so it can damage your reputation faster too.

Protect your sender reputation

Complaints, bounces and ignored mail count against you, and the damage can spill onto your ordinary email.

Use a sending domain you can afford to protect, authenticate it, warm new mailboxes up gradually (Scalry includes an email warm-up agent type for this), keep volume low and steady, and remove bounced addresses.

Know the rules where your prospects are

Rules on unsolicited email and text messages vary by country and sometimes state: some require prior consent, some allow business-to-business outreach with conditions, and most require a clear identity and a working opt-out. Text and WhatsApp are typically stricter than email. If you are unsure about a market, take qualified advice before contacting people in it.

Whatever the local rule, identify yourself, honour opt-outs on every channel and keep a record of where each contact came from.

Judging success from your own data

Do not judge an agent by how much it sends; volume is the easiest number to inflate. Judge it as you would a person in the role.

• Positive replies, counting genuine interest rather than auto-responses.
• Meetings booked and held; booked is not held.
• Complaints and opt-outs, which point at the source or the message.
• Bounces and failed actions, which can threaten deliverability.
• Human time saved, including the time you spend approving.

Give it weeks, not days, and compare against your own history with the same segment. If you have no baseline, start recording one now.

Common mistakes

• Too broad an ideal customer profile. The agent has nothing specific to say, so it says something generic. Narrow the target first.
• No approvals at launch. The first real message is also the first time you see how the agent interprets your brief. Look before it reaches a customer.
• A generic prompt. “Be helpful and sell our product” describes nothing. Write the constraints.
• Judging in three days. Replies lag, and the first batch teaches you about the prompt more than about the market.

How Scalry handles this

Scalry has AI agents of several types: sales, content, support, custom and email warm-up. Each has a name, a job title, a tone, and a status of draft, active or paused. That makes an SDR agent a defined role in your workspace rather than a loose automation.

New workspaces come with default agents, including an SDR agent, and they start paused. Nothing runs until you activate them, which gives you time to write the brief, choose sources and set limits first. The default SDR agent can auto-enrol contacts only from sources that imply interest, such as the lead finder, funnels, your website, referrals and ads, and it has a daily action cap.

Skills define what an agent can do, and each skill can require human approval. Proposed actions wait in the approval queue until you approve or reject them. Guardrails cover working hours, cooldown, follow-up limits, daily limits, leads per day, exclusion tags, never-contact lists, topic restrictions and approval requirements. An agent works one action at a time per contact, in cycles with a cooldown, and an enrollment can be active, paused, completed or exited, so you can see where each prospect stands and stop any of them.

Scalry gives you a controllable worker with a log. It does not choose your market, guarantee replies or replace the person who runs the conversations that matter.

Questions to ask before you commit

Whichever way you are leaning, these are the questions that change the answer. Work through them against your own numbers rather than anyone's feature matrix — including this one.

• Who exactly is the agent allowed to contact? If you cannot write the profile in two sentences, the agent cannot follow it.
• Which sources imply interest? Anything else stays out of automatic enrolment.
• Who takes over when someone replies, and how fast? Name a person and a time.
• Which actions must you approve for the first two weeks?
• What are the limits on follow-ups and daily volume?
• Do you know the rules for the countries you will contact?

If you cannot answer the first three, you are not ready to choose between Scalry and doing this by hand yet — you are still deciding what you need, which is a cheaper problem to solve first.

A short plan to get started

Before activating, write the profile, list the sources, draft the brief and set low guardrails. In week one require approval on every customer-facing action. In week two edit and approve while checking reply classification. From week three relax only low-risk approvals, one change at a time, and review results against your own baseline.

Frequently asked questions

What is an AI SDR agent?

An AI SDR agent does the repetitive work of a sales development rep: it researches prospects, scores fit, drafts and sends messages, classifies replies and hands interested people to a human. It works one prospect at a time in small steps you can review.

Should an AI sales agent send emails without approval?

Not at first. Require approval on every customer-facing action for the first week, then edit and approve in the second. Relax approvals only on low-risk actions such as notes and tasks, and keep a human check on first contact for as long as you want it.

Which contacts should the agent be allowed to enrol?

Contacts from sources that imply interest, such as website enquiries, funnel sign-ups, referrals, ad lead forms and lead finder results matching a narrow profile. Do not auto-enrol bulk imports, existing customers or support enquiries.

What guardrails should I set before launch?

Set working hours, a cooldown, a low maximum number of follow-ups, a daily limit, a cap on new leads per day, an exclusion tag and a never-contact list. Start tighter than you think you need and loosen gradually.

How long should I test before judging results?

Several weeks, not days. Replies arrive slowly and the first batch mostly teaches you where the prompt is weak. Compare against your own past outreach to the same segment.

Is automated outreach legal?

It depends on the country, and sometimes the state. Rules on unsolicited email, text and WhatsApp differ on consent, business-to-business exceptions and opt-out requirements. Prefer prospects who showed interest, honour opt-outs, and take qualified advice for markets you are unsure about.

When should the agent hand a conversation to a person?

When the prospect asks about price or terms, requests a call, sounds frustrated, raises a complaint or legal concern, or replies ambiguously. Include who they are, why they were contacted, the full conversation and a suggested next step.