AI Appointment Setter: What It Does and How to Deploy

An AI appointment setter is software that replies to leads, asks your qualifying questions and books calls on your calendar, then hands the conversation to a person when it should. It works on repeatable conversations with a clear goal. It does not replace judgment on pricing, anger, nuance or high-value deals. This guide gives you everything needed to build one yourself.

If you sell through calls, appointment setting is where leads quietly leak away. Replies sit unread, follow-ups slip, and a prospect who was interested on Tuesday has gone cold by Friday. This is a build guide, not an overview. By the end you will have the knowledge base template, the system prompt, the stage-by-stage scripts (text and voice), an objection library, a reply classifier, handover rules, booking and reminder templates, a 20-message test script, a KPI table, a rollout plan with exit criteria and a day-by-day first week. If the underlying idea is new to you, read what an AI agent is first. If you want the upstream half (finding and qualifying prospects before they reply), see our guide to AI agents for lead generation.

How an AI appointment setter works from first reply to booked call

AI appointment setting covers the conversation between "a lead showed interest" and "a call is on the calendar and the person shows up". Underneath, it is a chain of small, checkable jobs, not one magic prompt. Anthropic's engineering team distinguishes workflows (code defines the path the model follows) from agents (the model decides its own steps), and describes patterns such as routing, where an input is classified and sent down a specialised path, and prompt chaining, where each step feeds the next with checks in between. It also advises starting with the simplest approach and adding complexity only when it clearly helps. (Source: Anthropic, "Building effective agents", accessed 6 November 2026.) For appointment setting that translates into a mostly fixed path: classify the reply, answer or qualify, offer times, confirm, remind, recover no-shows. The model writes the words. Your rules decide the path.

Channels

The same stages run on email, LinkedIn or Instagram direct messages, messaging apps and voice. Each channel has its own rules, so decide up front where your setter is allowed to operate. Email carries sender requirements that affect whether your messages are delivered at all (see the compliance section below). Messaging platforms such as WhatsApp apply their own policies on opt-in and on which messages you may send outside a recent-conversation window, so read the current policy of every platform before you automate on it. Voice adds its own disclosure and consent questions. We recommend starting with one text channel, getting it right, and only then adding a second.

What an AI appointment setter cannot do

Being clear about limits is what separates a good deployment from a brand problem.

• It cannot judge nuance reliably. An angry, confused or emotionally loaded lead should go to a person, immediately.
• It cannot invent your offer. It knows only what you write down. If pricing, scope or availability is unclear in its instructions it will dodge or guess. The fix is a knowledge base (Asset 1) and a rule to hand over when the answer is not there.
• It cannot rescue a weak offer or a bad list. A setter amplifies the pipeline you give it. In our own outreach, a generic opener sent identically to every founder produced only two replies from roughly 375 contacts, and the fix was a sharper opener tied to one service, not more automation.
• It must not make promises. No guarantees, discounts, timelines or commitments unless you have written exactly what it may say.
• It is not a compliance shortcut. Rules on unsolicited messages and automated calls differ by country and channel. We are not giving legal advice. Check each market you contact and always honour opt-outs.
• It cannot fix a channel that does not fit. We have told prospects plainly that our outbound approach did not suit them when the signal they wanted to target (for example a life situation) could not be found in any professional database. If you cannot verify who qualifies from data you can see, do not automate the conversation, fix the targeting first.

AI setter vs human setter vs hybrid: how to decide

The honest answer to "ai setter vs human setter" depends on volume, deal size and how well your qualification can be written down. Use the table, then apply the decision rules under it.

We do not quote cost or performance comparisons, because they depend on your market, offer and set-up. Compare them on your own numbers after a trial.

Asset 1: Decision rules

Choose AI first when ALL are true:
- You get enough replies that some wait more than a few hours.
- Your qualification fits in 3 to 4 written questions.
- A wrong answer costs you an awkward email, not a lost relationship.
- Someone on your team will read conversations every week.

Choose a HUMAN setter when ANY are true:
- Deals are high-ticket or sensitive and leads expect a personal touch.
- Qualification depends on judgment you cannot yet write down.
- Fewer than a handful of replies arrive each week (do it yourself first).
- Your market expects voice notes, local language nuance or relationship building.

Choose HYBRID when you want speed AND warmth:
- AI: first response, qualification, scheduling, reminders.
- Human: pricing talk, objections beyond the library, anyone upset,
  anyone senior, every conversation the AI marks "maybe".

Re-decide every quarter using your own KPI table.

Before you build: the inputs the setter needs

Set the tooling aside. Write these four things in a document first. If you cannot write them, the setter cannot act on them.

1. What a booked call is. Call types, duration, who takes them, minimum notice, which days and hours are bookable, and what makes a lead worth a slot.
2. Who qualifies. A rubric a stranger could apply in 30 seconds (Asset 3).
3. What the setter may say. The knowledge base (Asset 2).
4. When it must stop. The handover rules (Asset 8).

Step 1: Write the knowledge base so the setter never invents answers

Most setter failures are knowledge failures. The model fills silence with plausible text. The cure is a short document it must use, plus an explicit instruction to hand over when the answer is not in it. Keep it to one or two pages, in plain language, and write every answer the way you want it said.

Asset 2: Knowledge base template

KNOWLEDGE BASE: <COMPANY_NAME>   Last updated: <DATE>   Owner: <NAME>

1. WHAT WE DO (max 3 sentences)
We help <ICP_TYPE> with <PROBLEM> by <HOW_IN_ONE_SENTENCE>.

2. WHO WE ARE FOR
Good fit: <3 to 5 bullet points>
Not a fit: <3 to 5 bullet points>

3. THE CALL
Name: <CALL_NAME>   Length: <MINUTES>   Who joins from our side: <ROLE>
What happens on it: <2 to 3 bullets>
What the lead should bring: <1 to 2 bullets>
It is free / it is paid: <STATE_ONE>

4. PRICING POLICY (choose one and write it exactly)
A) "Pricing depends on scope. We cover it on the call."
B) "Engagements typically start from <APPROVED_RANGE>." (only if you approve it)
C) "I cannot quote prices. A person will answer that before or on the call."
The setter must never state a number that is not written here.

5. HOW WE WORK (process in 3 to 5 steps, no promises of outcomes)

6. FAQ (write 10 real questions leads ask, with approved answers)
Q: <QUESTION>
A: <ANSWER_IN_2_SENTENCES>

7. PROOF YOU MAY MENTION
<Only case studies and facts you have permission to share, worded exactly>

8. HARD EXCLUSIONS (the setter must hand over or decline)
- Custom discounts, contracts, legal or compliance advice
- Guarantees of results or timelines
- Anything about named clients without permission
- Topics outside <SCOPE>

9. LOGISTICS
Timezones we book in: <LIST>   Working days and hours: <DAYS_AND_HOURS>
Minimum notice: <HOURS>   Reschedule rule: <RULE>
Booking link (fallback): <URL>   Human contact for escalations: <NAME_AND_CHANNEL>

How to check it worked: hand the document to a colleague and ask ten questions a lead might ask. If they can answer eight from the page alone, the setter can too. Every question they cannot answer is either a missing FAQ entry or a deliberate handover.

Step 2: Define qualification as a rubric

Qualification in a conversation is not an interrogation. Ask the minimum needed to protect a calendar slot, one question at a time, and score the answers against criteria you wrote in advance. Three to four questions is usually enough. More than that feels like a form and loses people.

Asset 3: Qualification script and rubric

QUESTIONS (ask one per message, in this order, skip any already answered)
Q1 Situation: "How are you handling <PROBLEM_AREA> at the moment?"
Q2 Fit: "Roughly how big is the team / business (a range is fine)?"
Q3 Need or timing: "Is this something you want to sort out this quarter, or later?"
Q4 Decision: "Will anyone else need to be on the call with you?"

RUBRIC (total out of 10)
Problem is real and current (Q1)            0 to 3
Size or type within ICP (Q2)                0 to 2
Timing within <WINDOW> (Q3)                0 to 3
Decision maker or access to one (Q4)        0 to 2

DECISION RULES
8 to 10: qualified. Offer times.
5 to 7: maybe. Hand over to a person with a summary; do not book.
0 to 4: not now. Send the polite close and add to nurture list.
Any answer that is unknown scores 0. Never infer.

ALWAYS ask Q4. If another decision maker exists, invite them to the call.
Booking someone who cannot decide is how closing calls end in
"I need to speak to my partner".

Step 3: The setter's system prompt (copy and paste)

The system prompt is the contract. It sets the role, the tone, what the setter may and may not say, when it must stop and how to use the knowledge base. Paste your knowledge base and rubric into the marked slots. Keep the structure: rules first, then context.

Asset 4: Appointment setter system prompt

ROLE
You are the scheduling assistant for <COMPANY_NAME>. You reply to people who
responded to our outreach or contacted us. Your single goal is to book a
<CALL_NAME> with a qualified person, or to hand the conversation to a human.
You are an AI assistant. If anyone asks whether you are a person or a bot,
say plainly that you are an AI assistant for <COMPANY_NAME> and offer to
connect them with a person.

SOURCES OF TRUTH
Use ONLY the knowledge base and conversation history below. If the answer is
not there, do not guess. Say: "Good question. I want to give you an accurate
answer, so I will ask <HUMAN_NAME> to reply." Then hand over.

KNOWLEDGE BASE
<PASTE_KNOWLEDGE_BASE>

QUALIFICATION RUBRIC
<PASTE_RUBRIC>

TONE
- Warm, brief, plain language. Two to four short sentences per message.
- One question per message. Never stack questions.
- Mirror the lead's formality. No hype words, no exclamation marks, no emojis
  unless the lead uses them first.
- Write like a courteous colleague, not a salesperson.

YOU MAY
- Answer factual questions from the knowledge base.
- Ask the qualification questions in order.
- Offer up to three time slots returned by the calendar tool.
- Confirm, reschedule or cancel a booking through the calendar tool.
- Send the reminder and recovery messages defined below.

YOU MAY NOT
- State prices, discounts or terms that are not in the knowledge base.
- Promise results, timelines or guarantees.
- Give legal, tax, financial or compliance advice.
- Mention other clients or share any personal data.
- Argue with, pressure or guilt a lead. Objections get one calm response,
  then you offer the call or step back.
- Invent availability. Offer only slots the calendar tool returns.
- Continue after "stop", "unsubscribe" or similar. Confirm once and end.

HAND OVER TO A HUMAN (stop replying and notify <HUMAN_NAME>) WHEN
- The lead asks for a person, is upset, rude or distressed.
- The lead raises pricing negotiation, contracts, legal or compliance topics.
- The lead is senior at a company named on our priority list: <LIST_OR_NONE>.
- The score is "maybe" (5 to 7).
- You are unsure, the question is outside the knowledge base, or you have
  sent <MAX_MESSAGES> messages without progress.
- The lead mentions media, a complaint, or any sensitive personal matter.
When you hand over, write a summary: who they are, what they said, what they
want, what was answered, recommended next step.

BOOKING
Collect: full name, email, timezone, preferred times. Book only after the lead
clearly chooses a slot. Then restate the time in the lead's timezone.

OUTPUT FORMAT
Return JSON only:
{"reply": "text to send or empty if handing over",
 "action": "reply" | "offer_times" | "book" | "reschedule" | "stop" | "handover",
 "stage": "qualify" | "objection" | "scheduling" | "confirmed" | "recovery",
 "score": null or 0-10,
 "handover_reason": "" or short text,
 "notes_for_crm": "one or two lines"}

Why it is built this way: Anthropic's guidance describes agents as best kept to a bounded task with clear stopping conditions and human checkpoints, with testing in a safe environment before wider release, because errors can compound. The maximum-message cap, the JSON action field and the handover list are those ideas applied to a setter. Treat the action field as the only way the model can cause something to happen, and let your automation (not the model's free text) perform the booking.

Step 4: Design the conversation, stage by stage

Below is the conversation as we run it, stage by stage, with full scripts for text channels (email or direct message). The structure follows a pattern we use in our own founder outreach: answer a question in two or three lines and then offer the call, and when someone shows interest, reply within hours with two time options rather than a long explanation.

Stage 1: First response

Goal: acknowledge, answer if asked, ask one question or offer times. Never open with a paragraph about yourselves.

Asset 5: Stage scripts for email and direct messages

FIRST RESPONSE: lead said "interested" or "tell me more"
Hi <FIRST_NAME>, thanks for getting back to me. In short: we help <ICP_TYPE>
with <PROBLEM>. Before I suggest a time, one quick question: how are you
handling <PROBLEM_AREA> at the moment?

FIRST RESPONSE: lead asked a factual question that is in the knowledge base
Hi <FIRST_NAME>, good question. <ANSWER_IN_TWO_SENTENCES>
If it helps, I can find a time for a short call. Would this week or next suit you?

FIRST RESPONSE: lead asked something not in the knowledge base
Hi <FIRST_NAME>, good question. I want to give you an accurate answer, so I
have asked <HUMAN_NAME> to reply to you directly. In the meantime, is there
anything else I can clarify?

QUALIFICATION (one question per message)
Thanks, that helps. Roughly how big is the team involved (a range is fine)?
...
Understood. Is this something you want to sort out this quarter, or later?
...
Last one: will anyone else need to join the call with you?

OFFERING TIMES (qualified)
Thanks <FIRST_NAME>, that sounds like a good fit for a <MINUTES>-minute call
with <HOST_NAME>. These work on our side (times in <LEAD_TIMEZONE>):
1) <SLOT_1>   2) <SLOT_2>   3) <SLOT_3>
Reply with a number and I will book it. If none suit, tell me a day that does.

CONFIRMING
Booked: <CALL_NAME> on <DAY_DATE> at <TIME> (<TIMEZONE>) with <HOST_NAME>.
Calendar invite sent to <EMAIL>. On the call we will <ONE_LINE_AGENDA>.
If you can, bring <ONE_THING>. Need to move it? Reply here and I will find another time.

NOT A FIT (score 0 to 4)
Thanks for the context, <FIRST_NAME>. From what you describe, we are probably
not the right fit right now, and I would rather say so than waste your time.
If <TRIGGER> changes, reply here and I will pick it up.

NOT NOW
No problem, <FIRST_NAME>. I will leave it there. If it becomes a priority,
reply to this thread and we will find a time. <OPTIONAL_SOFT_EXIT: link to
a useful resource or community, only if you have one>

Stage 2: Objection handling

The setter should do three things with an objection: acknowledge it, answer from approved content, and either offer the call or step back. It gets one attempt per objection. Pushing twice is how automated conversations become unpleasant. Anything not in this library goes to a person.

Asset 6: Objection-handling library (top 8)

Stage 3: Classify every reply

Classification is the routing step. Run it before the setter replies, so that unsubscribes, out-of-office replies and angry messages never get a salesy answer. Anthropic describes routing as classifying an input and sending it to a specialised follow-up, so each path can be tuned on its own. A small, cheap model call is often enough for this step, but test it on your own messages before trusting it.

Asset 7: Reply classification prompt (JSON output)

Classify the latest reply from a lead into exactly one label.

Labels:
interested | question | objection_price | objection_other | not_now |
not_a_fit | wrong_person | referral | pick_slot | reschedule |
cancel | unsubscribe | out_of_office | upset | asks_for_human | unclear

Conversation so far:
<CONVERSATION_HISTORY>

Latest reply:
<REPLY_TEXT>

Return JSON only:
{
  "label": "...",
  "confidence": 0 to 1,
  "summary": "one line",
  "needs_human": true or false,
  "extracted": {"timezone": "", "chosen_slot": "", "new_contact": "", "date_mentioned": ""},
  "suggested_next_action": "..."
}

Rules:
- Any request to stop contact is "unsubscribe", whatever else it says.
- Any sign of anger, a complaint or distress is "upset" and needs_human is true.
- If the lead asks for a person, label "asks_for_human" and needs_human is true.
- If confidence is below 0.7, use "unclear" and needs_human is true.
- Never invent a timezone or a slot. Leave empty if not stated.

Stage 4: Handover rules

Write the handover rules as a table that a person and the automation can both read. Name the owner and the maximum response time for each, or handovers become a black hole.

Asset 8: Handover and escalation rules

Hand-off message to the lead: "Thanks, <FIRST_NAME>. I am passing this to <HUMAN_NAME>, who will reply to you personally. You will not need to repeat anything." The summary to your team: name, company, channel, what the lead said in their own words, what was answered, score, recommended next step.

Stage 5: Scheduling that actually works

Most booking errors are calendar rules, not AI errors. Google's help documentation for appointment schedules illustrates the controls you should insist on, whichever calendar tool you use: set the appointment duration and availability, set a minimum lead time and a maximum booking window (Google's default minimum lead time is 4 hours), add buffer time between calls, cap bookings per day, check your connected calendars for conflicts, choose Google Meet, in-person, phone or "to be specified later" as the location, require email verification on the booking form to reduce spam bookings, and send confirmations and reminders (up to 5 reminders, with fixed text for the built-in emails). It also notes that co-host availability is not checked by default, so add co-host calendars to the check if two people take the call. (Source: Google Calendar Help, "Create an appointment schedule", accessed 6 November 2026.)

Our rules for the scheduling step:

• The setter offers only slots returned by the calendar tool, never from memory. Offer two or three, not a full grid.
• Always state the timezone in words next to the time. If the lead's timezone is unknown, ask before offering times.
• The booking action needs four values or it fails safely: lead email, chosen slot, timezone, call type. Make the tool hard to misuse, as Anthropic advises when designing tools for models: precise inputs, clear boundaries, tested on many examples.
• After booking, read the event back from the calendar and compare it with what the lead agreed. If they differ, hand over.
• Before an important call, a person confirms the lead is the decision maker, or that the decision maker is joining.

Stage 6: Confirmations, reminders and no-show recovery

A booked call is not a held call. Reminders protect the calendar, and a calm recovery message rescues a share of no-shows. These are our starting timings, adjust them with your own show-rate data.

Asset 9: Booking, reminder and recovery templates

CONFIRMATION (immediately)
Booked: <CALL_NAME> on <DAY_DATE> at <TIME> (<TIMEZONE>) with <HOST_NAME>.
Join here: <LINK_OR_PHONE_DETAILS>. Need to change it? Reply and I will find a new time.

REMINDER 1 (24 hours before)
Hi <FIRST_NAME>, a reminder that we speak tomorrow at <TIME> (<TIMEZONE>).
Reply "move" if you need another time.

REMINDER 2 (2 hours before)
See you at <TIME> today. Link: <LINK>. <HOST_NAME> will cover <ONE_LINE_AGENDA>.

NO-SHOW RECOVERY 1 (within 30 minutes after the start)
Hi <FIRST_NAME>, we missed you just now. No problem. Want me to find another time
this week? Here are two options: <SLOT_1> or <SLOT_2>.

NO-SHOW RECOVERY 2 (3 days later, only if no reply)
Hi <FIRST_NAME>, last note from me on this. If the timing has changed, reply and
I will book something. Otherwise I will close this out. Thank you.

RESCHEDULE REQUEST
Of course. These times are open (<TIMEZONE>): <SLOT_1> / <SLOT_2> / <SLOT_3>.
Which suits you?

CANCEL REQUEST
Understood, I have cancelled it. If you would like to talk later, reply here.

Voice variant: the outbound or inbound AI caller

If your setter makes or answers phone calls, the same stages apply with tighter rules. Our voice-agent build notes (using a voice platform and a workflow tool) set out the pattern we follow: the system prompt defines the persona, the goal and the constraints, the agent discloses that it is an AI assistant near the start of the call, answers a question before attempting to book, books only after the lead is qualified and the value has been explained, and never discusses price beyond "an expert can answer that on the call". After every call, an end-of-call event sends the transcript and call outcome to a workflow that classifies the disposition (booked, not interested, call back) and logs a row in a sheet or CRM. Test every call path end to end before real leads hear it. Check the telephone-marketing, recording-consent and AI-disclosure rules in each country you call, and keep calling hours conservative.

Asset 10: Voice setter opening and rules

OPENING (within the first 10 seconds)
"Hi, is this <FIRST_NAME>? This is an AI assistant calling from <COMPANY_NAME> on
behalf of <HUMAN_NAME>. You replied to our message about <TOPIC>. Do you have
two minutes?"

IF YES: ask Q1 and Q3 from the rubric only, then offer times:
"Based on that, a <MINUTES>-minute call with <HUMAN_NAME> would help. I have
<SLOT_1> or <SLOT_2>, which suits you?"

IF NO / BAD TIME: "No problem. When is a better time to call?" Log it. End politely.

IF THEY ASK "ARE YOU A ROBOT": "Yes, I am an AI assistant. I can book you with
<HUMAN_NAME> or have them call you directly. Which would you prefer?"

IF THEY ASK PRICE: "I cannot quote prices myself, but <HUMAN_NAME> can answer
that on the call. Shall I book a time?"

VOICE RULES
- Speak slowly. Short sentences. Do not interrupt.
- Never record or continue if they ask you to stop. Apologise and end the call.
- Confirm the booking by repeating the day, time and timezone, then send a text
  or email confirmation.
- End every call within <MAX_MINUTES> minutes or hand over.

A worked example (illustrative)

This example is illustrative. The company, people and details are fictional, invented to show how the pieces connect. It is not a client result.

The seller: Brightpath Advisory, a small consultancy that helps logistics firms (50 to 200 staff) design hiring processes. Offer: a 30-minute diagnostic call with the founder, free. Pricing policy: "depends on scope, covered on the call". The setter is named "the Brightpath assistant" and works over email.

Across the whole conversation the assistant never stated a price, never promised an outcome, and never answered the contract question. The founder spent a few minutes approving drafts instead of an afternoon in the inbox, and the person who took the call arrived with a summary. In review mode, step 2 would have been shown to a person first, which is exactly where you catch a wrong tone before a lead sees it.

Step 5: Red-team the setter before it meets a real lead

Anthropic recommends extensive testing in a sandbox, with guardrails, before agents act on real work. For a setter, that means sending it hostile and awkward messages in a test inbox and checking each result against the expected behaviour. Run all 20. Re-run the whole set after any change to the prompt or the knowledge base.

Asset 11: Red-team test script (20 messages)

Pre-launch QA checklist

• The knowledge base was tested: a colleague answered eight of ten sample questions from it alone.
• All 20 red-team tests passed, and results were saved.
• The pricing policy is written exactly and the setter never states any other figure.
• The setter identifies itself as an AI assistant when asked and, on voice, near the start of the call.
• Every handover rule has a named owner and a response target.
• Unsubscribe and "stop" replies suppress the contact across every sequence and channel.
• The calendar tool returns only valid slots; buffer time, daily cap and minimum notice are set.
• A test booking appears correctly in the calendar, in the CRM and in the lead's timezone.
• Reminders and recovery messages fire once, at the right times, and stop after a reply.
• Email sending domain is authenticated, not your main domain, and has an unsubscribe route.
• A person has read at least 30 full drafted conversations in review mode.
• A kill switch exists: you can pause all sending in one step.

Rollout plan: review mode, small slice, then widen

Do not switch everything on. Move through three phases and advance only when the exit criteria are met. The thresholds below are our suggested starting points; set your own from your risk tolerance.

An evaluator pass can help during phase 1: Anthropic describes the evaluator-optimizer pattern, where one model call produces a response and another critiques it against clear criteria. A second call that checks each draft against your "may not" list before a person sees it can reduce reviewer workload, though it does not remove the need for human review.

Common mistakes and fixes

What to measure and how to read it

Asset 12: KPI table

Read these weekly. Judge conversion metrics only after a few hundred conversations, since small samples swing wildly. Change one thing at a time (a script, a rule, a reminder time) so you know what moved the number. Always read a sample of actual conversations alongside the numbers, because metrics will not show you a rude sentence.

Deliverability and compliance: the non-negotiables

If your setter replies by email, the same sender rules apply to its messages as to your outreach. Google publishes requirements for sending to personal Gmail accounts. As of 6 November 2026 its guidance says:

• All senders should set up SPF or DKIM authentication for sending domains, use a TLS connection, and keep spam rates reported in Postmaster Tools below 0.30 percent.
• Senders of more than 5,000 messages a day to Gmail accounts must also set up DMARC (the policy may be set to none), align the sending domain with SPF or DKIM, and support one-click unsubscribe with a clearly visible unsubscribe link in marketing and subscribed messages.
• Google advises keeping the spam rate below 0.10 percent and never reaching 0.30 percent or higher.

(Source: Google, "Email sender guidelines", accessed 6 November 2026.)

Practical rules we apply regardless of volume: send from a secondary domain, never your main one; include a plain opt-out line; stop the setter the moment someone asks to stop; keep a single suppression list shared by every tool; tell people when they are talking to an AI if they ask or if local rules require it; and keep records of consent where your market expects them. Rules on unsolicited messages and automated calls differ by country and platform. This is not legal advice, so check the requirements for every market you contact.

Your first week, day by day

1. Day 1: Write the booked-call definition and the rubric (Asset 3). Write the knowledge base (Asset 2). Decide human, AI or hybrid using Asset 1.
2. Day 2: Fill in the system prompt (Asset 4). Set up the calendar rules: duration, minimum notice, buffer time, daily cap, location. Create a test inbox.
3. Day 3: Load the stage scripts (Asset 5), the objection library (Asset 6) and the reply classifier (Asset 7). Run ten sample replies through and read every output.
4. Day 4: Write the handover table (Asset 8) with owners and targets. Wire the notifications and the suppression list. Do a test booking end to end and check calendar, CRM and timezone.
5. Day 5: Run all 20 red-team tests (Asset 11). Fix the prompt or knowledge base for every failure and re-run the full set.
6. Day 6: Start phase 1: review mode on live replies, with a named reviewer. Log every edit you make and why.
7. Day 7: Review the week: edits, handover reasons, new questions to add to the knowledge base. Set up the KPI table (Asset 12). Decide whether phase 1 exit criteria are close and plan next week.

Frequently asked questions

Sources and further reading

• Anthropic, "Building effective agents" (workflows versus agents, routing, prompt chaining, evaluator-optimizer, human checkpoints, stopping conditions, tool design, starting simple). Accessed 6 November 2026.
• Google, "Email sender guidelines" (authentication, spam rate and one-click unsubscribe requirements). Accessed 6 November 2026.
• Google Calendar Help, "Create an appointment schedule" (duration, lead time, buffers, daily caps, conflict checks, conferencing options, email verification, reminders). Accessed 6 November 2026.
• SFC internal outreach playbooks, prompt library and voice-agent build notes (methods adapted and anonymised; no client data used).

What to do next

Do Day 1 today: write what a booked call is, then the knowledge base. That document is the foundation of any setter, human or AI. If you would rather have a person on the inbox, we place vetted human appointment setters from our global talent network. If you want the AI route built and running in your own accounts, we build AI outbound agents that handle outreach, qualification and handover into your CRM. Not sure which suits you? Book a short call and we will talk through your volume and your offer. If you are still choosing a provider, our guide to choosing an AI automation agency lists the questions to ask.