AI sales agents: choose the job before the model
An AI sales agent is software that takes a defined sales action using customer context and rules. Pick the job first.
ManySetter teamUpdated 9 min read
On this page
Chapter 01Start here
Four common AI sales-agent jobs
They share a name, not a job. Evaluate each one on the work it owns.
Prospecting and outreach
Researches accounts with CRM data and drafts outreach, like Salesforce Agentforce (opens in a new tab) or HubSpot’s Prospecting Agent (opens in a new tab).
Inbound qualification
Takes a live inquiry and decides what’s next, including “not enough information yet.”
Scheduling
Checks a call fits, proposes a time and records the event. A sent link isn’t a booking.
Deal assistance
Summarizes accounts and prepares reps, without ever talking to a prospect.
Concrete jobs by agent type
| Agent type | Starts from | Does | Hands off to |
|---|---|---|---|
| ProspectingStarts from: A CRM list or target accountsDoes: Researches accounts, drafts first outreachHands off to: A rep who reviews it | A CRM list or target accounts | Researches accounts, drafts first outreach | A rep who reviews it |
| Inbound qualificationStarts from: A message the lead sentDoes: Answers, asks one fit question, keeps the stateHands off to: A call or a person | A message the lead sent | Answers, asks one fit question, keeps the state | A call or a person |
| SchedulingStarts from: A qualified leadDoes: Proposes times, sends the link, waitsHands off to: A confirmed calendar event | A qualified lead | Proposes times, sends the link, waits | A confirmed calendar event |
| Deal assistanceStarts from: An open deal in the CRMDoes: Summarizes calls and emails, suggests next stepsHands off to: The rep | An open deal in the CRM | Summarizes calls and emails, suggests next steps | The rep |
Chapter 02Architecture
The system behind a useful sales conversation
A reliable agent is more than a prompt. Five layers decide what it can do, and each type builds them differently.
Architecture by agent type
| Layer | Prospecting | Inbound setting | Scheduling | Deal assistance |
|---|---|---|---|---|
| ModelProspecting: Research and outreachInbound setting: Writes each replyScheduling: Matches open slotsDeal assistance: Summarizes calls | Research and outreach | Writes each reply | Matches open slots | Summarizes calls |
| OrchestrationProspecting: Steps across a listInbound setting: Answer, ask, book, hand offScheduling: Propose, send, confirmDeal assistance: On a rep’s request | Steps across a list | Answer, ask, book, hand off | Propose, send, confirm | On a rep’s request |
| MemoryProspecting: Account historyInbound setting: The thread and its sourceScheduling: Past bookingsDeal assistance: Deal history | Account history | The thread and its source | Past bookings | Deal history |
| Lead stateProspecting: CRM stageInbound setting: Criteria met or unknownScheduling: Proposed, sent, bookedDeal assistance: Deal stage, risks | CRM stage | Criteria met or unknown | Proposed, sent, booked | Deal stage, risks |
| Learning loopProspecting: Accepted meetingsInbound setting: Corrections, calls heldScheduling: No-shows, reschedulesDeal assistance: Rep edits | Accepted meetings | Corrections, calls held | No-shows, reschedules | Rep edits |
- 01TriggerWhat started it
- 02MemoryWhat we know
- 03RulesWhat it may do next
- 04ModelThe reply itself
- 05OutcomeA verified event
Can’t inspect those layers? A natural-sounding reply tells you little. The method: AI lead qualification.
Chapter 03In practice
An inbound lead after a Reel
The inbound qualification job, step by step. A Manychat flow sends the guide from a Reel, then the lead writes back.
What the agent keeps after two messages
| Field | Value |
|---|---|
| SourceValue: Reel “3 DMs that book calls” · keyword GUIDE | Reel “3 DMs that book calls” · keyword GUIDE |
| SaidValue: Coaches 4 clients, asks if it fits | Coaches 4 clients, asks if it fits |
| Fit ruleValue: Coaches with 2 to 10 clients: met | Coaches with 2 to 10 clients: met |
| Still unknownValue: What stops their DMs from becoming calls | What stops their DMs from becoming calls |
| Next objectiveValue: Ask it, then propose a fit call | Ask it, then propose a fit call |
It answers first and asks the one missing question. It proposes a call only if the fit rule holds. A custom request goes to a person, with a summary. The meeting counts once the calendar confirms it.
Test Manychat first
Manychat already has flows and native AI (opens in a new tab). If a native flow solves the job, use it.
Chapter 04Compare
AI sales agent, chatbot or AI SDR?
A chatbot answers predictable questions or follows a script. An AI SDR prospects new accounts. An inbound setter qualifies people who wrote first and decides whether to propose a meeting.
- If you needFind and contact new accountsLook atA CRM prospecting agentResearch, enrichment and outreach from your CRM.
- If you needHandle incoming DMs and qualify people
- If you needBook after qualification
- If you needHelp reps understand deals fasterLook atA sales copilotSummaries and next steps inside the CRM.
A business may use several. Give each agent one clear responsibility.
Chapter 05Test
How to evaluate before deployment
Write five real test conversations and the next action you expect for each.
0 of 5 completed
Then check the failure path:
- Can a person pause the agent?
- Can you explain why a call was proposed?
- What happens when two messages arrive seconds apart?
- Does a clicked calendar link count as booked?
- Can you see each conversation’s source and cost?
On Manychat, test each channel
Instagram, WhatsApp and Messenger have their own triggers and messaging windows (opens in a new tab). Don’t generalize one channel’s flow to another.
Chapter 06Measure
Measure the part of the funnel it owns
One job, one metric
| Job | The metric that proves it |
|---|---|
| Inbound qualificationThe metric that proves it: Qualified lead rate | Qualified lead rate |
| SchedulingThe metric that proves it: Calendar-confirmed bookings | Calendar-confirmed bookings |
| ProspectingThe metric that proves it: Accepted opportunities | Accepted opportunities |
| Deal assistanceThe metric that proves it: Rep time saved, deals moved | Rep time saved, deals moved |
Track conversations, responses, qualified leads, handoffs, calls proposed and confirmed bookings. Read the lost conversations as closely as the booked ones.
More replies can still cost you
An agent that answers more but misclassifies high-value prospects hurts the business. Compare it with what you have today, on the same lead sources.
ManySetter
An inbound setter that remembers.
ManySetter uses memory, lead state and the rules you teach on inbound Manychat conversations. Not cold outreach or calls.
- Inbound DMs from Manychat
- Keeps each lead’s state
- Learns from your corrections
Build and test free · Manychat billed separately
Chapter 07Sources