Major

AI Agents for Sales: 6 Workflows Built to Run Reliably

Sales agents can do more than draft outreach. See how repeatable lead routing, follow-up, and account updates become deterministic apps, with the model reserved for judgment and the work governed where it runs.

Jose GironUpdated
Sales team mapping a lead pipeline with sticky notes on a whiteboard

What AI agents do in sales

AI agents for sales carry multi-step work across your CRM, inbox, calendar and data sources: qualifying and routing leads, researching accounts, drafting follow-ups, cleaning the pipeline and flagging renewal risk. The agents that hold up in production use a model only for the judgment calls. Repeatable steps run as deterministic apps that keep their own state and logs.

That split is where Major fits. Major is the enterprise platform where agents build the software they run on. When an agent works out how to route a lead or clean a stale deal, it builds an app for that step, then runs the app instead of reasoning through the step again. Reason once, run forever. The app holds the account history, and RevOps and IT can read every action it took.

The audience for this guide is RevOps leads and sales operations owners who want agents doing real work in Salesforce or HubSpot without a black box deciding who owns a lead.

Why re-reasoning every record fails in sales

Salesforce's guide to AI sales agents describes the agent as software that uses sales and customer data to carry out work such as lead response, outreach and quoting. IBM's overview recommends focused objectives, clear handoff points for people and governed CRM records.

An agent that reasons from scratch on every lead behaves a little differently each time. It spends tokens on rules that never change, and it leaves no durable record of why a lead went to one rep. Put the territory rule in code and the thousandth lead gets the same treatment as the first.

One limit applies to every approach. A routing app cannot enforce a territory rule nobody has written down, and an agent cannot fix a CRM full of duplicate accounts.

Six sales workflows agents can run

Each workflow below splits into what code does every time and what the model decides.

  • Lead qualification and routing. A new lead event triggers the app. Code validates fields, normalizes company names, matches territory and assigns an owner. The model handles conflicts, such as a lead from a subsidiary of an existing customer. Ambiguous leads go to a queue with the reasons attached.
  • Account research and enrichment. The app pulls context from sources your team has cleared and stores each fact with its source and retrieval date. Code flags stale facts. The model reconciles conflicting ones, and the rep decides what to say to the buyer.
  • Follow-up drafting. After a meeting, the app gathers the CRM record, notes and open next steps. Template choice, merge fields and calendar slot lookup are code. Tone and objection handling go to the model. The rep approves every draft before it sends.
  • Pipeline hygiene. Rules find deals with no activity in a set number of days, missing close dates or stages that contradict the fields beneath them. The app logs before and after values for every proposed change. The model suggests a stage change or close-date push, and the owner accepts or rejects it.
  • Renewal and risk signals. The app combines usage and support data you permit it to read and applies scoring rules. The model reads the unstructured parts, such as a tense support thread, and drafts a note. High-risk accounts route to the owner with evidence attached.
  • Meeting prep and SDR to AE handoff. The app assembles recent activity, open opportunities and prior commitments with links to every source. The model condenses it and proposes next steps.

Pipeline hygiene rewards determinism most. The before and after log is what separates a cleanup from a corruption: if a proposal turns out wrong, the original value is still on record.

Build the lead-to-owner agent on Major

This is an illustrative build, not a customer deployment. It covers lead routing with an approval-gated follow-up, using Salesforce, Gmail and Slack.

  1. Connectors. Connect Salesforce for lead and account reads and task writes, Gmail for drafts, and Slack for owner notifications. Credentials are scoped per connector, so the app can read leads and accounts and write only approved fields and tasks.
  2. The app that holds state. A routing app on Major's managed database keeps a table of leads, rule outcomes, assignment history and enrichment results with source and status. A view for managers shows why each lead went where it did.
  3. The deterministic path. A Salesforce lead event triggers the app. A daily sweep catches anything the event missed. Code applies territory and qualification rules, assigns the owner, creates the CRM task and posts the Slack notification.
  4. The agent's judgment. The agent handles exceptions: conflicting rules, a likely duplicate of an account owned elsewhere, or enrichment data that contradicts the CRM. It routes the case and attaches its recommendation.
  5. Approval gate. Gmail drafts are held for the rep. Send permission stays off until your policy and templates are validated, and outreach never goes out on its own.
  6. Governance. SSO and role-based access come from the platform, and every transition is written to the audit log. A manager can see when a lead moved and which rule or agent decision moved it.

The model does less here, not nothing. The routine path is code with its own state, so cost is front-loaded when the agent builds the app and then flat as lead volume grows. Salespeople still own the conversation and the decision to send. Start building on Major and describe the routing rules in plain language to get the first version.

Where to start

Pick one high-volume, rule-bounded task. Lead routing and pipeline hygiene both qualify, because the rules can be written down and the output can be checked. Outbound sending is the wrong first project.

  1. Baseline the task today: time per lead and how often it goes wrong.
  2. Map the permissions each action needs, and grant read before write.
  3. Let the agent build the repeatable path as an app.
  4. Test bad data, duplicate leads and missing fields.
  5. Send outputs to a review queue instead of customers.
  6. Review the logs weekly, comparing what the rules decided with what the model decided.
  7. Widen scope one permission at a time.

This guide does not cover multi-currency quoting, forecasting or call coaching. Those have their own data requirements.

Use cases by tool

Step-by-step builds for the tools sales teams already run, each using the tool's own API.

Gong

HubSpot

Outreach

Salesforce (Sales Cloud)

Salesloft

ZoomInfo

Start building on Major

Build the lead-to-owner routing app on your own Salesforce data, with scoped credentials, a review queue and an audit trail you can read.

Related articles

Frequently asked questions

can ai agents do sales?
Yes, for bounded multi-step tasks such as lead routing, account research, follow-up drafting and pipeline cleanup. Repeatable steps belong in deterministic apps with logged state, and the model handles exceptions. Negotiation, pricing decisions and relationship judgment stay with salespeople, who approve anything that reaches a customer.
how do i create an ai sales agent?
Pick one rule-bounded task, map the permissions it needs, and have the agent build the repeatable path as a deterministic app that stores its own state. Test exceptions, stage outputs in a review queue behind human approval, and monitor the logs. Expand permissions one action at a time.
which ai agent is best for lead qualification?
Judge by criteria instead of rankings. Look for fit with your CRM, narrowly scoped read and write actions, durable state that records every routing decision, and audit logs a manager can inspect. Territory and qualification rules should run as code, with the model reserved for ambiguous leads.
can ai agents send sales emails automatically?
Technically yes, but human approval is the sound default. Start with drafts reviewed by the rep, and withhold send permission until templates, policy and validation are in place. Automate only narrow, well-tested cases afterward, with logs and a way to roll back.