What is an AI BDR? Workflow, controls and measurement

An AI BDR is software that uses AI to carry out parts of a business development representative's prospecting workflow: finding accounts, researching people, preparing outreach and managing early conversations. An autonomous agent can also take actions, such as sending messages or routing replies, within the permissions you give it. An assisted workflow leaves selected decisions with a person.

For a sales leader, founder or rep, the useful starting point is the job you want to delegate. Researching a prospect, approving a message and booking a qualified meeting are separate jobs. A product labeled “AI BDR” may cover some of them, while relying on your team or another system for the rest.

Extrovert covers the LinkedIn discovery-to-conversation workflow. It finds prospects, prioritizes comments and DMs, and prepares drafts from prospect context and your voice. You can review the work or delegate permitted actions to a connected AI assistant. That scope differs from an email-sending or meeting-booking platform.

For product-by-product capabilities and prices, use the AI BDR tools comparison. This guide focuses on designing the work, controlling execution and judging the result.

AI BDR, AI SDR and sales assistant: what changes?

BDR stands for business development representative; SDR stands for sales development representative. Teams use the human titles differently, often reserving BDR for outbound account development and SDR for outbound, inbound qualification or both. The AI labels do not define a standard set of capabilities.

Artisan calls Ava an AI BDR and describes lead discovery, enrichment, outreach and meeting booking. 11x calls Alice an AI SDR and describes prospect engagement, reply handling and qualified meetings. These are vendor descriptions of scope, not evidence that every AI BDR performs the same tasks or delivers the same results.

A writing assistant can produce a draft without having permission to send it. An autonomous agent combines decisions with tool access, so it can act on a prospect record, message queue or calendar. Check which actions a product can take, what triggers them and what still needs approval.

Operating mode	Who decides and acts?
Assisted	AI researches, prioritizes or drafts. A person reviews the output and approves the next action.
Partly delegated	The agent handles specified actions. A person reviews exceptions, selected accounts or sensitive replies.
Autonomous within a scope	The agent executes an approved workflow without per-action review, with boundaries, logs and escalation rules.

Autonomy can differ by step. A team might automate account research, approve first messages and hand every interested reply to a rep. Ask about the actual workflow rather than treating “human-in-the-loop” as a guarantee that a person sees every send.

Choose the first job to delegate

Pick the bottleneck you can observe in your current process. That makes the pilot easier to evaluate and defines when the job is complete.

• Research takes too long. Delegate account research and qualification notes first. The output is a reviewed list with sources and a reason each person fits, not a larger sending queue.
• Reps struggle to stay in touch. Delegate activity monitoring, draft preparation and follow-up reminders. Keep existing threads and account ownership visible so the next message continues the conversation.
• Interested replies wait for a rep. Start with reply classification and routing. Name the receiving rep and define when the agent must stop talking and hand over.

Before launch, record the current result, the main measure for your chosen job and the condition for continuing the pilot. Compare research time per reviewed prospect, follow-up effort per conversation or time from an interested reply to rep acceptance, depending on the job.

A bad list needs a targeting fix before more outreach. If drafts repeatedly need the same correction, improve the sales material or instructions. If good replies arrive but nobody owns them, fix handoff before increasing volume.

Give the workflow usable inputs

Before connecting an agent to a sending channel, define what a qualified prospect and a useful conversation look like. The agent needs enough material to decide whom to include and what it can truthfully say.

Prepare:

• An ideal customer profile: company characteristics, buyer roles, relevant problems and explicit exclusions.
• Account and contact data with source, freshness and existing ownership.
• Approved product claims, sales material, proof and messaging guidance.
• CRM context for customers, open opportunities, prior outreach and suppression requests.
• A handoff definition: who owns an interested reply, what information they need and when the agent must stop.

Separate account fit from interest. A person can match your ICP without having a current need. A reaction to a LinkedIn post can show attention to a topic without proving purchase intent. Use the buying-signals guide to distinguish a reason to investigate from a reason to pitch.

Keep customer nurture, new-account prospecting and open-deal follow-up separate when they require different rules. Otherwise, the same agent can send an acquisition message to an existing customer or interrupt a conversation another rep already owns.

Map the work from discovery to handoff

The workflow below is a design checklist. Confirm which steps your product supports natively, which require integrations and which remain with a person.

1. Discover and qualify prospects

Use a database, an approved imported list or activity signals to find candidates. Check company fit, role, exclusions and ownership before enrolling them. Keep the reason for inclusion beside the prospect record so a reviewer can inspect it.

In Extrovert, Lead Finder searches a B2B database by person and company criteria without using your LinkedIn account for that search. Intent Signals finds people behind selected LinkedIn activity and shows what each person did. Those are different inputs: one starts from fit, the other from activity you want to investigate.

2. Research and prepare a relevant message

Give the agent the source material it needs to draft a message: a prospect's post, company context, conversation history and approved sales content. Check that every personal reference comes from the right person and says what the source supports.

An agent should not infer budget, urgency or a buying project from a job title or a generic post. Keep uncertain conclusions out of the message. Review product claims, promised outcomes and the next step as carefully as the personalized opening.

Inspect a source-to-message example

The Conversation Restarters illustration shows a prospect's post about attending SF Tech Week. It suggests: “Seen you went to SF Tech Week. How did it go?” The message uses the event the person shared to reopen a conversation. It does not assume they have budget, need your product or want a demo.

Review that recommendation in three parts: does the post belong to the right person, does the thread make this a sensible follow-up, and does the draft stick to what the post says? A generic thought-leadership post may offer no useful reason to restart. Leave it alone rather than forcing a sales opening.

3. Execute only the allowed channel actions

Document whether a step creates a task, prepares a draft or sends a message. “Multichannel” can mean email plus LinkedIn actions, or a sequence that queues manual calls. It does not establish that the agent can autonomously execute every step.

For email, verify the sending infrastructure and provider requirements. Google's sender guidelines require SPF or DKIM for all senders to personal Gmail accounts, and SPF, DKIM and DMARC for bulk senders. Google requires keeping the Postmaster Tools spam rate below 0.3%. That rate is distinct from a study's broader “spam-flag” metric.

For LinkedIn, account limits and message review are operational controls, not permission from the platform. LinkedIn's User Agreement prohibits bots and unauthorized automated access and messaging. Neither a human approval step nor a low-volume setting guarantees that an account cannot be restricted.

4. Handle replies and hand off the conversation

Define separate routes for interest, questions, objections, out-of-office replies, opt-outs and messages the agent cannot interpret. Stop irrelevant follow-ups after a reply or suppression request. Escalate claims about pricing, security, contracts or commitments when they need a responsible person.

A booked calendar event is only one stage. Your team still needs to confirm fit, attendance and whether the meeting advances an opportunity. Specify the receiving rep, the conversation summary and the qualification evidence before treating handoff as complete.

Set review, permissions and stopping rules

Start with a narrow workflow whose decisions a reviewer can inspect. Expand the scope when the targeting, drafts and reply routes behave as intended.

Write down which actions the agent may perform: read records, enrich contacts, edit drafts, enroll prospects, send messages, update CRM fields or create calendar events. Grant only the access needed for that workflow, and confirm how to revoke it.

Keep a review step for high-value accounts, unsupported claims, sensitive information, ambiguous replies and commitments to the prospect. Log the source context, draft, approval or delegated permission, executed action and subsequent response.

Check the stopping behavior before scaling:

• What happens after a reply, an opt-out or a disqualified lead?
• Does a campaign pause cancel queued actions, or only prevent new ones?
• Can the agent change targeting, cadence or sending limits without asking?
• Who sees failed actions and messages waiting for review?
• Can you trace a sent message back to its source and permission?

Turn the rules into a pre-launch test. Check classification and routing with internal records in draft-only mode. Test queue behaviour and sending permissions separately in a sandbox or with an owner-controlled internal recipient:

1. Give it an existing-customer record. Confirm it stays out of new-account acquisition.
2. Add an opt-out. Confirm it suppresses future outreach, including queued follow-ups.
3. Give it a security or contract question. Confirm it routes the question to the named person instead of inventing an answer.
4. Pause the workflow with an action pending. Check the pending queue and cancel anything you no longer want sent.
5. Revoke its sending permission. Confirm an attempted send is rejected; a draft-only test cannot establish that.

A failed stop or permission test is a reason to keep sending disabled while you fix that control. Review and delegation govern what your team sends. They do not establish deliverability, legal compliance or platform authorization on their own.

Use Extrovert for LinkedIn discovery and conversations

The Extrovert product overview shows a workflow built around prospects and relationships: discovery, relevant comments, connection requests, DMs and follow-up. Lead Finder and Intent Signals supply candidates. Campaign settings and your Voice guide the priorities and drafts.

For warm outreach, engage with a prospect's relevant content before making a direct ask when that context warrants it. Existing conversations need their own follow-up rules. Read the history rather than resetting everyone into the same first-touch sequence. The LinkedIn DM workflow explains the timing, unread replies, skips and unanswered follow-up limits in more detail.

Review the reason a DM is due

Recommended DMs brings conversations into a daily queue with a reason and a prepared draft. The illustration shows first outreach, follow-up and restarter reasons beside the conversations, with Send or Snooze controls in the desktop view. Read the reason and thread before deciding whether to send the message today.

Choose what you review or delegate

Extrovert supports manual review and optional automatic commenting for selected campaigns. Through MCP (Model Context Protocol), a connection that gives an AI assistant access to Extrovert tools, a compatible assistant can read priorities and conversations, edit drafts, and approve or send actions you allow. Define what the assistant may execute and what it must bring back to you.

The product overview includes a concrete connected-assistant exchange: “What needs attention today? Approve the comments you're sure about and show me the rest.” The assistant reads four comment drafts and two due DMs, approves three comments and brings one back for a decision. The user then says: “Skip it. The first DM: shorter, then send it.”

That exchange separates delegated approval from an explicit send instruction. For your own workflow, define what “sure about” means: source-backed, relevant to the campaign and within approved claims. Keep pricing promises and sensitive replies in review.

That gives you control over the LinkedIn workflow without requiring every task to begin from scratch. It does not turn Extrovert into a cold-email sender or establish native autonomous meeting booking. Keep email and calendar responsibilities explicit in your wider sales process.

To apply this on LinkedIn, start with one campaign, set your Voice and review the first suggested comments and DMs. Check the source context and due reasons before widening what you delegate.

Measure the outcome, not the amount of activity

Judge the workflow against the job you assigned it. Research automation should produce qualified records with usable evidence. Outreach automation should produce relevant conversations and qualified opportunities, with review effort and channel problems visible alongside them.

Track results for a defined account cohort and observation window:

• Qualification accuracy and the reasons people were rejected.
• Positive replies and useful conversations, separated from any reply.
• Meetings held and accepted opportunities, separated from meetings booked.
• Review time, edits, escalations and failed actions.
• Email bounces, complaints and opt-outs, or LinkedIn warnings and restrictions.
• Software, data and human-review costs relative to the outcomes you value.

Compare cohorts with similar fit, channel, offer and sending conditions. Do not compare a cold-email reply rate with a LinkedIn comment response and call the difference an AI performance effect. If you change the list, offer and cadence at the same time, a before-and-after result will not isolate which change mattered.

Run a pilot with a usable scorecard

Choose one segment, one offer and one workflow. Record the starting list, the assigned rep and the permission settings. Review the first qualification notes and drafts before expanding execution. Keep the same definitions throughout the pilot and give meetings time to happen before evaluating the result.

Measure	How to record it	What it helps you decide
Qualification accuracy	Reviewed prospects that meet your rules ÷ all prospects reviewed	Whether to fix discovery or exclusions
Positive-reply rate	Unique people with a relevant positive reply ÷ unique people contacted	Whether the targeting and message open useful conversations
Held-meeting rate	Unique contacted people who attended a qualified meeting ÷ unique people contacted	Whether outreach produces meetings your team can use
Handoff completion	Interested conversations accepted by the named rep ÷ interested conversations routed	Whether replies are reaching an owner
Review effort	Minutes on checks, edits and exceptions per reviewed prospect or handled conversation; setup time recorded separately	Whether delegation saves work or moves it into a review queue
Cost per accepted opportunity	Allocated software, data and human-review cost ÷ opportunities accepted by sales	Whether the workflow is worth continuing

Count each person once in a reply rate, even if you sent several messages. Log out-of-office replies, opt-outs and objections separately. Keep booked meetings, held meetings and accepted opportunities in separate columns. If the denominator is zero, leave the rate unreported rather than calling it zero performance.

Set your pass criteria against your team's current workflow before launch. No universal reply-rate target can settle whether the agent works for your segment. Continue when targeting is accurate, stop rules work, handoffs reach a rep and the cost of useful outcomes fits your business. Revise the message or targeting when the relevant measure fails; stop execution immediately when suppression or permission controls fail.

What does published AI outreach research tell you?

Digital Applied reports an analysis of 100,000 paired cold emails: 50,000 AI-generated and 50,000 human-written, matched on factors including persona, company fit and sender-domain age. It reports reply rates of 4.1% versus 5.2%, and meeting-booked rates of 0.7% versus 1.1%. Those figures describe that publisher's email dataset; they do not measure every AI BDR or the LinkedIn engagement workflow.

The report also uses a “spam-flag” measure covering recipient-reported or filter-detected classification. It is not interchangeable with Gmail's user-reported Postmaster Tools spam rate. Neither measure shows that manually approving an AI-written message removes deliverability risk.

Start with one delegated job, a named owner and a scorecard agreed before launch. Expand the workflow when it produces the conversations your team can use, with costs and failures you can account for.