Where should the line sit between human reps and AI agents?

Last month I sat in on a discovery call with a global enterprise software company. They had 90 BDRs. Each one was swivel-chairing between 5 to 8 tools a day. The BDRs didn't trust the contact data, the intent tools disagreed with each other, and plenty of them had quietly started writing their own AI emails outside the cadences.

Their question wasn't "should we use AI?" They were already using it, badly. Their question was the one every sales leader I talk to is asking right now: who should be doing what?

Here's my honest answer. The line between people and AI agents is not a philosophy question. It's a data question. And most teams can't answer it yet because they don't know what each account is worth.

Why do sales teams need a line between people and AI at all?

Without a clear line, reps and agents step on each other. Reps ignore accounts the agent is working, agents email accounts a rep is already talking to, and nobody owns the result. A written line tells everyone, people and software, which accounts are theirs and when to hand off.

The old model was simple. Hire more SDRs, give each one a list, measure activity. That time is done. Inboxes are flooded, buyers do their research alone, and adding headcount adds cost faster than pipeline.

But "just let AI do it" breaks too. Picture an agent sending a generic sequence to your biggest prospect while the account executive is mid-negotiation. That's not an AI problem. That's a missing rule.

The fix is a lane system:

• The human lane. High-value accounts, live deals, and anyone who has replied, booked or asked for something.
• The AI lane. The long tail: accounts worth working but not worth a rep's week. The agent researches, drafts, follows up and watches for signals.
• The handoff. A written rule that moves an account from the AI lane to a person, fast, with the context attached.

How do you decide which accounts get a person?

Rank every account by revenue potential, what it could be worth to you, then set a cutoff. Accounts above the line get a named rep. Accounts below it start in the AI lane. Any account that shows a strong buying signal jumps to a person, no matter where it ranked.

This is why we almost always start with revenue potential. If you don't know what an account could be worth, you can't decide who should work it. We wrote a whole answer on what revenue potential is and why it beats click-based scoring, and another on how to calculate it for each account.

Once every account has that number, the line becomes a simple question: how many accounts can each rep work well? Divide your top accounts across your reps until each has a book they can actually cover. Everything below that line goes to the agent. Our answer on designing sales territories that scale covers how to size those books.

A few rules I use to set the cutoff:

• Start conservative. Put more accounts in the human lane at first. Move the line down as you trust the agent.
• Use the client's data, not a vendor default. Your best customers tell you where value sits. A generic model doesn't.
• Revisit it every quarter. As reps ramp and the agent proves itself, the line should move.

What should AI agents do, and what should stay with people?

AI agents should do the reading, researching, drafting and watching: account research, first-draft emails, follow-ups, meeting reminders and signal monitoring. People should own relationships, live conversations, pricing, negotiation and anything where a wrong word costs a deal. When in doubt, the agent drafts and a person approves.

At InCheck, we set it up exactly that way. The AI drafts follow-up emails, and a human approves them before anything goes out on its own. Each sender is capped at 20 emails a day, so nothing scales faster than the team can review. Routing is automated, so the fastest-response leads get worked first.

Anthropic's own guidance on building effective agents makes the same point from the engineering side: choose the simplest setup that works, and build in checkpoints where an agent pauses for human feedback. In sales, that checkpoint is usually "a person approves before a customer sees it."

A useful way to split the work:

• AI lane jobs: research the account, find the right contacts, draft the first email, send follow-ups, book reminders, watch for hiring, funding or website signals.
• Human lane jobs: discovery calls, demos, pricing, contracts, escalations, and every reply that needs judgment.

When should an AI agent hand an account to a rep?

Hand off the moment a real person engages: a reply, a meeting booked, a pricing page visit from a known contact, or a request the agent can't answer. Set a clock on it too. If the agent can't book a meeting within a set window, the account goes to a person instead of waiting forever.

At Terraboost, the AI SDR answers inbound leads in under 5 minutes. If it can't book a meeting inside a 12-hour, 3-touch window, the lead escalates to a human rep. That rule is the whole system. Without it, good leads sit with a bot that keeps trying.

The handoff also has to carry context. When the account lands with a rep, they should see what the agent sent, what the buyer said and why it escalated. Otherwise the rep starts cold and the buyer repeats themselves.

Speed matters on both sides of the handoff. In a technical assessment we ran, the goal was a response in about five minutes, with automatic reassignment when the owner is out. You can't hit that if reps don't log activity, so we capture timestamps from the dialer and email systems instead of relying on manual entry. Our answer on B2B speed to lead goes deeper.

What results can you expect from an AI lane?

When the lane is set up well, the AI handles volume at a fraction of the cost, and people spend their time on accounts that can actually close. The best proof is cost per meeting and closed revenue, measured separately for each lane so you can see what each one really produces.

At Terraboost, the numbers made the case on their own:

• The AI Email SDR helped close $109K in won business.
• The team booked 246 meetings.
• Cost per meeting was about $150 with a human SDR and about $10 with the AI email SDR.

Those results came after we consolidated about 16 workflows into 2 engines and cleaned up a pipeline where 77 of 82 open deals had no next step. The agent didn't fix a broken process. The cleanup did. The agent then ran on top of it. See the full Terraboost case study.

That's the pattern I'd copy. Fix the data and the process, draw the line, then let the agent work the long tail. If you do it in the opposite order, you just automate the mess.

Frequently asked questions

Will AI agents replace SDRs?

Not the good ones. AI agents replace the repetitive parts of the SDR job: research, first drafts, follow-ups and list building. People are still better at live conversations, judgment calls and building trust. Most teams we work with keep fewer, more senior reps and let agents cover accounts that never got attention before.

How do we keep AI emails from hurting our sender reputation?

Start small, authenticate every sending domain, keep volume per sender low and make unsubscribing easy. Google's email sender guidelines spell out what large senders need, including authentication and one-click unsubscribe. Add human approval on early messages so a bad draft never reaches a buyer.

Should the AI lane be email only?

No. Email is usually the first channel because it's the easiest to measure, but the same lane can include LinkedIn, ads and website follow-up. Surround the right accounts on several channels when a signal says they may be ready, instead of blasting thousands of contacts through one channel.

Who should own the line between the lanes?

A sales leader should own the rule, with RevOps maintaining the data behind it. The rule should be written down, reviewed every quarter and visible to reps. When reps can see why an account is in their book or the agent's, they stop arguing about it.

What does IT need to approve?

IT usually wants to know where data goes, which AI keys are used and who can turn it off. We build in the client's own systems and accounts, so nothing holds them hostage. The NIST AI Risk Management Framework is a helpful shared checklist for that conversation.

Sources

• A global enterprise software company, discovery call, 2026: 90 BDRs working across 5 to 8 tools, distrust of contact and intent data, and the need for a human lane and an AI lane.
• Terraboost, 2025 to 2026: AI SDR replies in under 5 minutes, the 12-hour, 3-touch escalation window, $109K won, 246 meetings, about $150 versus about $10 per meeting, about 16 workflows consolidated into 2 engines and 77 of 82 open deals with no next step. See the Terraboost case study.
• InCheck, 2026: AI-drafted follow-up approved by a person before sending, automated routing and a cap of 20 emails a day per sender.
• RevOps Shop technical assessment, 2026: a response goal of about five minutes, reassignment when the owner is out, and timestamps captured from dialer and email systems.
• Anthropic, Building effective agents: use the simplest solution that works and pause for human feedback at checkpoints.
• National Institute of Standards and Technology, AI Risk Management Framework.
• Google, Email sender guidelines: authentication and one-click unsubscribe for bulk senders.