How to staff AI agents as a GTM roster (not one mega-bot)

One mega-bot that "does GTM" fails.

Founders buy a single agent pitched as research, enrich, write, send, book, and follow up. It works in the demo. In production it sprays. Wrong titles. Soft signals. First lines that could go to anyone. Domains heat up. Calendar stays empty.

One mega-bot that "does GTM" is a spam factory with a nicer UI. Staff a roster; humans keep the knife.

What is a GTM agent roster?

A roster is a set of narrow agents with named jobs and named handoffs.

Think sports, not magic. Each role has a lane. The play still needs a coach.

AI as operator means tools run inside rails. A roster is how you draw the rails for agents instead of hoping one prompt holds the whole motion.

Which roles belong on the first roster?

Start with four. Add later only if these stay clean.

1. Signal watcher. Watches de-anon (RB2B, Vector), G2 category hits, competitor LinkedIn engagement, PostHog stuck states, Factors.ai multi-page paths. Emits fit + intent + recency. Does not send.
2. Enricher. Adds title, stack clues, email, LinkedIn URL after the signal clears. See enrichment in outbound. Does not invent intent.
3. First-touch drafter. Writes a signal-matched note for humanson, signal type, timestamp, fit score.

Enricher attaches: verified email, one stack clue, one recency fact.

Drafter returns: three-line note, ask for a meeting, no feature dump. Human edits or kills.

Logger closes the loop: sent / held / no-reply / rage. Kill-switch watches bounce and complaint.

If any step can fire the next without the gate, you rebuilt the mega-bot in four costumes.

What should I refuse to automate first?

Refusal list for week one.

• Final send without human eyes on a new template
• Thaw after a kill-switch trip
• Pricing and discount language
• Claims you have not published on the site

Automate watching and drafting. Delay autonomy on trust and money.

Would you rather four boring agents that book meetings, or one flashy bot that "handles GTM" while your reply rate dies?

Adapt or fail. Staff the roster. Keep the knife human.

Start Signals, Convert, Grow

FAQ

Can one model power all four roles?

Yes, as separate prompts and tools with separate permissions. Same model, different jobs and gates.

When do I add a fifth agent?

When a role has clear volume and a failure mode the others should not share. Example: deliverability watcher separate from drafter.

Do I need Clay for this?

No. Clay can help enrichment waterfalls. The roster is the design. Tools are interchangeable.

How does this fit Signals → Convert → Grow?

Signals own watcher and enricher. Convert owns message and meeting ask. Grow owns spend on the agents that produce held meetings, not activity.

What is the first metric?

Held meetings per week from signal-matched sends. Not drafts generated. Not tokens burned. approve or tight template with human QA sample. Never auto-sends on day one. 4. Meeting logger. After a book or a no, logs source, template, outcome. Feeds the kill-switch and the weekly review.

Human owns message final and the kill-switch. That is not optional.

Why does the mega-bot pattern break?

Because GTM is a system with different failure modes.

Signal quality fails differently than copy quality. Enrichment fails differently than deliverability. One agent optimized for "activity" will maximize sends. Sends are not meetings.

Ehrenberg-Bass / LinkedIn B2B Institute: about 5% in-market. A mega-bot does not create that 5%. It can burn the domain while talking to the 95%.

Gong's cold email floor is ugly enough without an agent that never sleeps and never doubts.

How do I staff the roster without a huge team?

Founder plus one operator is enough to start.

• Map each role to a tool or a prompt file, not a headcount.
• Shared inbox or Slack channel for "draft ready" and "switch tripped."
• Weekly 20 minutes: which signal source produced held meetings? Demote the rest.

This is signal-based outbound with agents as sensors and scribes, not as your VP Sales.

What does a clean handoff look like?

Signal watcher posts: account, per