Sales and marketing in an agentic-AI-first company

"What does outbound look like without an SDR team?"

That's the question a founder of a 12-person infra-tools company asked me last month. He'd done the math. Two SDRs at $80K each plus management plus tooling was $200K all-in for a pipeline they couldn't reliably forecast. He had heard "AI is going to replace SDRs" so many times he'd stopped listening. He wanted to know what the actual team looked like on the other side. Role by role. With the hand-offs named.

So here it is. The sales-and-marketing team inside an agentic-AI-first company. Seven named agent roles, the hand-offs between them, and the specific places where a human is still the right answer.

TL;DR

The decomposition: an SDR is actually six jobs

Before naming the agents, name the work. A traditional SDR does six distinct things in a day.

  1. Source prospects (search, list-build, scrape, filter).
  2. Enrich each prospect (find email, check ICP fit, gather context).
  3. Draft a personalized opener.
  4. Send it.
  5. Read and reply to inbound replies.
  6. Hand off the hot ones to AE / founder for the close.

Jobs 1, 2, 3, 4, and 5 are mostly text-and-data work. AI handles them well when wired with the right verification layer underneath. Job 6 still needs a human, at least for high-ticket closes. The maths is brutal: if AI can absorb 5 of the 6 jobs and a human still needs to do job 6, the SDR role isn't replaced, it's decomposed and most of the parts get re-assigned.

That decomposition is the model.

Role 1: Content Writer, "produces the inbound surface area"

What it owns: writing one substantive post per day on the company blog, plus weekly long-form on a paid channel (Substack, gated newsletter) when relevant. Long-tail SEO, AI-engine optimization, evergreen pages.

What it does NOT own: distribution to other channels (the Distributor does that), paid promotion (the Ads Manager does that), syndication formatting (a sub-agent does that).

Concrete: at OperatorIQ, the Blog Writer is one Sonnet-driven cycle that runs daily at 06:00 ET. It reads a topic queue, picks the highest-EV post, drafts, lints, publishes to the site. Output: one MDX file pushed to GitHub, which auto-deploys to operatoriq.io.

Hand-off: writes a DISTRIBUTION_TRIGGER_<slug>.md file when each post ships. The Distributor picks it up on its next cycle.

What changes when you wire it this way: the company starts producing the same volume of content as a 3-person content team. The cost is mostly model spend (~$30/month for ~30 posts) plus the upfront build (~2 weeks of engineering).

Role 2: Distributor, "pushes content to every channel"

What it owns: every channel where the post should also appear. Dev.to, Bluesky, IndexNow, Hashnode, relevant subreddits (when Reddit OAuth lands), LinkedIn (via API), Substack Notes (via Playwright, no public API).

What it does NOT own: writing the content. It re-formats and re-frames per channel; it never writes from scratch.

Concrete: at OperatorIQ, the Distributor is a Haiku sub-agent. It reads every DISTRIBUTION_TRIGGER_<slug>.md file in the autonomous folder, generates channel-specific copy (a 280-char Bluesky hook is not a 1,500-word dev.to syndication), posts via each channel's API, and logs the result.

Hand-off: writes back to a distribution_log.jsonl so the Analyst agent can correlate channel performance against pipeline.

What changes when you wire it this way: content syndication stops being a job that someone has to "find time for." Every post hits every channel it should within an hour of publish. Indexing happens faster because IndexNow gets pinged automatically.

Role 3: Lead Sourcer, "finds the prospects"

What it owns: finding new prospects that match ICP. Pulls from Apollo, LinkedIn (via Apollo + manual cookie sessions), Crunchbase, public Stripe customer lists, Upwork job posts, sub-Reddit threads where the buyer self-identifies their pain.

What it does NOT own: drafting outreach. It produces enriched contact rows and hands them to the next agent.

Concrete: at OperatorIQ, the Lead Sourcer runs every morning at 08:00 ET. It pulls 50-200 fresh prospects, scores them against ICP (industry, headcount, signal indicators like "we're hiring an Ops person"), filters to the top 25, and writes a lead_tracker_<date>.csv file.

Hand-off: the CSV is the input to the Outreach Closer's morning cycle.

What changes when you wire it this way: the founder stops being the bottleneck on "who do we email this week." The team has 25 fresh, qualified, enriched prospects in queue every morning by 9am. The cost is Apollo's API plan (~$200/month) plus the agent itself (~$20/month in model spend).

Role 4: Outreach Closer, "drafts and sends"

What it owns: drafting personalized cold outreach (email + LinkedIn), sending it inside daily quotas, reading and triaging replies, escalating hot ones.

What it does NOT own: the close itself when the prospect wants a real conversation. Anything above ICP-defined deal size still routes to a human. Anything that asks about specific terms, custom pricing, contracts.

Concrete: at OperatorIQ, the Outreach Closer pulls the morning lead CSV, drafts a personalized opener per prospect using the Voice Calibrator's profile (so it sounds like the founder, not like AI), lints each draft against the founder's rules (no em-dashes, no banned phrases, no jargon), sends up to 25 per day via authenticated SMTP, and watches the inbox for replies. Replies under a confidence threshold get queued for the founder; replies above the threshold get handled directly with a follow-up.

Hand-off: writes to a pipeline.csv file with every prospect, every touch, every status. The founder reads this end-of-day.

Authority envelope: can send up to 25 emails per day, max 3 follow-ups per prospect, can never offer a discount, can never schedule a call without escalating.

What changes when you wire it this way: outbound stops being a job that someone has to grind through. The pipeline runs every day whether or not the founder is at her desk. Reply quality is the metric, not volume.

Role 5: Voice Calibrator, "makes the agents sound like a real person"

What it owns: mining the founder's actual sent folder weekly, extracting the voice profile (avg sentence length, contractions, openers, closers, warmth markers, banned phrases), publishing the profile as a reference file that every other customer-facing agent reads on each cycle.

What it does NOT own: writing. It produces the calibration target, not the copy.

Concrete: at OperatorIQ, the Voice Calibrator runs every Sunday at 06:00 ET. It pulls a stratified sample from the founder's Outlook Sent folder via COM, extracts patterns, writes christine_voice_profile.md. Every other agent reloads the profile on cycle start.

Why this role exists: the moment your agents sound like AI in a prospect's inbox, the whole stack falls apart. The reply rate drops to near-zero and the founder's reputation takes a hit. This role is what prevents that.

What changes when you wire it this way: the AI-flavored phrasing that gives away an agent ("I'd love to leverage", "delve into", "robust solution") gets caught at the source. Drafts come out reading like the founder typed them, because they were calibrated against what she actually typed last week.

Role 6: Sales QA, "catches the mistakes before they ship"

What it owns: reading every outbound draft before it sends. Linting against rules (no banned phrases, contains the right CTA, links work, the prospect's name is actually correct). Flagging anything weird for human review.

What it does NOT own: writing. It only reads and approves or kicks back.

Concrete: at OperatorIQ, the Sales QA is a 100-line Python linter plus an optional Haiku sub-agent for the high-confidence drafts. The linter is the floor (cheap, fast, catches 80% of issues). The sub-agent is the ceiling (catches the 15% the linter misses, costs ~$5/month in model spend).

Hand-off: anything that fails QA gets logged to qa_failures.jsonl with the reason. Anything that passes goes to the send queue.

What changes when you wire it this way: the rate of "oh god I can't believe we sent that" drops to near-zero. The Outreach Closer can run at higher volume because the QA layer catches what the human would have caught.

Role 7: Human, "closes the deals worth closing"

What it owns: the calls, the demos, the contract negotiations, the "I want to talk to the founder" moments. The strategic decisions about ICP shifts, pricing changes, new offers.

What it does NOT own: the upper-funnel grunt work. The CRM updates. The reply triage. The list building. The follow-up scheduling.

Concrete: the founder spends ~30 minutes a day on sales. She reads end-of-day pipeline status, handles any escalations, replies to any prospect who asked specifically for her. Everything else routes through the agents.

Why this role still exists: people who are about to spend real money still want to talk to a real person. They want to hear conviction. They want to ask "what happens if it doesn't work in 60 days." That conversation needs a human. If you try to fully automate it, your close rate drops by half because you stripped out the moment where the prospect decides they trust you.

The hand-off map, all seven roles in one paragraph

The Content Writer ships a daily post (handoff: DISTRIBUTION_TRIGGER file). The Distributor reads the trigger and pushes to channels (handoff: distribution_log update). The Lead Sourcer fills the morning queue (handoff: lead_tracker CSV). The Outreach Closer pulls the queue, drafts via Voice Calibrator's profile, lints via Sales QA, sends, watches replies (handoff: pipeline CSV). Voice Calibrator runs weekly to keep the voice profile fresh. Sales QA reads everything before it ships. The human reviews end-of-day and takes the calls worth taking.

Total monthly cost for this stack at our scale: under $400, mostly Apollo subscription and model spend. Replaces what was a 4-to-6-person GTM team.

Where most teams get this wrong

The single biggest failure mode: replacing the SDR without building the Voice Calibrator. The agents draft outreach. The outreach sounds like AI. Reply rates collapse. Founder concludes "AI doesn't work for sales." Most companies that pulled the plug on agentic GTM in 2024 and 2025 pulled it at this exact point.

The second failure mode: skipping Sales QA. The agents draft, send, and 1-in-30 messages goes out with the wrong name or the wrong reference. Each one is a small thing. Compound them across 500 sends and you have a pattern. Always run QA between draft and send.

The third failure mode: trying to remove the human from job 6 (close). Don't. For deals under ~$2K-$5K the agent can carry the whole flow. Above that, the human shows up at the close, even if every other touch was the agent. The economics still work because the human only shows up for the 5% of deals worth closing, instead of for every cold opener that's never going to convert.

If you want this built

This is what we ship. The seven-role agentic GTM stack, in seven days, for a flat fee. The agents, the voice calibrator, the QA layer, the integrations to your CRM and your Apollo. See our blueprints for the full scope and the price.

What to read next

If you got value from this, the cornerstone post is What is an agentic-AI-first business?. The infrastructure piece is the 5 layers of an agentic AI stack. The maturity model is from copilot to colleague.

Coming next in this series: customer support reimagined, the autonomous CS agent. The role most teams should automate first, why, and the exact envelope that keeps it safe.

If you want to talk about what your sales team looks like on the other side of this, email christine@operatoriq.io. Tell me how many SDRs you have today. I'll tell you what the agentic version of your team looks like.

Cheers, Christine