AI · Prospecting · 2026 An applied workflow talk

A talk in fourteen panels

From cold to connected.

Prospecting is now an operations discipline. Here’s what changed — and what one operator running the new pattern looks like.

Presented for business leaders Today
From cold to connected · 2026

Where the hours go

70%
of a typical sales-development rep's week is spent on activities that aren't selling — research, list-building, data entry, scheduling, follow-up admin.

Industry composite. Numbers vary by org but the shape is consistent: the people you hired to sell are mostly assembling.

02 / 14

A short history

It used to be wishful.
Now it’s a workflow.

2018
AI suggests subject lines.Predictive text. Marginal lift.
2022
AI writes the email.ChatGPT. Useful, still single-shot.
2024
AI completes the task.Tool use. Function calling. Agents start to act.
2026
AI runs the workflow.End-to-end execution. Humans at the decision gates.
03 / 14

Anatomy of the stack I run

Two agents. One operator.
Distinct roles.

01 — Cognitive
Claude Code
Judgment, copy, narrative, decision-making. Drafts the email, scores the lead, writes the brief. The agent doing most of the thinking-shaped work.
02 — Verification
Codex
Independent code review. Audits the scripts that run the pipeline. Catches the bugs the first agent would ship past you. The audit layer.
03 — Accountability
Operator
Strategy and ratification. Picks the ICP, sets the voice, owns the relationship. The agents pull the operator out of execution and put them back into judgment.
04 / 14

Before

How prospecting used to run.

01 Define ICP Human
02 Source leads Human
03 Enrich data Human
04 Personalize copy Human
05 Send outreach Human
06 Track replies Human
07 Hand off to sales Human

~40 hours per week per SDR. ~6 of those hours are selling.

05 / 14

After

Same seven steps. Different operator.

01 Define ICP Human gate
02 Source leads Agent
03 Enrich data Agent
04 Personalize copy Agent
05 Send outreach Agent
06 Track replies Agent
07 Hand off to sales Human gate

~3 hours per week of operator decisions. The other 37 are agent execution.

06 / 14

Before I show this working

Three things it won’t do.

LIMIT 01
It won’t fix bad targeting. It amplifies it.
Two hundred touches to the wrong segment is worse than thirty to the right one — because the agent will keep going until you stop it. ICP definition is the operator’s job, and it doesn’t get delegated.
LIMIT 02
It won’t sound like you on day one.
Models converge on a flavorless corporate-AI register unless you actively train them against it. Three or four rounds of voice review during onboarding is the cost of getting the model out of the median.
LIMIT 03
It won’t ship claims you haven’t approved.
The agent that drafts is not the agent that reviews. The operator makes the final call on every artifact that leaves the system. In regulated industries this is non-negotiable — and the architecture is designed around the limit, not around hiding it.
07 / 14

A discovery cycle, end-to-end

Nine minutes. From signal to drafted outreach.

00:00
Trigger event surfaces. News, filing, hire, expansion — any public signal worth a touch
00:45
Firm enriched. Website, leadership, recent news, geo, sector
02:30
ICP scored. Tier assigned. Star · Cash Cow · Question · Dog
04:00
Personalized microsite generated. Static page, firm-specific framing
06:15
Outreach drafted with three voice variants. A/B/C cohort, deterministic distribution
08:30
Queued for operator review. Quality audit gates: 89% kill rate, by design
09:00
Operator approves. Send queued for batch. Same task that used to take 90 minutes
08 / 14

What lands in the inbox

The artifact, not the chore.

From · Operator Personalized · v3
On your Q3 expansion — a quick observation

Saw the announcement on Aug 14. Quick note that’s probably already on your radar: most teams making a move at your scale typically leave material value on the table in the first six to twelve months — not because the math is hard, but because the operational layer isn’t usually packaged for teams in your size band.

I built a twenty-second view of what that gap looks like against your specific situation, no signup: example.com/[your-firm]

→ Reply “walk me through it” if useful.
Trigger
Pulled from a specific public signal — not from a list.
Personalization
A custom view built against that specific firm — not a token in a template.
Voice
Drafted in the operator’s register. No “I hope this email finds you well.”
Ask
Small. One-line reply. No 30-minute calendar pitch.
09 / 14

The decision gates

Two places where humans still win.
That’s the whole list.

Gate 01 · Targeting

Who are we chasing, and why?

The agent will faithfully pursue a bad ICP at industrial scale. Volume is not strategy. The operator’s job is to choose the segment the agent then executes against — and to revise it when the data says it’s wrong. Humans pick the slope; agents climb it.

Gate 02 · Voice

Does this sound like us, or like AI?

The recipient should recognize the firm in the first sentence. The agent gets close; the operator closes the gap. Three rounds of voice review during onboarding is usually enough for the model to stop sounding like everyone else’s outbound stack.

10 / 14

Cost per qualified meeting

One operator. Seven times the surface area.

Traditional · 1 SDR · 1 week 30 touches $130K loaded comp ÷ 1,560 touches/yr ≈ $83 per touch
Agentic · 1 operator + stack · 1 week 200+ touches ~$200/mo software · operator focused on the 3-hour judgment slice
Net effect on cost per qualified meeting −65% to −80% Range. Depends on conversion. The agent doesn’t fix bad ICP — it amplifies whatever’s there.
11 / 14

How to start without betting the firm

The trust ratchet.

I Read-only The agent surfaces opportunities, drafts notes, builds lists. A human ships everything. Risk: near zero. Yield: ~30% of the eventual value. Weeks 1–4
II Drafts for approval The agent prepares the artifact end-to-end — copy, microsite, queue. A human reviews each batch before send. Risk: low. Yield: ~70%. Weeks 5–12
III Bounded autonomy The agent ships within an explicit “green lane” — categories of action it’s allowed to execute without prior approval. Human reviews after the fact. Risk: managed. Yield: full. Month 4+
12 / 14

Where your motion is agent-shaped

Two axes. Four answers.

High volume · Low variability First to delegate. Research, enrichment, draft assembly, list dedup, follow-up scheduling. The agent runs these unsupervised once voice is tuned.
High volume · High variability Drafts with mandatory review. Personalized outbound to mid-market accounts. Agent does 90% of the work; human spends ten minutes per batch confirming voice and fit.
Low volume · Low variability Automate, but skip the agent. Rules engines and templates are cheaper, faster, and more auditable than an LLM. If the answer is the same every time, you don’t need a thinker.
Low volume · High variability Keep the human. Strategic accounts, executive communication, anything bespoke. The agent assists; it does not lead. Cost-per-touch is irrelevant here.
13 / 14

In closing

Less time on the chase.
More time on the conversation.

That’s the whole point.

Thank you.   ·   Questions →

14 / 14
01 / 14