# The AI SDR model

> What an AI SDR automates and what it cannot — autonomy levels from drafts-for-approval to supervised autonomy, and the limits that keep humans in the loop.

**An AI SDR automates the mechanical span of outbound — list building, research, drafting, sending, reply triage, and meeting booking — while strategy, offer, and edge-case judgment stay human. Autonomy is graduated — drafts for approval first, then supervised sending, then autonomous operation that escalates anything uncertain.**

_Status: Reviewed — human-edited, facts not yet confirmed against the product._

## What it is

An AI SDR is software that runs the outbound loop a human SDR runs: build a prospect
list, research each person, draft a personalized message, send it on a schedule, triage
what comes back, and book the meeting. What it does not supply is the strategy
underneath — who to target, what to offer, whether the product fits the market, and what
to do when a reply falls outside every pattern it knows.

## Why it works this way

Most of an SDR's day is pattern work: find people who match a profile, gather the same
few facts, adapt a template, log what happened. Language models make each of those steps
automatable — but they are probabilistic, and occasionally confident and wrong. So
autonomy is granted in levels, each one widening what the system may do alone:

- **L1 — drafts for approval.** The system prepares everything; a human approves each message before it sends.
- **L2 — sends with a review queue.** Routine messages go out automatically; flagged ones wait for review.
- **L3 — autonomous with escalation.** The system runs end to end and escalates only edge cases — unusual replies, sensitive accounts, low-confidence drafts.

## How it behaves in practice

- **Approval gates exist because of hallucination risk.** A fabricated detail in a "personalized" opener is worse than no personalization — review catches what confidence scores miss.
- **The quality floor is the data.** Research and drafting inherit the accuracy of the underlying data; enrichment errors become personalization errors, at volume.
- **Autonomy is earned per playbook.** Teams typically start at L1, watch precision for a few weeks, and widen autonomy where the system proves reliable.
- **Escalation keeps humans on exceptions.** The human role shifts from writing every message to judging the cases the system cannot.

## Common misconceptions

| Belief | Reality |
| --- | --- |
| "An AI SDR replaces sales strategy" | It executes strategy; targeting, offer, and positioning remain human inputs |
| "Full autonomy is the goal" | Matching autonomy to proven reliability is — some steps stay gated on purpose |
| "AI personalization is always accurate" | Generated details can be wrong; that is why approval gates exist |
| "Good AI fixes bad data" | Output quality has a floor set by list and enrichment quality |

## How Reply implements this

Jason AI is Reply's implementation of the model
([AI SDR capability](/capabilities/ai-sdr)): it works from an
[offer](/specifications/offers) you define, uses
[playbooks](/specifications/playbooks), and grounds drafts in a
[knowledge base](/specifications/knowledge-bases). Autonomy maps onto two separate
controls:

- **Sending autonomy.** In **Approval mode**, every generated message lands in a Pending approvals queue and nothing sends until you approve it — you can edit it, regenerate it with a short prompt, or delete it, which removes the prospect from the sequence. In **Automatic mode**, Jason generates and sends without waiting for approval.
- **Sourcing autonomy.** **Autopilot mode** keeps the sequence fed by adding new contacts every day based on the filters you set. Source limits bound it: the Sales Navigator source caps at 500 contacts per day, with defaults of 10 per day and 3 contacts per company, and signal-based sources filter candidates by ICP fit — 60% by default.

Scoring is relative to the sequence, not absolute per contact: the ICP score is
computed against that sequence's offer, so the same contact can score differently in
two sequences promoting different products. A LinkedIn Activity score (Frequent,
Consistent, or Occasional) guides which channel to lead with. The full setup path is
[build an AI SDR](/workflows/build-an-ai-sdr).

## Where this breaks down

An AI SDR amplifies the strategy it is given — a weak offer sent at volume fails at
volume. Edge cases never disappear entirely, so an escalation path is permanent, not a
transition phase. And trust is asymmetric: one hallucinated claim to a key account can
cost more than the automation saved all quarter.

## Related

- [Reply classification](/learn/reply-classification)
- [Sequence design](/learn/sequence-design)
- [AI SDR in Reply](/capabilities/ai-sdr)
- [Build an AI SDR](/workflows/build-an-ai-sdr)

## Build with Reply

- REST API: [docs.reply.io](https://docs.reply.io/api-reference/introduction) — offers, playbooks, and knowledge bases
- MCP: [agents.reply.io/mcp](https://agents.reply.io/mcp) — Jason AI setup and approval operations