AI Sales Development Agent: Specification

The specification for an inbound sales development agent: what it is for, what it needs, what it produces, how the work flows, and the approval gate it can never route around.

4 min read

Complexity: High
Estimated savings: Estimate: 8-20 hrs/week of SDR administration at steady inbound volume
Time to deploy: Typically 4-6 weeks

What this agent is

A sales development agent that works an inbound lead from first contact to a scheduled meeting, drafting every outbound message for a person to approve and sending nothing on its own.

It is not an autonomous prospector. It does not buy lists, it does not decide who to contact, and it does not write to a prospect without a human-owned approval step. Those constraints are the design, not a first-release limitation: outbound email is regulated, brand voice is expensive to repair, and an agent that can send is an agent that can send the wrong thing at scale.

Why it is specified rather than demonstrated

A demo shows the good path. A specification tells you what happens on the twentieth run, when enrichment returns nothing, when two reps both claim the account, and when the model has no supported answer to a pricing question. The sections below are written to be checked against your own systems before anyone books a call.

Work an inbound lead from first contact to a scheduled meeting — drafting every outbound message for a human to approve, and sending nothing on its own.

Inputs

Form submissions, inbound email and chat transcripts
Firmographic enrichment from a licensed provider
Your ideal customer profile, held as an editable rubric owned by sales
CRM records: accounts, open opportunities, ownership and territory rules
An approved collateral set the agent is allowed to quote from
Rep calendar availability for the meetings it proposes

Outputs

A structured qualification brief for each lead, shown beside the raw message
A fit score with a written justification against the rubric
A drafted first reply and follow-up sequence, held for approval
A proposed owner derived from deterministic routing rules
A booked meeting, once a human has approved the outreach
A run log of every retrieval, score, decision and draft

Workflow

Ingest the submission, de-duplicate it against existing accounts and open opportunities
Enrich the company from licensed sources; sparse results lower confidence rather than being filled in
Summarise the company and the request into a brief
Score fit against the rubric and write the justification in the same pass
Route to an owner using deterministic rules, with a named tie-break owner when two rules match
Draft the first reply from approved collateral only, and write it to the CRM record
Hold for human approval; the rep edits or rejects, and the edit is recorded
On approval, send, propose times from real calendar availability, and book
Write the outcome back and close the loop on the run log

Benefits

First-touch latency drops from hours to the length of one human review
Every lead gets the same research, including the ones that arrive at 11pm
Scoring is explained rather than asserted, so sales can argue with it and improve the rubric
Routing is deterministic and auditable, which removes the most common source of disputes
The edit rate on drafts becomes a measurable signal of whether the agent is actually useful

A queue-backed worker per lead. Enrichment and CRM reads run first and cache against the account; the model is called once per decision with retrieval limited to approved material; every draft is written to the CRM record rather than to an outbox, and approval is a state transition a person makes on that record. The run log is a first-class store, subject to the same retention and erasure rules as the CRM itself.

Oversight and failure modes

Approval fatigue. If every draft is nearly right, reviewers start approving without reading, and the human gate becomes decorative. Measure the edit rate — the share of drafts changed before sending. A very low edit rate is a warning, not a win.

Ownership conflicts. Routing must be deterministic and must lose gracefully. When two rules both match, the lead goes to a named tie-break owner, not to whichever rule ran first.

Quoting beyond the approved material. The agent may only cite content from an approved collateral set. Pricing, contractual terms and delivery commitments should be treated as never-quote unless explicitly cleared, because a drafted number that survives review becomes a number your customer heard from you.

Enrichment gaps. Sparse firmographics must lower confidence and route to a person. The failure to avoid is a confident brief assembled from almost nothing.

Data protection. Inbound contact data is personal data. Retention, lawful basis and the right to erasure apply to the agent's run log exactly as they apply to the CRM, and the log is the part teams forget.

What it might save

A range, dependent on your inbound volume and how much of the work is currently manual: at a steady few dozen leads a week, plan for something in the region of eight to twenty hours of SDR administration per week — research, first drafts, CRM hygiene and scheduling. That is an estimate for planning, not a result we are reporting.

Would this specification survive contact with your stack?

Bring us your CRM, your enrichment source and your ICP, and we will tell you where it breaks.

Talk to us

Frequently Asked Questions

No. There is no path from the agent to an outbox. Drafts are written to the CRM record and a rep approves them, which makes the human gate a property of the architecture rather than a configuration flag someone can switch off under pressure.

It may only draft from an approved collateral set. Pricing, contractual terms and delivery commitments are treated as never-quote unless explicitly cleared, because a number that survives review becomes a number your customer heard from you.

A falling edit rate. If reviewers almost never change a draft before sending, they have stopped reading, and the gate is decorative. Measure it from the start and treat a very low rate as a warning rather than a success.

More in this category

ERP Assistant Agent: Specification

The specification for a read-only assistant over ERP data: plain-language questions answered from a curated query surface, run under the asking user's own permissions, with the filters shown beside every figure.

High complexity · 5 min read

AI Research Agent: Specification

The specification for a research agent that answers one written business question from allowlisted sources, carries a citation on every claim, and is allowed to report that something is not established.

Medium complexity · 5 min read

AI Customer Support Agent: Specification

The specification for a grounded support agent: it answers from your documentation, cites what it used, refuses what it cannot support, and hands off with the whole conversation attached.

Medium complexity · 4 min read

Oversight and failure modes

Ownership conflicts. Routing must be deterministic and must lose gracefully. When two rules both match, the lead goes to a named tie-break owner, not to whichever rule ran first.

Enrichment gaps. Sparse firmographics must lower confidence and route to a person. The failure to avoid is a confident brief assembled from almost nothing.