A professional services firm was receiving a consistent volume of inbound inquiries through email. The challenge was not generating leads. It was responding to them well, quickly, and at scale. Sales reps were spending significant time reading each inquiry, locating the relevant case material or service documentation, and drafting a contextualised response. Responses that required matching a specific inquiry type to a specific service area were taking anywhere from 45 minutes to several hours per thread.
The risk with automating this was clear from the start: a response that misreads the inquiry, misrepresents a service, or strikes the wrong tone does not just fail to convert. It actively damages the relationship. Any workflow here had to keep a human in control of what actually gets sent.
Zone Design
AI Zone: lead ingestion, prequalification, and context matching
The workflow begins when a new inbound email arrives in the firm's sales inbox. A read-only Gmail adapter, scoped to inbox reading only, fetches the message and passes it to a prequalification layer.
Prequalification determines whether an inquiry is a genuine lead: it filters out automated responses, out-of-office replies, and clearly off-scope messages. This classification runs in the AI Zone. It is bounded, the stakes of a misclassification are low (a missed lead gets surfaced for manual review, not discarded), and the logic is straightforward enough to be validated by volume.
Qualified leads then go through a context matching step. The inquiry is analysed to determine the service area it most plausibly relates to. The workflow queries two sources: an off-the-shelf CRM for existing client records and past interaction history, and internal Notion documents that describe specific service offerings, pricing tiers, and relevant past work. The combination allows the AI to construct a draft brief: what is this person asking about, what have we offered similar clients, what context is most relevant to include in a response.
Interface: AI Zone draft to Human Zone review
The matched context and the original inquiry are passed to a drafting step that produces a Gmail draft using a custom adapter that strips away the permission to send emails. The draft is visible to the sales rep in their Gmail account and is never sent automatically.
Each draft is reviewed by a sales rep before sending. The rep sees the original inquiry alongside the draft, with the matched context that informed it. They can edit freely, send as-is, or discard and write from scratch.
The AI handles the research, structuring, and initial drafting. The human makes the decision to send and owns the communication.
Feedback loop: continuous tone improvement
We implemented a feedback process between drafts created and drafts sent. When a rep edits a draft significantly before sending, those edits are captured as a signal. Over time, patterns emerge: certain phrasings that are consistently softened, certain service descriptions that are regularly supplemented, formality adjustments that go in the same direction across multiple reps.
This feedback does not retrain a model. It populates a structured prompt library that evolves the drafting instructions over time. The sales team, not the AI system, controls which feedback signals are incorporated. Reviews happen on a defined cadence; the team approves changes before they affect future drafts.
Human Zone: all send decisions
No message is sent without a sales rep reviewing and actively choosing to send. This is a hard architectural constraint, not a guideline. The firm's position was clear: AI involvement in client-facing communication is only acceptable if a human is accountable for every message that leaves the account. The compose-only adapter design enforces this.
What Changed
Before this workflow, the time from inquiry receipt to first response was up to four days. Reps reported spending up to two hours on research and drafting per lead for mid-complexity inquiries, time that came at the cost of other pipeline work.
After deployment, reps report spending an average of 8 to 12 minutes reviewing and editing drafts. The research and structuring work, which previously consumed most of the drafting time, is completed by the workflow before the rep opens the message. Response times to inbound leads improved by approximately three times on the median.
Draft acceptance rate (the proportion of AI drafts sent with only minor edits rather than discarded) reached 83% and continues to improve as the feedback loop refines the tone library and services to focus on. Reps report that even when they edit a draft, having the structure and context assembled is valuable.
Zone Split
- AI Zone: Read-only inbox monitoring, lead prequalification and filtering, CRM and Notion context matching, draft assembly
- Human Zone: Rep reviews Gmail draft, edits, and decides whether to send; all send decisions, all final message content, feedback approval for prompt library updates
Design Notes
Creating a custom adapter to limite permissions to reading and drafting emails without any sending capabilities was crucial. Any broader permission, including read-send without drafts, would have crossed a line the firm was not willing to cross, and correctly so. Constraining the adapter to draft creation meant the Human Zone handoff was enforced by the system architecture rather than by discipline alone.
The CRM and Notion integration was scoped narrowly. The workflow queries for context relevant to the inquiry type. It does not have access to full client records, financial terms, or relationship notes beyond what is needed to draft a response. This was both a privacy consideration and a scope management decision: wider access would have increased the surface area for drafting errors.
The feedback loop is deliberately slow and human-gated. We could have designed a tighter loop where edits are reflected in drafts within a day. We chose not to. The goal is tone consistency over time, not rapid adaptation. The sales team's ability to review and approve changes before they affect live drafts is more important than the speed of improvement.