A multi-agent outbound engine that researches, qualifies, personalizes, and books, with human sales development reps reviewing exceptions rather than running the process.
The challenge
Outbound sales at scale is a research problem disguised as a writing problem. For every prospect, someone has to find the company context, confirm the person is still in the role, decide whether they match the ideal customer profile, write a message that reflects all of that, and then handle the reply, the objection, the scheduling back-and-forth, and the CRM update. Done by hand, each step is slow and inconsistent, and the inconsistency is what kills reply rates.
The platform needed to run that entire loop autonomously for high volumes of prospects while staying accurate enough that a human would be comfortable letting it send on their behalf. Single-prompt approaches had been tried and produced generic messages, hallucinated facts about companies, and no audit trail when something went wrong.
The solution
InsightSol designed a multi-agent orchestration layer with specialist agents for research, qualification, personalization, reply handling, and scheduling. Each agent has a narrow system prompt, a short list of typed tools, and a confidence threshold. An orchestrator holds the state of every prospect as a durable state machine, so a crash or a provider outage resumes from the last committed step instead of restarting the sequence.
A critic agent scores every outbound message and every proposed action against written policy before it executes. Messages that cite facts the research agent could not verify are blocked. Replies that indicate a complaint, a legal concern, or an unsubscribe request route to a human with the full conversation attached. Everything else, including the final calendar booking and the CRM write-back, runs without a person in the loop.
Headline result
<80s
from inbound signal to a verified, booked outcome
System architecture
Orchestrator
Durable per-prospect state machine in PostgreSQL with Redis for hot state and locks
Research agent
Web and CRM enrichment through typed tools, with source citations stored per fact
Qualification agent
ICP scoring against structured criteria with an explain-your-score requirement
Personalization agent
Message generation constrained to verified facts and approved tone
Reply and scheduling agents
Intent detection on replies, objection handling, calendar negotiation, and booking
Critic and policy gate
Pre-send scoring, restricted-action detection, and human escalation routing
Integrations
Email and calendar APIs, CRM sync, and n8n for notifications and side effects
Quantified results
01
Splitting research, qualification, personalization, and scheduling into separate agents made each one cheaper to run, easier to evaluate, and far less likely to invent facts than any single-prompt approach that was tried before it.
02
Storing a source for every fact let the critic block unverifiable claims before sending. That protected reply rates more than any copywriting change, because prospects notice a wrong fact faster than a dull sentence.
03
Resuming from the last committed step with idempotent actions turned provider outages from incidents into non-events. No sequence restarted, no message went out twice, and nobody had to be paged.
Disciplines involved
The audit produces the same diagrams and unit economics this system started with, for your workflow and your data.
