01 — Discovery
Attributes lived in tribal memory.
The right field for a high-value segment was spread across three docs, two analysts' heads, and a dashboard nobody owned. Average attribute discovery: 3.5 days per request.
3.5 days · per audience
From 4.5 days
to 4.5 seconds.
How RevForgeHQ shipped a production-grade Audience Agent for a Fortune 500 FinTech client — collapsing a multi-day, multi-team workflow into a single English-language conversation, with traceability the compliance team actually trusted.
Results · First 90 days post-launch
−98%
Cycle time
Audience build collapsed from 4.5 days to under 5 seconds.
$4.2M
Attributable lift
Incremental quarterly revenue from previously un-shipped audiences.
87%
Marketer adoption
Of the campaign team self-serving in the first 90 days.
0
Rogue pushes
Every CDP write gated by human approval. Compliance signed off in week one.
01 — Challenge
A team of 12 analysts
was the bottleneck.
The client's CDP held a decade of beautifully instrumented behavioral data — but a single audience took days to build, every campaign waited in a queue, and the marketers who knew the brand best couldn't speak the schema. We were brought in to fix the workflow, not the data.
02 — Build
Hand-written, fragile SQL.
Once attributes were found, an analyst hand-wrote SQL, validated counts, and pushed to the CDP. One typo missed 30% of the segment. One forgotten consent join failed compliance review.
1 day · per audience
03 — Trust
No one trusted black-box AI.
Two prior text-to-SQL pilots had failed compliance — no lineage, no reproducibility, no approval gate. Marketing Ops refused to push any agent-built segment live without explanation.
2 failed pilots · prior 18 months
02 — Approach
A graph, an agent,
and a careful door
to production.
We treated audience-building as a language problem wearing a data problem's costume. The fix was vocabulary on demand — and a hard wall between agent and CDP.
Phase 01
Weeks 1–3 · Pilot
Map the graph.
Working sessions with marketing ops, analytics, and compliance to extract the tacit ontology. Curated 412 nodes, 1,108 edges, every consent edge double-signed.
- Attribute audit · 6 source systems
- Ontology workshops · 11 sessions
- Synonym index · 2,300 phrases
Phase 02
Weeks 2–3 · Pilot
Build the agent.
LLM planner with structured tool-use against the graph. Every reasoning step logged. Self-critique loop catches edge cases before they ship.
- Plan → Resolve → Query → Critique
- Cypher-against-AKG, never raw SQL
- Eval suite · 240 golden audiences
Phase 03
Weeks 4–6 · Rollout
Earn trust, then ship.
HITL approval gate on every CDP write. Full reasoning trace and data lineage surfaced inline. Compliance, Ops, and Legal all approved the workflow before GA.
- Approval gate · zero auto-writes
- Lineage view · 100% of attributes
- Phased rollout · 4 marketer cohorts
03 — Solution
The Audience Agent,
in five layers.
A planner over a curated knowledge graph, with explainability surfaced as a first-class UI — not an afterthought. Every artifact is reproducible and auditable.
The marketer phrases targeting in plain English. The LLM decomposes intent into sub-goals. The agent resolves each phrase against the Audience Knowledge Graph — a curated ontology of customer entities, attributes, events, and consent edges — then compiles a multi-hop Cypher query.
Results are returned with counts, sample profiles, and a full data-lineage view tracing every attribute back to its source system with a confidence score. The agent self-critiques edge cases and surfaces them for human review.
Nothing reaches the CDP until a marketer clicks Approve. The write is signed, logged, and reversible. The compliance team's only ask in week one: keep the gate, and we'll sign.
04 — Results
90 days after launch.
The metric that surprised the room wasn't speed — it was adoption. When marketers could trust the lineage, they shipped audiences they'd never have asked an analyst for.
We've stopped queueing audience requests. The agent is the queue.
−98% cycle time
From 4.5 days to ~4.5 seconds
Median end-to-end build time across 1,840 audiences shipped in Q1.
$4.2M
Attributable revenue lift · Q1
Driven primarily by long-tail audiences that previously cleared no business case to build.
4× volume
Audiences shipped vs. pre-launch
2,400 marketers self-serving — 12 analysts redeployed to high-leverage modeling work.
0 incidents
Zero compliance flags · zero rogue writes
HITL gate held. Every CDP write traceable to a named approver and reasoning chain.
05 — What we built
Deliverables.
A working agent in production, the curated knowledge graph that powers it, and the operating model the client's team now runs on their own.
The Audience Knowledge Graph
412 curated entities, 1,108 edges, full lineage to source systems. Built collaboratively; owned by the client's data team.
The Agent (production deploy)
LLM planner + tool-use + HITL workflow. Deployed inside the client's existing MarTech stack. Average response: 4.3s.
Eval & observability suite
240 golden audiences as regression tests. Per-step latency, token, and cost telemetry. Reasoning traces archived for audit.
The operating model
Approval routing, escalation rules, weekly KG curation ritual. Documented, transferred, and now run by the client's MarTech team.
Work with RevForgeHQ
Have a workflow
worth 4.5 days of waiting?
We build production-grade agentic systems for marketing, sales, and revenue teams. If your queue looks like the one in this case study, let's talk.