Mastering the RUSHA AI Insights Engine

RUSHA is LedgeSpace's proprietary AI layer, designed to augment analytical capabilities and reduce manual data interpretation time.

How to Use Conversational Querying

1. Open the RUSHA Chat Interface.
2. Ask natural language questions about connected performance data (e.g., "Why did our CPA increase this week?").
3. RUSHA has direct tool-calling access to live data across all connected integrations to formulate an answer.

How Intent Classification Works

Before querying data, an Intent Classifier categorizes your prompt:

• query_simple: Single metric lookup.
• query_complex: Comparisons, root-cause analysis, cross-platform queries.
• Unrelated Block: If you ask a non-marketing/non-data question, the system immediately cuts off the LLM and triggers a graceful fallback, preventing prompt injection.

How the Dual-Voice System Protects You

RUSHA runs a dual-pass generation pipeline to protect agency-client relationships:

1. The Internal Version: Generates a blunt, exact, root-cause diagnosis for internal agency staff.
2. The Client-Voice Rewrite: Before the user ever sees the output on a public dashboard, a secondary LLM pass rewrites the internal version to be diplomatic and forward-looking, framing issues as "opportunities".
3. Factual Consistency Lock: A programmatic guard verifies that every single numeric token in the internal version is present in the client version—tone is altered, but data cannot be lost or hallucinated.

How Zero-Hallucination Guardrails Work

Every RUSHA response passes through a multi-step validation loop:

• Strict Number Verification: Any number output by the LLM is mathematically verified against the raw integration data. If it cannot be proven, the response is rejected and retried.
• Deterministic Citations: Citations are generated deterministically based strictly on which API tools were successfully called.
• Forbidden Language Checker: The AI is strictly barred from using absolute words like "confirms," "proves," or "definitively."

Automated AI Features

• Proactive Anomaly Detection: A background scan compares your client's last 7 days against their historical baseline (4–8 weeks) across spend, conversions, revenue, contacts, deals, and email metrics. When something moves more than 2 standard deviations from baseline, RUSHA writes a diagnosis and files it as a draft insight.
• Correlation Detection: A separate background scan looks for strong statistical relationships (Pearson correlation, |r| > 0.7) between metrics across your connected platforms over an 8-week window — e.g. "email click rate rises in lockstep with email open rate." Results are cached for 24 hours per client so the same correlation isn't re-surfaced repeatedly.
• Executive Summaries: RUSHA writes concise (under 1,000 characters) executive summaries specifically for PDF or web reports, factoring in the report's exact date range.

Reviewing and Approving Draft Insights

Both anomaly and correlation detection create drafts, not published insights — RUSHA never shows a client-facing insight without a human approving it first.

1. Go to Dashboard → RUSHA AI → Insights Feed.
2. Click any notification to expand it — you'll see both the Client View (diplomatic, forward-looking) and Internal View (blunt diagnosis) tabs, exactly like a chat response.
3. Choose Approve (1 insight credit) or Dismiss.
   • Approving costs 1 insight credit from your quota (see Pricing, Quotas, & Billing) — the background scan itself is always free; you only pay when you decide the insight is worth surfacing.
   • Dismissing costs nothing.
   • An unreviewed draft sits at pending_review indefinitely — it never auto-approves or auto-expires.

Public "Vault" Display

When a client views their live, tokenized web report, RUSHA's latest approved correlation insight is injected at the top of their dashboard as continuous "proof of value." This only happens for insights approved before that specific report was generated — reports capture an immutable snapshot at generation time, so approving something after a report has already been sent won't retroactively update it. Generate a new report to pick up a newly-approved insight.