How AI4Love's Nightly Agent System Works

AI4Love runs seven specialized intelligence agents every night that analyze your supporter data and surface actionable insights by morning. The agents use deterministic mathematical rules to detect patterns — the AI generates readable text only after a pattern is confirmed. Your team starts each day knowing who needs attention.

The Three-Layer Decision Model

Understanding how the agents work requires understanding the three layers:

Layer 1: Deterministic Analysis (Rule-Based)

Each agent applies fixed mathematical rules to your supporter data:

Recency, Frequency, Monetary (RFM) scoring
Activity trend calculations (rolling averages, decay functions)
Threshold-based triggers (e.g., "no activity in 60+ days" flags as at-risk)
Cohort comparisons (individual performance vs. segment averages)
Eligibility filters (suppress insights for supporters below minimum activity thresholds)

This layer is not AI. It is rollups, formulas, and conditional logic. The same inputs always produce the same outputs.

Layer 2: Text Generation (LLM-Constrained)

Once an agent identifies a pattern worth surfacing, it calls the Claude API to generate the insight's human-readable text. The LLM is constrained:

Input: A scoped prompt containing the supporter's name, the detected pattern, and relevant metrics — not the full record
Output: A short headline and recommendation, validated against the pattern type before being saved
Scope: The LLM writes text. It does not decide what to surface — that decision was made in Layer 1

The LLM cannot access data beyond what the prompt contains, trigger actions, or override suppression rules.

Layer 3: Human Action

Insights appear in the dashboard. Staff read them and decide which to act on, how to respond, and whether to dismiss or defer. AI4Love does not close the loop — there is no auto-send, auto-enroll, or auto-assign.

The Seven Agents

At-Risk Detection

Flags supporters showing declining engagement, ranked by urgency and relationship value. Catches drift before the relationship is lost.

Conversion Opportunities

Finds supporters ready for deeper commitment — volunteers ready to donate, donors ready for monthly giving, supporters giving below capacity based on their engagement signals.

Recognition Triggers

Detects milestones worth celebrating: giving streaks, volunteer hour thresholds, participation anniversaries, and moments that should be acknowledged before they pass.

Campaign Intelligence

Surfaces individual-level campaign signals — which campaigns are converting, which channels perform best, and who is being over-solicited across multiple campaigns simultaneously.

Relationship Deepening

Finds pure connection-building opportunities with no ask attached. These are the insights that donor-centric systems miss entirely.

Cross-Agent Intelligence

Synthesizes patterns that only emerge when you look across all dimensions at once. A supporter may not be at-risk by any single metric but show a combined pattern that warrants attention.

Knowledge Enrichment

Enriches every insight with best practices from nonprofit sector research sources. Grounds recommendations in what works across the sector, not just your organization's data.

What the Agents See

Each agent's prompt template declares exactly which fields it needs. Only those fields are sent to the LLM. Adding a new field to Airtable does not automatically expose it.

Example of what is sent:

Supporter name
Pattern detected (e.g., major donor engagement drop)
Days since last activity
Lifetime giving amount
Previous activity trend

What is never sent:

Email address
Phone number
Street address
Date of birth
Payment method
Government ID

Failure Behavior

The agents are designed to suppress rather than guess:

If an agent cannot determine a pattern with confidence, no insight is generated
If the LLM returns text that does not match the expected pattern type, the insight is discarded
If a supporter has insufficient data, eligibility filters suppress the insight
If the LLM is unavailable, pattern detection completes (it is math), but text generation is queued for the next run
Duplicate insights are deduplicated within a configurable window

Silence is safer than noise. A missing insight costs nothing. A wrong insight erodes trust.

Scheduling and Orchestration

The agents are orchestrated by Make.com, running on a nightly schedule. Each agent:

Queries your Airtable base for eligible supporters
Applies deterministic pattern detection rules
Filters out supporters who do not meet minimum activity thresholds
Sends scoped prompts to the Claude API for insight text generation
Validates the generated text against the expected pattern
Writes confirmed insights back to your Airtable base

By morning, your team sees fresh, prioritized insights ready for action.

Common Questions

Are the insights the same every morning? No. Each nightly run analyzes current data. As supporter activity changes, different patterns emerge. Deduplication prevents the same insight from being surfaced repeatedly.

Can we adjust how sensitive the agents are? Yes. Thresholds are calibrated during implementation and tuned at the midpoint review. If your average donor gives twice a year, a 6-month gap is treated differently than for an organization where donors give monthly.

What happens if the nightly run fails? If an agent cannot reach Airtable, no insights are written and the error is logged. The next nightly run retries automatically. No partial or incorrect insights are ever saved.

Can we turn off specific agents? Yes. Each agent runs independently. If your organization does not need campaign intelligence or recognition triggers, those agents can be disabled without affecting the others.

How many insights does the system typically generate per night? This depends on your supporter base size and activity levels. The system is designed to surface the most important patterns, not generate volume. Most teams see 3-5 high-priority actions per day.