For 20 years, loyalty programmes have been operated by teams of marketers who manually build segments, design campaigns, write copy, schedule sends, and chase reports. The cognitive load is enormous; the throughput is low; the work is repetitive in ways that should make any analyst suspicious that software could do it.
It can. AI agents — autonomous software that observes data, makes decisions, takes actions and learns from outcomes — handle 80% of loyalty operations today, and the share is climbing. The right operating model for 2026 is human-led strategy, agent-executed tactics.
The 10 agents that matter
Fundle Brain runs 10 production agents, each owning a specific operational domain:
| Agent | Owns |
|---|---|
| Loyalty Agent | Points credit/debit, tier qualification, reward catalogue management, liability tracking |
| Segmentation Agent | RFM scoring, 11+ behavioural cohorts, cohort migration tracking |
| Engagement Agent | Gamification mechanics, streaks, challenges, micro-moment triggers |
| Campaign Agent | Audience selection, channel arbitration, send-time optimisation, A/B variants |
| Retention Agent | Churn prediction, win-back triggers, tier-protection logic, lifecycle nudges |
| Personalisation Agent | Per-member offer/channel/timing choice, content variant selection |
| Analytics Agent | Anomaly detection, KPI alerts, natural-language data interface |
| Monetisation Agent | Retail-media placements, sponsored rewards, brand partnership orchestration |
| ROI Agent | Control-group setup, incrementality measurement, programme-level ROI reporting |
| Store Support Agent | Real-time customer lookup at POS, suggested offers at till, training recommendations |
How an agent actually works
An agent is more than a model. It's a loop: observe, decide, act, measure, learn.
- Observe: read the relevant data (member behaviour, programme state, environment)
- Decide: pick the next-best-action from a policy (e.g., "for at-risk members, trigger win-back journey with 25% offer")
- Act: execute through APIs (send WhatsApp template, credit points, publish audience, etc.)
- Measure: track outcome (did the member return? did churn risk drop?)
- Learn: update the policy based on outcomes; over time, the agent gets sharper
The agent doesn't replace the team — it replaces the routine. The team sets policy (what cohorts matter, what guard-rails apply, what budget ceilings exist), and the agents execute within the policy at scale.
The human-agent division of labour
- Humans own strategy: programme design, earn-burn structure, brand voice, regulatory compliance, executive reporting
- Agents own execution: per-member decisions, channel choice, send-time, content variant
- Humans review escalations: edge cases the agents flag for review, novel patterns the policy doesn't cover
- Humans monitor agents: review agent decisions monthly, audit policy drift, recalibrate when business shifts
The shift is hard for teams used to manual control. Done well, it frees a team of 8 to operate a programme that previously took 30 — the cost saving is real, but the strategic capacity gain is bigger.
Trust and explainability
Agents that operate without explanation are agents the team will mistrust. Every Fundle agent decision carries a "why" — the rule, model output, and policy variables that produced the action. When a member is moved to At-Risk and triggers a win-back, the agent can show the team exactly which signals fired and the predicted ROI of the intervention.
Don't deploy 10 agents at once. Start with Segmentation + Campaign + Retention (the high-volume routine work). Add Personalisation in Month 3 (depends on baseline). Add Analytics + ROI in Month 4 (most teams want explainability before incrementality). Monetisation and Store Support last (require ecosystem buy-in).
See 10 AI agents in production
30-minute walkthrough — live demo of each agent on real data, with the policy controls a human team uses to govern them.
FAQs
Are these LLM-based or model-based?
Both. The decision policies use traditional ML models (gradient boosting, neural nets) for prediction tasks. LLMs power the natural-language interface, content generation, and reasoning for novel cases. Each agent uses what works.
What if an agent makes a bad decision?
Agents operate within policy guard-rails set by the team (budget caps, frequency caps, brand-safety rules). Bad decisions are flagged in monitoring; the policy is recalibrated. The error rate at scale is lower than human teams; the error cost is bounded by guard-rails.