Retention AI Maturity Model

🧠 What this is: A methodology to build and execute retention agents leveraging the full capabilities of the technology (not automating old workflows).

👤 Who it's for: Senior leaders responsible for the retention number at recurring revenue businesses: CCOs, CMOs, and commercial heads who want to move beyond rules-based logic towards outcome-based autonomous agents operating within strict trust-defining guardrails.

📈 What it does: It provides a structured methodology for moving from generative AI co-pilots to operationalising retention agents.

Introduction

In framework 4, I described what needs to be true and in what order, for retention efforts to compound rather than cancel each other out. At the top of the hierarchy of needs are moments of truth. These are moments in the relationship that make or break retention (covered in detail in framework 3).

Retention AI efforts operate at the moments of truth layer, supporting brands in managing their base across the lifecycle to maximise retention outcomes, without eroding margins.

In this framework, I cover 5 levels of retention AI maturity:

(1) Rules-based logic

(2) Predictive targeting

(3) Behavioural understanding 

(4) Assisted agents

(5) Autonomous agents

The reality is that different organisations are on a different trajectory with AI, whether it is already deeply embedded or an emerging practice. The aim of this framework is to assist recurring revenue brands in navigating the relevant levels with confidence, without breaking customer trust or causing organisational disruption.

The Retention AI Maturity Framework

The Retention AI Maturity framework is essentially a staircase, where each LEGO brick builds on the last. 

At level 1 is rules-based logic (AI still plays a role here). At level 2, maturity advances towards predictive retention and targeting, and while this has been leveraged for many years, recent AI advancements can massively accelerate execution. In levels 1 and 2, we are using AI to enhance or automate existing processes. In levels 3-5, we start leveraging AI to drive new interventions based on desired outcomes rather than predefined journeys or processes. 

Level 1: Reporting & Rules (Generative AI for Fixed Journeys)

The Core Focus: "Tell me what happened" Retention Outcome: Low

Level 1 represents the foundational, yet highly reactive, state of retention management. At this level, retention strategies are governed by rules-based logic and retrospective reporting. Brands rely heavily on lagging indicators, channel-centric execution, and static segments (e.g., "Send an email if the user hasn't logged in for 7 days").

At Level 1, AI is used primarily to write copy and generate imagery for fixed, manually built CRM journeys, digital experiences or CX interactions. Teams use Generative AI tools as creative co-pilots, drafting subject lines or iterating on body copy. 

While this speeds up content production, the underlying logic remains heavily reliant on human effort and rigid decision trees.  However, generative AI tools and co-pilots do help teams execute level 1 programs more efficiently and effectively. 

Level 2: Predictive Targeting (Machine Learning & Automated Ingestion)

The Core Focus: “Tell me who might churn” Retention Outcome: Incremental

Level 2 moves from reporting on the past to predicting the future. Organisations at this tier leverage classic Machine Learning and Deep Learning models to generate churn scores, risk cohorts, and propensity models. Predictive churn models have been used for decades with mixed success. Best case, you see a small incremental lift; worst case margin erodes. 

While targeting improves, execution remains largely batch and journey-based. 

However, operational execution at Level 2 has recently undergone a massive technological leap. Rather than spending months manually building templates, loading copy, and structuring journeys, we can leverage advanced AI (such as Claude and Claude Code) to fully automate the production and ingestion of journeys into your CRM or ESP. 

This is operationalised in three steps:

1. We define a client-specific retention strategy and high-level communications framework.

2. We feed our proprietary retention knowledge base (in my case built on 26 years of retention & CRM IP) into Claude and Nano Banana to generate 90-95% send-ready copy and imagery across unlimited communications and lifecycles.

3. We then use Claude Code to write the scripts that load these templates, journeys, and content components directly into your CRM platform (like Airship or Optimove) via APIs.

This automated production and ingestion removes the months-long execution bottleneck, often enabling recurring revenue brands to move from strategy to send-ready lifecycle execution in under 30 days.  

Level 3: Behavioural Understanding (The Embeddings Concept)

The Core Focus: “Tell me what’s changing” Retention Outcome: Meaningful

Level 3 is where the true paradigm shift occurs. We move away from rules and predictive scoring and introduce the concept of Embeddings.

An embedding is a deep learning methodology that turns messy, human behaviours: what a customer uses, how often, in what order, and with what intensity into a mathematical "behavioural fingerprint". 

Instead of relying on rigid rules, embeddings place these fingerprints onto a multi-dimensional map; customers experiencing similar emotional states and behavioural shifts naturally cluster near one another in this "meaning space".

This enables the system to detect behavioural drift a gradual movement away from value that happens weeks before standard churn signals appear. 

At Level 3, we map the four universal drivers of churn directly into embedding inputs:

• Delayed Value Realisation: We embed time-to-first-meaningful-action and feature discovery depth. Embeddings ask, "Does this customer behave like others who struggled to 'get it'?"

• Usage Guilt: We embed declining frequency slopes, interrupted streaks, and session efficiency. Embeddings ask, "Is this customer drifting out of a routine rather than explicitly rejecting the product?"

• Novelty Decay: We embed repetitive usage paths and shrinking exploration radii to identify customers who are bored but not yet absent.

• Lack of Recognition: We embed relational distance metrics, such as declining response rates to comms or silence after service recovery, to detect when a relationship has gone cold emotionally.

By leveraging embeddings to achieve true behavioural understanding, we can finally intervene based on what is changing in a customer's motivation state, making interventions lighter, earlier, and vastly more effective.

Level 4: Assisted Agents (Human-in-the-Loop)

The Core Focus: “Recommend the right action” Retention Outcome: Compounding

At Level 4, we introduce Assisted AI Agents. While embeddings at Level 3 allow us to detect drift, Level 4 agents leverage advanced planning capabilities (such as reasoning and tool orchestration) to recommend the exact right intervention.

However, these agents do not act with total autonomy; they operate under strict Human-in-the-Loop (HITL) oversight.

When the embedding detects that a customer's behavioural fingerprint is drifting towards "Usage Guilt," the Assisted Agent evaluates the state.

Using Retrieval-Augmented Generation (RAG), the agent retrieves the approved intervention playbook from your secure internal systems. It then recommends a tailored response, such as a habit-scaffolding message or a low-effort usage alternative.

Strict guardrails and governance define this level:

• Human decision-makers define explicitly which embedding-driven actions require approval.

• High-value customers, regulatory edge cases, and complex pricing boundaries always escalate to a human agent.

• The agent provides decision support, but execution proceeds only after a human verifies that the action is safe and proportionate.

By combining machine-scale pattern recognition with human judgment, Level 4 creates compounding retention outcomes without risking brand trust.

Level 5: Autonomous Retention Agents

The Core Focus: “Act proportionately, learn continuously” Retention Outcome: Structural Advantage

Level 5 represents the pinnacle of the AI retention maturity model. At this stage, we deploy moment-specific Autonomous Retention Agents that seamlessly monitor, reason, and execute end-to-end within pre-approved boundaries.

Instead of relying on a single monolithic system, Level 5 utilises Multi-Agent Collaboration, where specialised agents are deployed for specific moments of truth. 

Examples include:

• The Activation Rescue Agent: Detects when a user is experiencing early-life friction and dynamically reorders onboarding steps or triggers just-in-time SMS guidance.

• The Value Realisation Pathing Agent: Spots novelty decay (e.g., endless scrolling without consumption) and silently re-ranks homepage content or injects fresh relevance.

• The Habit Monitor Drift Agent: Identifies a declining slope in usage and surfaces low-effort alternatives to gently scaffold the habit back into place.

Crucially, Level 5 agents are self-improving. They utilise a Reinforcement Learning Flow: if an agent tries a contextual reminder and the customer's usage slope recovers, the agent receives a positive reward signal.

Over time, the agent's long-term memory recognises which sequences of nudges, content, and micro-rewards work best for each drift type or moment of truth, continuously optimising your retention strategy with limited manual oversight.

Conclusion: Navigating the Curve

The journey from Level 1 to Level 5 is a curve. 

The most common and catastrophic mistake recurring revenue brands make is attempting to skip Level 2 (Prediction) and go directly to Level 5 (Autonomous Agents). 

Without the foundational semantic mapping of Embeddings (Level 3) and the strict human-defined guardrails of Assisted Agents (Level 4), automation breaks trust, hallucinates interventions, and alienates customers.

By deliberately anchoring your transformation in drift prevention, mastering the embeddings concept, and strictly defining where autonomy helps versus harms, you can build a fleet of retention agents that scale your execution, protect your margins, and create a structural advantage in your market.

I am working with clients at all levels. If retention focussed AI is on your agenda, we should talk. Email: [email protected] or DM me on LinkedIn.

Until next time, 

Tom