Most organisations have built conversion models. They work reasonably well. They forecast who is likely to buy what products, and sometimes even when. So the natural question follows: why not use them to steer your Next Best Action strategy towards the outcomes that matter?
The answer is: you can. But the mechanics matter.
There is no established best practice here. What there is: first principles, trial and error, and hard-won insight from those who have tried. Three practical approaches stand out. Each has merit. Each has traps.
Approach 1: Applicability and suitability rules
The simplest path is to use your conversion propensity scores as gates within your arbitration framework. You might suppress certain actions for customers in the lower deciles of conversion likelihood, or exclude contacts altogether when propensity falls below a threshold. Think of this as soft guardrails: suitability rules that adjust contact frequency based on your model’s signal.
This approach has clear benefits. It sits cleanly within the arbitration framework. And because these are soft rules—dropped during control groups or exploration periods—adaptive models still get to explore edge cases when needed. That exploration is valuable. It helps validate whether your conversion models are actually telling you something about customer readiness.
The risk is straightforward: heavy exclusion by such criteria can be counterproductive. You exclude actions entirely rather than making them less likely. Your customer may be ahead of your model and be in the market for your product before your scoring refresh cycle has executed.
Before rolling out such exclusions, ensure you are monitoring Impact Analyser. This is not optional. It is how you know whether your strategy is genuinely maintaining lift or simply removing optionality. Run it. Measure what actually happens to your relevance and business value. Then trial different decile thresholds. Use the tool to flag where exclusions are too restrictive. Measure the second-order effects: are you moving customers towards conversion, or removing the optionality they need?
Approach 2: Steering with context weights
A more subtle mechanism is context weighting within the arbitration formula itself. Rather than excluding actions, you amplify the relevance of those actions for customers where your conversion model signals intent.
This is a sensible approach. It bends the NBA ranking without breaking it. It works within the system rather than around it. And it lets you exploit your conversion intelligence without pretending to replace adaptive propensity altogether. Conversion models may contain signals that adaptive models do not have access to. And vice versa of course - adaptive models are there to spot those real-time buying signals that your batch conversion models have not yet seen. Conversion models will often lag adaptive models.
But context weighting has a psychological trap embedded in it. If you weight heavily towards conversion, you are, in effect, always trying to “close the deal”. This sounds sensible—you have high-intent signals; move customers towards purchase. In practice, it often fails.
Sales psychology tells us that customers need “permission to buy”. They need to feel ready. Persistent pressure to close triggers resistance. The best conversions come after you have built readiness. A customer may show all the conversion signals and still reject an application form if they are simply unaware that the product exists or unconvinced of its relevance. At that point, an education message is far more valuable. It builds readiness. It moves them closer to genuine purchase intent.
This is where the journey construct becomes essential. Rather than optimising for a single outcome—conversion—break the sales process into a series of micro-conversions. Each customer sits at a stage: unaware, aware, educated, evaluating, ready to purchase. Different actions are appropriate at each stage. An education message competes fairly with a sales message not because they are equally likely to convert, but because they are moving the customer through a necessary progression.
Each stage can be weighted by value. A customer who is aware is more valuable than one who is not. Moving a customer from unaware to educated has genuine business value. It increases their propensity to convert. Weighting your arbitration to reflect this—making actions appropriate to their current stage—means adaptive models can work properly.
If you use context weighting from conversion models, ensure the semantics align. A model that predicts near-term conversion (say, within 30 days) is a legitimate signal of intent. A model that predicts product affinity—“customers like this tend to have a credit card”—is not. The latter contains no contextual information about the customer’s readiness. Using it to weight arbitration is very likely to be counterproductive. It will oversuppress educational content that actually builds the funnel.
Approach 3: Predictors in adaptive models
Your conversion models are excellent as predictors within the adaptive layer itself. That is best practice. They add signal. Adaptive models treat predictors equally; they calculate normalised propensity across all the factors they see. If your conversion signals help predict engagement, they will show up in the model weights. If they don’t, they won’t. The adaptive engine will tell you the story.
What you gain from this approach is data insight. Run models and inspect them. Do your conversion signals influence engagement? Are other factors more important? More critically: do your conversion models correlate with successful engagement across different action types? Two customers might show identical conversion propensity but be at very different journey stages. One needs education; one is ready to buy. Your adaptive models and engagement reporting should tell you whether your conversion signals are predicting readiness or merely purchase likelihood. If they are not, you are steering blind.
Can you use your conversion models to validate your adaptive approach? Your reporting and analytics tools should be part of your normal data science workflow. They will show whether the data in your conversion models actually matters, and where.
The common thread
There is a pattern across all three. Conversion models are not a silver bullet. They are valuable signals. But pulling arbitration too hard towards immediate conversion—through heavy exclusions, aggressive context weighting, or overweighting in the adaptive model—typically decreases engagement. And engagement is, in the long term, how you build value in customer relationships.
This does not mean you should not steer towards conversion. You should. Weights are there to steer towards specific business aims: profit, products sold, retention, satisfaction. But the steering should be a gentle nudge, not an override. Adaptive propensity should still dominate.
The practical advice: start with soft rules or modest context weighting. Measure impact rigorously using Impact Analyser and simulation. A/B test. Let your data tell you how much steering is optimal. If you find that excluding or heavily weighting certain actions improves value without harming engagement, good. If exclusions become too restrictive—if you start discovering that lower-propensity customers actually engage well, or if engagement collapses—pull back.
And map your actions to journey stages. Make sure your sales funnel has steps, and that your customers are moving through them. That focus on progression—not just on purchase—tends to improve outcomes substantially. It is the difference between a transaction and a relationship.