Beyond Customer Propensity: Incorporating Business Value in AI‑Driven Next Best Action

In almost every Customer Decision Hub engagement I’ve been involved in, the same conversation surfaces sooner or later. Sometimes it appears early, sometimes only once a programme is already live, but it is there. The question is not whether to use PxV arbitration, but what we actually mean by value, and how should we implement it?

As practitioners, we are fortunate to work with the most sophisticated real-time AI decisioning platform available. Pega’s adaptive machine learning does an exceptional job of understanding customer behaviour and producing propensities to guide empathic interactions. For many organisations, that capability alone is transformative. It replaces static targeting rules with evidenced relevance, and guesswork with AI-driven learning.

But eventually, organisations realise that delivering relevance is not the same as business success. The AI may be good at predicting what customers are likely to accept, while being silent on whether those outcomes are the ones the business should be caring about. That is usually when PxV enters the conversation, and when opinions start to diverge!

In theory, PxV arbitration sounds straightforward. In practice, it quickly exposes a deeper organisational tension: many businesses are not aligned on what customer value actually means. Marketing, engagement, service, and sales each contribute to value in different ways, over different time horizons, and often with different success measures. When these perspectives collide inside a single arbitration framework, uncertainty about value is inevitable, not because the technology is unclear, but because the organisation itself is still negotiating what it truly values.

There are familiar fears. How do we assign value to non‑revenue generating actions like service, reassurance, or education? How do we value a retention initiative? Won’t high‑value products (mortgages and broadband are the usual suspects) dominate arbitration even when customer readiness is low? Should value reflect finance, strategy, customer satisfaction, or brand intent? And who gets to decide?

What I’ve seen repeatedly is that these debates rarely fail because teams lack sophistication. They fail because V may be quietly asked to do too many jobs at once: represent contribution, encode urgency, express brand tone, fix prioritisation, and support reporting — all in a single number.

Different organisations resolve this tension in different ways. Some deliberately stay p‑only longer than feels comfortable. Others lean heavily on weights, or push journeys to do most of the sequencing work. The least successful response is to abandon AI‑driven arbitration altogether and fall back to targeting rules. In all of these cases, the organisation fails to realise the full promise of Customer Decision Hub: an AI‑driven, continuously optimising customer experience engine and reduce it to little more than a marketing campaign tool.

The purpose of this post is not to prescribe a single “correct” way to design V in CDH. Instead, it lays out a set of principles that I’ve found useful when advising clients — principles that help teams disagree productively and evolve safely.

These principles are offered deliberately as conversation starters, not commandments.
If you’ve solved this problem differently — or think some of these principles are wrong — I’d actively encourage you to say so in the comments. PxV is one of those areas where healthy disagreement can be a sign of innovation, not confusion.

The Ten Core Principles of PxV Value Design

The principles below are not configuration tips. They are a “constitution” for PxV arbitration. They are constraints that protect learning, reporting, and optimisation as sophistication and strategies evolve.


1. Separate customer relevance from business importance

Propensity measures what the customer is likely to do. Value expresses how much the business cares if it happens.

This separation is foundational. When customer relevance and business importance are conflated, decisioning flips between being customer‑led and sales‑driven, with no stable centre. PxV exists precisely because neither perspective is sufficient on its own.


2. Keep value to one question

“If this action is accepted, how good is that outcome for the business?” Nothing else belongs inside value.

The moment value starts encoding urgency, timing, preference, or journey stage, it loses meaning. Those signals belong elsewhere. Value must stay conceptually clean to remain interpretable.


3. Treat value as a journey contribution, not eventual product payoff

In multi‑step outcomes, action value represents progress toward success, not the end reward in isolation.

Most meaningful business outcomes require multiple experience decisions over time. If all value is assigned to the final step, arbitration becomes structurally biased toward premature closing and undermines experience.


4. Conserve value across the journey

Distributed action values should add up coherently so reporting, simulation, and optimisation remain meaningful.

Tools like Value Finder, Scenario Planner, and Impact Analyzer assume value is additive and comparable. If value is double‑counted or inflated, these tools faithfully optimise the wrong thing.


5. Value must include non‑revenue outcomes

Service, reassurance, risk mitigation, and journey‑enabling actions carry value because they enable long‑term success.

Service is not the opposite of sales. In many journeys, it is a prerequisite. Excluding non‑revenue actions from value design systematically biases decisioning toward short‑term gain.


6. Make value universal and comparable

Value should not vary by customer, moment, or context; it must remain stable enough to compare across populations and time.

Customer‑level variation belongs in outcome measurement, not in action value. Stability is what allows learning and optimisation to mean anything.


7. Use weights to express strategy and brand intent

Weights define emphasis, balance, and timing — they are not a correction for incoherent value.

Weights are where judgement lives. They allow organisations to decide how sales and service should coexist, without corrupting the value signal used for learning.


8. Use journeys to control the time when actions should dominate

Journeys manage timing and sequencing, so the right contributions win at the right stage.

Journey stage weighting changes priority without redefining contribution. This is what allows decisioning to respect customer readiness while remaining analytically sound.


9. Design value for learning: it must be explainable

Value does not need to be precise, but it must be defensible as a measurement currency for optimisation and impact analysis.

Learning fails not when models are wrong, but when meaning becomes blurred and incoherent. Explainable value is more important than “accurate” value.


10. Start simple — then evolve under governance

It is valid to begin with relevance‑led decisioning (p‑only) and refine value deliberately as a governed asset over time.

P‑only decisioning can be a strategic phase. What matters is recognising when it breaks — and evolving value thoughtfully, not reactively.


Closing thought

PxV arbitration is not about aggresively forcing business outcomes onto customers. It is about making customer relevence, value and strategy explicit — clearly, transparently, and governably — inside an adaptive machine‑learning decisioning system.

A continuously optimising customer experience depends on a clean separation of concerns: propensity to represent customer relevance, value to express business importance, weights to encode brand intent, and journeys to manage time and sequencing. When those responsibilities are blurred, decisioning still functions — but learning, trust, and optimisation inevitably suffer.

I hear and feel the “fears” you mention all the time @LEWIA . Great to see some guidance on this. Would be good to learn how to assign Value to ‘non-revenue outcomes’ as this gets raised often.

Yes Chris, that’s a fair question. What I’ve shared here is really a set of principles rather than a how‑to. They don’t remove the need to actually design a value scheme across actions and outcomes, they are more about putting some guardrails and principles around how that thinking evolves.

Non‑sales actions definitely still have value. Sometimes that’s cost avoidance such as stopping things going wrong in the first place (complaints, churn, repeat contacts). Sometimes it’s about enablement,educating customers so products are used properly, journeys don’t stall, or the chance of conversion later on is higher.

In practice, consistency matters more than accuracy. Value doesn’t need to be “right” in a finance sense, it needs to work as a stable, explainable value unit that people can reason about. I tend to think of it as a kind of internal value accountancy rather than pricing outcomes. In some verticals/appliations it might even be right to think of “value” as nothing to do with money, but “engagement”, “satisfaction” etc. As long as you are consistent.

A really practical first step is getting the action taxonomy right. You want to be able to compare apples with apples e.g. “what is my next best mortgage education message?” rather than education versus closing. Once you’ve got that structure, you can apply a coherent set of values across the taxonomy and journeys and then use weights to let those groups compete sensibly with late‑stage sales actions.

That’s usually where the conversation becomes much more productive and where different organisations quite reasonably land in somewhat different places.

Great article and guidance! Let me abuse it to add one of my pet topics as it is related, looking forward to comments/discussion.

If I can make one ‘quant’-oriented comment: defining priority as PxV is not an optimal strategy (we also are doing something more subtle under the covers I think but that’s a different story).

Of course there valid debates on how to balance customer centricity versus business needs and relevance versus value, but here is the thing: even if we would assume we want to maximize the sum of value, priority shouldn’t be PxV but should lean more towards P.

And that might seem strange at first sight for the mathematically inclined. They know PxV as expected value, say with P the likelihood of an event, says response/conversion, and V the pay-off on that event. All the good customer-centricity talk aside, why should that not be mathematically optimal?

Here is the catch: making offers are not independent trials or experiments. The propensity to accept an offer for a given customer is influenced by past propensities. We are like Pavlov dogs, if we get offered lots of food (actions) we don’t like, we stop drooling (subsequent propensities will be lower).

Assisted channels are a huge additional catalyst here. If I make ten offers in a row that get refused, even if customers are different, I will stop making offers (and will fool the poor CDH folks by gaming the response buttons). So here propensities are even negatively influencing each other accross customers.

There is only one situation where given a value centric objective, the optimal strategy is PxV: when this is the very last time you ever will make an offer to the customer. Not exactly something you would want :winking_face_with_tongue:.

What does the AI hive mind think? Any war stories about tanking results, even in total value of conversions, when cranking up the influence of V? How to balance the trade-off then?

Thanks Peter. Not surprised to see you jumping in as I know this is a subject dear to yout heart :slightly_smiling_face:.

I think you’re absolutely right on the non‑independence point. We’re not running isolated experiment, we are engaging customers repeatedly over time. If we keep putting things in front of them that they don’t want, we’re effectively training them to say “no”, which drags propensities down across the board. Trying to shoot elephants too often is a great way to come home empty handed. We need to train our customers to say “yes”.

Assisted channels make this even more obvious. Enough rejected offers and the call centre agents themselves will quite reasonably stop offering, and NBAs are not offered at all.

I’m hoping to write a follow‑up post that digs into this a bit more, because once you think in terms of long‑term behaviour rather than one‑off decisions, simple PxV as a priority rule starts to look pretty crude. A discussion of alternative utility functions, and why you might deliberately lean more towards P, would be a good place for me to go next.

None of that really changes the principles above though. Value in CDH isn’t just there to “win” decisions — it’s also the unit of account for reporting, simulation, and understanding what’s going on. So even if the decision logic is more subtle under the covers, we still need value to be stable, coherent, and explainable.

Interested to hear others’ experiences here too.

I have now written about this problem - and also some notes on the proper time required to defrost a woolly mammoth.

Beyond PxV: Accounting for Non-Spherical Customers in Next-Best-Action