From Campaign Blasts to Autonomous Customer Journeys
Why point-in-time CRM scoring erodes customer lifetime value, and how enterprise brands are shifting to continuous relationship planning.
The Executive Case
- Most CRM systems optimize the next interaction, not the relationship over time. Point-in-time scores can miss whether a customer is gaining interest, cooling off, becoming discount-conditioned, or simply being over-contacted—creating margin leakage and avoidable churn.
- The opportunity is better timing, sequencing, and restraint. Instead of asking only who is most likely to respond now, the system evaluates how each action affects future customer value. In our controlled benchmark, this produced comparable customer revenue with 64% fewer contacts, more than 3× higher revenue per contact, and materially fewer unsubscribes.
- The approach can be proven before it is scaled. RunTime operates alongside the existing CRM and activation stack, first in shadow mode and then on a controlled production cohort. Its role expands only when measured incremental economics justify it.
The standard for modern CRM is straightforward: every customer decision should reflect what the enterprise knows, the history of the relationship, the actions available now, and how those actions are likely to affect future value.
You probably would not build a collection of independent propensity scores and next-best-action models. Nor would you simply load customer history into an LLM and ask what to do next.
The ideal system would maintain a continuously updated understanding of each customer: not only who they are and what they have done, but where the relationship is heading. Engagement rises and falls. Price sensitivity changes. Trust accumulates. Frustration compounds. And every intervention changes the customer state that determines what happens next.
The question therefore becomes: What should we do now—if anything—to maximize the value of the relationship over time?
- Conventional Aggregate-Feature Scoring: Standard XGBoost model trained on 30-day aggregate activity totals.
- RunTime Customer Decision System: Customer-history decision system optimizing cumulative 30-day net customer value.
Methodology note: The benchmark uses synthetic customer trajectories designed to isolate differences between decision approaches under controlled conditions. Production value should be established against each enterprise's incumbent system using a controlled customer cohort.
1. History & Sequence Dictate Customer Readiness
Customer decisions are inherently sequential. Each intervention changes the customer state that determines which actions will be effective next.
Many enterprise decision systems compress rich customer histories into engineered features and scores optimized for the next interaction. That can obscure the order, timing, and cumulative effect of prior interactions. In practice, two customers with identical 30-day activity counts can occupy completely opposite relationship states, requiring completely different actions:
2. Where Better Decisions Create Economic Value
A system focused only on immediate conversion can overuse discounts or outreach. RunTime explicitly accounts for the downstream effect of those actions, can prioritize service or education before an incentive, and treats no action as a valid decision when intervention would reduce expected customer value.
Point-in-time propensity scoring suffers from two major structural commercial flaws: over-discounting active buyers (which triggers ad fatigue and erodes margin) and over-contacting every customer on every cycle. Better timing and selective intervention can preserve customer value while materially reducing unnecessary contact:
3. Controlled Benchmark: Economic Impact
In the controlled benchmark, RunTime captured comparable customer revenue with materially fewer interventions—showing that much of the economic value associated with aggressive contact can be preserved through better timing and restraint.
| Model / Operating Strategy | Revenue / Customer | Contacts / 100 Cust | Revenue / Contact Sent* | Uncaptured Opportunity |
|---|---|---|---|---|
| Maximum Attainable Value | $19.10 | 50 | $38.56 | $0.00 |
| RunTime Customer Decision System | $17.38 | 36 | $48.22 | $1.73 |
| Next-Best Action Model (Immediate Focus) | $17.36 | 100 | $17.36 | $1.74 |
| Conventional Aggregate-Feature Scoring | $15.61 | 100 | $15.61 | $3.50 |
Executive implication: The benchmark does not suggest that more sophisticated decisioning creates value simply by generating more responses. Its principal advantage is identifying when intervention adds value—and when restraint preserves more of the relationship.
4. Evidence-Gated Enterprise Rollout
Enterprise teams should not need to commit to broad deployment before incremental value is established.
RunTime operates alongside the existing CRM and activation stack, supplying decision recommendations through the data and activation interfaces already used by the enterprise.
Start with a controlled production cohort. Measure incremental economic value against the incumbent decision system. Increase RunTime's allocation only after statistically and economically meaningful improvement is established.
How Proverify.ai Can Help
Proverify makes this practical without replacing the systems already in place. RunTime uses frontier AI to learn from customer histories, policies and existing workflows, then turns what works into a controlled decision process that can run quickly, consistently and with a full record of what happened. TokenOps tracks the economics around it—what models are being used, what they cost and what business value they produce.
Together, they let companies use the best available AI to reason over the full customer relationship, improve decisions as evidence accumulates, and change models or vendors without losing the operating knowledge they have built.
RunTime does not assume broader decision authority.
It earns it through measured performance.
Appendix
When might this not work?
When might this not work?
Consider Maya, a meal-kit customer. She skips a delivery, ignores a discount, contacts support, and later orders again.
Should the company send another offer, recommend cheaper meals, wait, or do nothing? Looking at her journey might help—but only if the journey is real, the order matters, and the next action changes what happens later.
Here are five situations where that may not be true.
There is no real journey to manage
Maya buys a meal-kit gift card for a friend. It is a one-off purchase with no meaningful follow-up decision.
There is little reason to plan several steps ahead. A simple prediction or rule may work just as well.
The order of events does not change the answer
Maya has skipped two boxes, opened three emails, and used one discount.
If the same offer is best whether the discount was used yesterday or six months ago, the sequence adds little. Totals and recency may tell the company everything it needs to know.
A useful check is to shuffle the events. If the model performs just as well, order is probably not helping.
The recorded journey is wrong
The data shows that Maya ignored an email and then placed an order.
What it misses is that she called support between those events and received free delivery. The model may wrongly credit the email for the sale.
Missing calls, disconnected identities, delayed timestamps, and unrecorded human actions can create journeys that never actually happened.
Past decisions hide cause and effect
Customer-service agents usually offer large discounts to customers most likely to leave. Those customers still churn more than average.
The data might suggest that discounts cause churn. In reality, the discount may have prevented even more customers from leaving.
Historical patterns alone cannot reliably show what a different action would have caused. That requires good action logging and controlled testing.
The model wins while the business loses
The system learns that a 30% discount gets Maya to order again. Response rises, so the policy looks successful.
But the order makes little profit, Maya learns to wait for discounts, and other customers receive so many messages that complaints increase.
Success must mean total net value—not clicks, response, revenue per message, or prediction accuracy. Costs, margin, churn, complaints, inventory, and customer fatigue all count.
Three questions to answer before proceeding
- Does order matter? Compare the true history with the same events shuffled in time.
- Do the recommended actions cause better outcomes? Test them in a randomized customer group.
- Does the policy create net value in the real world? Include costs and constraints, and begin in shadow mode before giving it control.
Any one of these issues can be enough to stop or simplify the approach. Passing all three tests means the idea is ready for a controlled production trial—not that benefit has already been proved.
Appendix
Should we expect Move 37s in CRM?
Should we expect Move 37s in CRM?
AI has repeatedly overturned expert assumptions in strategic domains—from Go to poker—by finding approaches humans had not seriously considered.
CRM may offer the same opportunity. The space of possible customer strategies is enormous, while historical CRM has explored only a small fraction of it.
Even six decisions with three choices each creates 729 possible sequences. Traditional CRM may have repeatedly used only a handful.
That means today’s models can become very good at learning what we usually did without discovering what we should do.
The opportunity is to systematically explore beyond the historical playbook and find high-value strategies humans may never have tried.
We should not assume every CRM problem contains a Move 37. But given what AI has already shown in other complex decision environments, we should expect some.
Your CRM learns what you did.
It needs to learn what you should do.
The problem: the actions you have historically tried are only a small part of the actions you could take.
Learns inside the old playbook
It gets very good at predicting outcomes from the tactics the company already used.
Searches the decision space
It asks how different actions and timings change the customer — and therefore what becomes best next.