Decision orchestration for marketing automation: whether, when and where to send

1Context and motivation

Marketing automation began with batch campaigns and simple triggers: a newsletter on a schedule, an email after an abandoned cart. A store that sells across several regions and channels now runs many such programmes at once, across email, SMS, push, in-app messages and paid media, each with its own rules. The rules are written per programme, so no single place weighs one message against another, and none asks whether a given customer should receive anything at all today.

We treat that question as a decision problem. For each customer and each moment, the system chooses among a set of actions that always includes silence: no message. It chooses the action whose expected incremental value, net of costs and risks, is highest, subject to consent, legal limits, contact caps and budgets. This paper describes the architecture we propose for that decision engine and the literature each part rests on. We describe the design; we do not report its effects, because we have not measured them.

2Background

2.1Incremental effect, not response

A model that predicts who will buy ranks customers by propensity. A message, however, only creates value for customers whose behaviour it changes. Uplift modelling estimates that incremental effect directly, as the difference between the outcome with and without the action, and the literature groups the methods into a two-model approach, a class-transformation approach and direct modelling of uplift¹. Tree-based estimators with honest sample splitting extend this to valid confidence intervals for effects that vary across subgroups².

The difference matters in practice. In two field experiments on proactive churn programmes, the customers at highest risk of churning were not necessarily the best targets; targeting by sensitivity to the intervention did better than targeting by risk³. A related point holds for volume: in data from a home improvement retailer, the number of emails sent had a nonlinear effect on both short-term and long-term profitability⁴.

2.2Sequential and constrained decisions

A single message is one step in a relationship. Myopic mailing policies send to the customers most likely to order from the current catalogue and ignore how today's decision changes future orders; a dynamic policy that accounts for those long-run effects was evaluated in a large field test⁵. Reinforcement learning is the general framework for choosing actions to maximise cumulative reward over such sequences⁶, and it has been applied to ad recommendation to optimise customer lifetime value rather than the next click, together with methods to check a policy's safety on historical data before deployment⁷.

When only the next action matters, contextual bandits choose among actions using the customer's context and learn from the reward of the chosen action alone⁸. Marketing decisions also face supply and budget limits: a contact cap per week, a spend budget per campaign. Bandits with knapsacks model exactly this, a learner that explores and exploits while respecting one or more resource limits⁹.

2.3Measuring what a policy does

Advertising effects are hard to measure. Across 25 large field experiments, the median confidence interval on return on investment was over 100 percentage points wide¹⁰, and observational methods often failed to recover the effects measured in randomised advertising experiments at Facebook¹¹. A decision engine therefore needs randomised holdouts as its ground truth¹², and offline estimates of a new policy's value from logged decisions, for which doubly robust estimators combine a reward model with a model of the logging policy¹³.

3Design

3.1Inputs and state

The engine reads four groups of inputs. Customer state covers behaviour (views, carts, purchases), channel interactions (opens, clicks, unsubscribes), lifecycle stage, region, language, time zone, channel preferences and consent. Content covers templates, offers, creative constraints and the cost of each channel. Context covers time, season, holidays and stock. Constraints cover merchant budgets, contact caps and regional rules.

3.2Action space

For each decision the feasible actions are: no message; a message on one eligible channel now; or the same message deferred to a later window. Silence is a first-class action with its own estimated value, not a fallback.

3.3Eligibility filter

Before any model is consulted, a rule layer removes actions that are not allowed: channels without valid consent, contacts above a cap, sends outside permitted hours, and merchant exclusions. Where a customer has objected to direct marketing, the engine stops processing their data for that purpose, which the General Data Protection Regulation requires in the EU, including profiling related to direct marketing¹⁴. These are hard constraints; no learned score can override them.

3.4Incremental value estimation

For each remaining action, the estimator predicts the change in the outcome relative to silence, per customer, using uplift methods¹ and heterogeneous effect estimators with confidence intervals². The net value of an action is the estimated incremental margin minus the message cost minus a fatigue term: the estimated increase in unsubscribe or complaint risk, multiplied by the value of keeping the customer reachable. Customers are ranked by this net value, not by propensity³.

3.5Policy layer

High-frequency choices, such as which variant or channel to use for an eligible message, are made by a contextual bandit⁸ with resource constraints for caps and budgets⁹. Choices with long horizons, such as how often to contact a customer over a quarter, are framed as sequential decisions^{6, 5}. A new long-horizon policy is evaluated on logged data before it serves any traffic^{7, 13}.

3.6Monitoring, explanation and override

Every decision is logged with the action taken, the actions rejected and the reason for each (for example, "push chosen; email capped this week"). Operators can force or suppress actions for a segment and set merchant-level exceptions. The interface follows published guidance for human-AI interaction: make clear what the system can do and why it acted, and support efficient correction and dismissal¹⁵.

Table 1 maps each component to the problem it addresses and the evidence behind it.

Component	Problem it addresses	Evidence
Silence as an explicit action	Messages sent to customers they do not change	1, 3
Per-customer effect with intervals	Average effects hide groups that are not helped	2
Volume as a decision, not a setting	Email volume has a nonlinear effect on profit	4
Long-horizon contact policy	Myopic sends ignore effects on future orders	5, 6
Contextual bandit with resource limits	Choosing channel and variant under caps and budgets	8, 9
Offline policy evaluation	A new policy cannot be tried on customers untested	13, 7
Randomised holdouts	Observational lift estimates are unreliable	11, 10, 12
Objection stops direct-marketing use	Legal requirement in the EU	14
Logged reasons and operator overrides	Operators cannot see or correct why the engine acted	15

Source: our design, with the works cited in the table.

4A worked decision: send or stay silent

Table 2 shows the net-value calculation for three hypothetical customers. All probabilities and money values are inputs chosen to illustrate the arithmetic; none is a measurement.

Customer	P(buy), send (input)	P(buy), silence (input)	Uplift (computed)	Added unsubscribe risk (input)	Net value (computed)	Decision
A	0.050	0.030	0.020	0.002	0.56	Send
B	0.120	0.118	0.002	0.004	-0.01	Silence
C	0.010	0.004	0.006	0.002	0.14	Send

Source: Worked model; inputs in the table. Net value = uplift x 30 - 0.01 - added unsubscribe risk x 15.

Customer B has the highest purchase probability, so a propensity model would contact B first. B would buy almost as often without the message, and B's added unsubscribe risk is higher, so the net value is negative and the engine stays silent. Customer C has the lowest propensity but a positive net value. The example shows why the design ranks by estimated incremental value; whether the estimates would be accurate enough for this ranking to hold is an empirical question.

5Evaluation protocol

Before any claim about outcomes could be made, the engine would need to pass the following protocol.

• Persistent holdout. A randomly chosen share of customers receives only the rule-based journeys, so the engine's effect is measured against the current practice and not against no marketing at all¹².
• Pre-registered metrics. The target metric (incremental margin per customer over a fixed horizon) and guardrails (unsubscribe rate, complaint rate, contact volume per customer) are fixed before launch.
• Offline first. Each new policy is scored on logged decisions with a doubly robust estimator before it serves traffic¹³.
• Heterogeneity. Effects are reported per region, channel and customer group, not only on average².
• Power. Because advertising effects are small relative to the noise in individual sales, the sample size is planned before launch; some questions may need more traffic than a single merchant has¹⁰.

6Ethics, fairness and privacy

An engine that optimises incremental margin will learn to concentrate contact where it pays. That can mean over-targeting customers who respond to pressure and neglecting regions or groups where data is thin. The design addresses this in three ways: hard rules exclude sensitive attributes and vulnerable groups named by the merchant; contact volume and offer value are reported per group so that concentration is visible; and the decision log supports audit and recourse. Consent and objection are enforced in the eligibility filter, before modelling, and data use for direct marketing stops when a customer objects¹⁴. Whether these measures are sufficient is not something the design can show; it has to be checked in the reports a deployment produces.

7Limitations

• No measurements. We have not run this engine on real customers. Every benefit implied above is a hypothesis for the protocol in the previous section.
• Estimation error. Uplift estimates for individual customers are noisy, and the send-or-silence ranking in Table 2 depends on them. With little data per customer, the engine may act on noise.
• Fatigue term. The value of keeping a customer reachable and the added unsubscribe risk are hard to estimate; a wrong value biases the engine towards over-sending or under-sending.
• Measurement cost. Advertising experiments can require very large samples¹⁰; small merchants may not be able to measure the engine's effect at all.
• Regional rules. The eligibility filter encodes the rules we know of. Legal requirements differ by country and change, and the rule set needs legal review per market.
• Scope. The design covers outbound messages. Inbound and organic traffic appear only as context.

8Conclusion

We have described a decision engine that treats silence as an action, ranks customers by estimated incremental value rather than propensity, learns channel and timing choices under contact and budget limits, and logs every choice with its reason. Each part rests on published methods, summarised in Table 1. What the design lacks is evidence of its own effect. The next step is a holdout experiment on one programme, such as win-back messages, measured with the protocol above.