MT362 Unit 10 AI Adoption Proposal Example

Rule Everywhere, Model Under Test: An MT362 Unit 10 AI Adoption Proposal Built From Pilot Results

[Student Name]

Purdue University Global

MT362: Artificial Intelligence Applications for the Marketing Professional

Unit 10 AI Adoption Proposal

[Instructor Name]

[Date]

The chain, pilot results and estimates are composite; results were computed with the pilot's stated script, and the sources cited are real.

Recommendation

Three sentences carry the recommendation. Adopt the schedule rule for all 71,400 Repeat Delivery subscribers from July 2027. Continue the churn model on a randomized 25 percent of subscribers for two quarters, with discounts removed from its offer ladder. Name Renee Castillo, director of lifecycle marketing, as the accountable owner. A decision is requested by June 1 so the rule can launch with the July delivery cycle.

What the Pilot Measured

The Unit 8 pilot randomized about 6,220 subscribers into each of three arms, stratified by state and tenure, and followed them for 90 days. The control arm kept the current $10 cancel-page offer. The rule arm added the schedule calculator and a skip offer whenever a subscriber edited a schedule. The model arm layered offers timed by the churn score, drawn from the Unit 5 ladder, over everything the rule arm received. The primary measure was cancellation within 90 days, with guardrails on discount cost, unsubscribes, complaints and the age-band detection ratio.

Pilot Results, Printed From the Script

The control arm canceled at 16.40 percent, the rule arm at 14.91 percent and the model arm at 13.61 percent. Against control, the rule reduced cancellations by 1.49 percentage points and the model by 2.80 points. Both differences clear the adjusted threshold of 0.025 set in advance for each comparison.

The model's edge over the rule, 1.30 points, is a different matter. Its interval, computed at the same adjusted level, runs from minus 2.71 to plus 0.11 points. Because that interval includes zero, the pilot cannot rule out that the model adds nothing beyond the rule. The proposal reports the edge as unproven rather than rounding it into a win. Kohavi et al. (2020) caution against declaring a winner when the interval around a difference includes zero, and the proposal follows that caution.

No guardrail was breached. Complaints stayed below 0.2 percent of recipients in every arm, unsubscribes in the model arm ran 1.3 times the control rate, and the age-band detection ratio held at 0.86 after the Unit 7 fix.

The model beat the old practice; it did not beat the cheap rule, and the proposal says so.

Kept Subscribers Against Spent Discounts

Value is computed per enrolled subscriber over the 90 days, relative to control. Each retained subscriber is valued at $120 in margin, an estimate from finance and labeled as such. Discount spending per enrolled subscriber was $1.85 in control, $1.15 in the rule arm, where fewer subscribers reached the cancel page, and $3.65 in the model arm. The rule therefore adds $1.79 in retained margin and saves $0.70 in discounts, a net of $2.49 per enrolled subscriber. The model adds $3.36 in retained margin but spends $1.80 more on discounts, a net of $1.56, before its fee. The vendor's $48,000 annual fee, spread over a quarter of subscribers, adds about $2.69 per covered subscriber per year.

If the margin estimate is halved to $60, the rule nets $1.59 and the model loses $0.12, so the ranking does not change. Lemmens and Gupta (2020) argue that retention programs should be judged by profit after incentives, and on that standard the rule is the stronger program at either margin.

Why the Model Stays, Narrowly

The model is not dropped, for one reason. Its edge over the rule, if real, likely comes from timing rather than money: it reached some subscribers before they opened the cancel page. Ascarza (2018) found that the value of retention targeting lies in reaching customers whose behavior an offer can change, and timing may be how the model does that. The continuation tests that idea directly by removing discounts from the model's ladder and keeping only delay, bag-size and free-delivery offers. If prediction earns its fee without discounts, the model has shown real value; if not, it ends.

Alternatives Considered

The proposal rejects two other courses. One was adopting the model for every subscriber, on the strength of its 2.80-point reduction against control. That would commit the chain to a $48,000 annual fee and to the model's higher discount spending for an edge over the rule that the pilot could not confirm, and it would remove the comparison group needed to learn whether the edge is real. The other was dropping the model entirely, on the grounds that its net value trailed the rule's. That would be the safer financial choice, but it would end the test just as the pilot suggested a specific, testable reason the model might help, its timing. Continuing the model on a quarter of subscribers, without discounts and under automatic exit terms, costs half a year's fee and answers the question the pilot left open.

Costs of the Rollout

The rule requires no new spending beyond the web team's maintenance of the calculator, estimated at a few hours a month. The model continuation costs $24,000 in vendor fees over two quarters, about 40 analyst hours a quarter for score exports and the age-band audit, and the discounts its ladder no longer includes are a saving rather than a cost. If the continuation ends early under the exit terms, the quarter-to-quarter contract limits the loss to one quarter's fee.

Rollout

The rule launches for all subscribers in July 2027. The model continues for two quarters, July through December, on a randomized 25 percent, so its results remain measurable against the rule alone on the other 75 percent. The pilot's guardrails carry over unchanged.

Governance

Renee Castillo owns the program and reports results to the chief marketing officer each quarter. Three conditions end the model's continuation automatically, without a further decision: an age-band detection ratio below 0.8 in any quarterly audit, a complaint rate above 0.5 percent of recipients, or net value per covered subscriber that fails to exceed the rule's by at least the fee per subscriber. The vendor contract runs quarter to quarter. The governance design follows the federal AI risk management framework's call to assign clear accountability and to monitor risks continuously rather than once (Tabassi, 2023).

What the Proposal Does Not Claim

The pilot covered 90 days in two states. Whether saved subscribers stay for a year, and whether the rule's effect holds at full scale, remain open. The first two quarters of rollout will answer the second question, and a twelve-month follow-up of pilot subscribers will begin to answer the first.

References

Ascarza, E. (2018). Retention futility: Targeting high-risk customers might be ineffective. Journal of Marketing Research, 55(1), 80-98. https://doi.org/10.1509/jmr.16.0163

Kohavi, R., Tang, D., & Xu, Y. (2020). Trustworthy online controlled experiments: A practical guide to A/B testing. Cambridge University Press.

Lemmens, A., & Gupta, S. (2020). Managing churn to maximize profits. Marketing Science, 39(5), 956-973. https://doi.org/10.1287/mksc.2020.1229

Tabassi, E. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). National Institute of Standards and Technology. https://doi.org/10.6028/NIST.AI.100-1