Agencies, Churn Models & AI Disclosure: 3 New Studies
If your team is using AI to draft ad copy, score churn risk, or clean up client reports, three questions are getting harder to answer: what exactly is the human doing, how well does the AI actually perform, and who is accountable when it goes wrong? This week's AI marketing research pulls at all three threads — and the picture that emerges is less about capability and more about accountability.
We looked at three papers: a conceptual argument about how advertising agencies survive AI-driven automation, a machine learning study claiming 96% accuracy predicting e-commerce churn with a GAN model, and a mixed-methods analysis showing that AI disclosure norms in professional work are, essentially, broken.
Each has a real signal for marketers, brand managers, and agency leaders. Each also has limits worth naming out loud before you cite the findings in a client deck.
Quick Takeaway
- Agencies losing repetitive work to Meta and Google automation may need to rebuild services around AI oversight and cultural judgment.
- A GAN-based churn model hit 96% accuracy on a Kaggle dataset — promising, but untested in live e-commerce environments.
- Customer satisfaction scores outperformed purchase history as a churn predictor in the model tested.
- 98% of AI disclosures at one major CS conference lacked any statement of human accountability — a template gap marketers should not repeat.
- Two of the three papers offer frameworks, not proof. Treat them as hypotheses worth testing.
What This Research Means for Marketers
The through-line across these three papers is accountability. AI is doing more of the actual work — drafting creative, scoring customers, shaping strategy — but the systems around it haven't caught up. Nobody is auditing outputs at most brands, few teams are logging what AI did, and the tools available (from Meta Advantage+ to churn models) don't ask you to document anything. That's a compounding risk.
For marketing leaders, the practical move isn't to slow down AI adoption. It's to build the missing layer: a review process for AI-generated work, a data pipeline that captures the signals that actually predict retention, and a disclosure habit that goes beyond "AI tools were used." The teams that install these layers now will be the ones clients trust when something inevitably goes sideways.
Papers Covered
Paper 1: A Discussion on the Future of Advertising Agencies in the Impact of Artificial Intelligence
- Source / venue: Intermedia International e-journal
- Link: https://doi.org/10.56133/intermedia.1740804
- Source type: Peer-reviewed journal article (conceptual)
- Method: Conceptual/review synthesis of existing literature and platform examples (Meta Advantage+, Google Ads automation). No primary empirical data collection.
- Sample: No empirical sample — literature and industry example synthesis.
- Main finding: AI is absorbing the repetitive tasks agencies traditionally billed for (ad variations, targeting, media planning), but underperforms on cultural nuance, ethical judgment, and creative decision-making. The paper argues surviving agencies will restructure around new specializations — AI output auditing, brand safety oversight, culturally intelligent creative — rather than compete on speed.
- Evidence strength: Peer-reviewed but conceptual only; single author; lower-profile Turkish communications journal. Useful as a thinking framework, not validated practice.
- Limitation: No empirical data, no case studies, no measured performance of the hybrid model it proposes. Prescriptions are normative, not evidence-based.
- Practical implication: Audit which deliverables your team produced last quarter that AI could now draft — that's your automation exposure. The remainder (cultural, ethical, strategic judgment calls) is what clients still need humans for. Package accordingly.
Paper 2: Generative AI for Personalized Marketing and Customer Experience in E-Commerce
- Source / venue: International Journal of Emerging Research in Engineering and Technology
- Link: https://doi.org/10.63282/3050-922x.ijeret-v7i1p103
- Source type: Peer-reviewed journal article (model comparison study)
- Method: Built a GAN classifier on a public Kaggle e-commerce customer dataset and benchmarked against Logistic Regression, Random Forest, and Naïve Bayes. Included data cleaning, feature selection, and exploratory analysis.
- Sample: A public Kaggle e-commerce customer dataset. Exact record count is not specified in the available text.
- Main finding: The GAN model classified loyal versus at-risk customers at roughly 96% accuracy, outperforming the baseline models. Customer satisfaction scores were the strongest single predictor of loyalty in the dataset — more than purchase history. GANs were especially effective at detecting behavioral changes over time.
- Evidence strength: Peer-reviewed but in a lower-tier venue, single independent researcher, no institutional affiliation stated. Public dataset (not proprietary business data). Not a field test.
- Limitation: Accuracy figures come from one Kaggle dataset with unreported size and no deployment. Higher classification accuracy has not been shown to translate into better retention outcomes in a live environment.
- Practical implication: Prioritize collecting customer satisfaction signals in your CRM, and structure your data pipeline to capture behavior sequences over time, not just totals. Treat the 96% figure as a research artifact, not a benchmark to promise stakeholders.
Paper 3: Expectations and Practices around AI Disclosure in CS Research
- Source / venue: arXiv (preprint — not peer-reviewed)
- Source type: Preprint
- Method: Mixed methods: (1) policy analysis of 65 top computer science conferences; (2) survey of 109 CS researchers on when AI disclosure feels necessary; (3) computational analysis of 13,867 AI disclosure statements from EMNLP 2025 and ICLR 2026.
- Sample: 109 CS researchers surveyed; 13,867 disclosure statements analyzed; 65 conferences reviewed.
- Main finding: 35 of 65 conferences have AI disclosure policies, but they're vague. Researchers say disclosure matters most for high-judgment tasks (experiment design, analysis), but in practice most disclosures cover low-stakes work like text polishing. Statements of human accountability for AI-assisted work were missing from 98% of EMNLP disclosures and 77% of ICLR disclosures.
- Evidence strength: Preprint, not yet peer-reviewed. Survey sample is small (N=109) and CS-specific. Corpus analysis is large and quantitative — that portion is the strongest evidence in the paper.
- Limitation: Preprint status; findings limited to CS venues; survey self-report may not match actual disclosure behavior; text parsing may miss nuanced disclosure language.
- Practical implication: If your team uses AI in client deliverables, don't rely on a generic "AI tools were used" note. Specify which tasks AI performed and confirm human review and accountability — especially for strategy, research, and analysis work.
Plain-English Payoff
AI is getting faster at the work marketers used to charge for. What it still can't do well is judge — culturally, ethically, or strategically. The value in the next few years is in the review layer: who checks the AI's work, who owns the results, and who can prove it to a client. That's true whether you're rebuilding an agency, deploying a churn model, or writing a disclosure line at the bottom of a deck.
Money Move
Package the accountability layer as a service. Three angles worth testing: (1) an AI output audit for brands running automated Meta and Google campaigns — human review for cultural fit, brand voice, and ethical risk before ads go live; (2) a churn dashboard that combines satisfaction-signal capture with human-reviewed retention triggers for e-commerce clients; (3) an AI disclosure toolkit — templates, boilerplate, and audit language — for agencies producing client-facing research and strategy work. Each addresses a gap the research surfaces, and none require you to out-build the platforms.
Evidence Check
- Paper 1 is a conceptual review with no empirical data — the hybrid agency model it proposes is untested.
- Paper 2's 96% accuracy comes from one public Kaggle dataset of unspecified size, not live e-commerce data or an A/B test.
- Paper 2 is titled around "Generative AI" but the actual contribution is a GAN classification model — not LLM personalization.
- Paper 3 is a preprint; the 98%/77% accountability-gap figures come from corpus analysis, which is its strongest component.
- None of the three papers demonstrate causal business outcomes. Treat findings as directional.
- Two of three venues are lower-tier journals; peer scrutiny is limited.
What to Test Next
- Action step. Run an internal audit of last quarter's deliverables. Tag which ones AI could now draft (automation exposure) and which required cultural, ethical, or strategic judgment (irreplaceable value). Use the split to reshape service pricing and positioning.
- Action step. Add a customer satisfaction capture point to your CRM if you don't already have one. Compare its predictive power against purchase history on a segment of at-risk customers over one quarter.
- Action step. Draft a two-tier AI disclosure template — one for high-stakes work (research, strategy, analysis) and one for low-stakes work (editing, formatting) — and require it on every client deliverable that used AI.
- Action step. Pilot a human-review checkpoint before any fully AI-automated Meta or Google campaign goes live. Track catches (cultural, tonal, ethical) as a measurable QA metric.
How This Connects to AI Marketing Strategy
The pattern across these three papers is not that AI is more or less capable than expected — it's that the professional infrastructure around AI is missing. Agencies don't have a service model that reflects what AI actually does. E-commerce teams don't have data pipelines that capture the signals AI models actually need. Research and marketing professionals don't have disclosure norms that reflect the judgment risk AI introduces.
That's the story we keep tracking on the AI & Marketing Research Radar: capability is racing ahead of accountability, and the commercial opportunity is closing the gap. Every service, tool, or workflow that installs a human review layer, a documentation habit, or a quality signal on top of AI is defensible work — because the platforms won't build it, and the clients need it.
FAQ
Will AI replace advertising agencies?
The research reviewed here argues no — but it does suggest agencies will need to restructure. AI is absorbing repetitive tasks like ad variations, targeting, and media planning. What remains defensible is cultural judgment, ethical oversight, and creative strategy. The paper is conceptual, so treat the survival framework as a hypothesis worth testing, not a proven model.
How accurate is AI at predicting customer churn?
One recent study built a GAN-based deep learning model that classified loyal versus at-risk e-commerce customers at roughly 96% accuracy on a public Kaggle dataset — outperforming Logistic Regression, Random Forest, and Naïve Bayes baselines. The catch: it hasn't been tested on live business data or in an A/B test, so real-world performance is unknown.
What predicts customer churn better — purchase history or satisfaction scores?
In the GAN churn model study, customer satisfaction scores were the single strongest predictor of loyalty, outperforming other features including purchase history. That suggests marketing teams should prioritize capturing satisfaction signals in their CRM data, not just transaction totals.
What should an AI disclosure statement include?
Based on the disclosure research, a useful statement specifies which tasks AI performed (research, drafting, analysis, editing), how much human involvement occurred, and — critically — an explicit statement that the humans take responsibility for the output. 98% of disclosures in one major venue lacked that accountability language.
Is 'AI tools were used' enough as a disclosure?
The research suggests no. Generic disclosures tell readers almost nothing about which tasks AI handled or who reviewed the results. For marketing work involving strategy, research, or client-facing analysis, a task-level disclosure with an accountability statement is closer to what professional norms are moving toward.
Can small businesses use deep learning churn models?
Not easily on their own — GAN models require data science capability and clean behavioral data. The practical path for smaller brands is either (a) a plug-and-play tool that abstracts the model behind a simple dashboard, or (b) starting with satisfaction score capture and simpler predictive models before scaling up.
Where can agencies find new revenue as AI absorbs traditional services?
The research points to specializations AI can't replicate well: output auditing, brand safety oversight, culturally intelligent creative strategy, and AI disclosure and compliance work. These are review-layer services — they sit on top of AI output rather than competing with it.
Listen to the Episode
Sources and Further Reading
- A Discussion on the Future of Advertising Agencies in the Impact of Artificial Intelligence
- Generative AI for Personalized Marketing and Customer Experience in E-Commerce
- Expectations and Practices around AI Disclosure in CS Research
Related Big Plans Media:
About Big Plans Media
Big Plans Media helps marketers, educators, entrepreneurs, consultants, and business leaders translate AI marketing research into practical strategy. AI & Marketing Research Radar is produced by Big Plans Media and hosted by Evita, an AI-generated research briefing avatar trained on Dr. Eva Wolf's research framework.
