I85PREP
⌘ K
DMI Expert study materials
DMI Expert study resources

DMI Expert: AI in Digital Marketing

DMI Expert · Work through your exam, one topic at a time.

All guide sections

AI in Digital Marketing

What this domain tests

AI in digital marketing tests practical judgement: where machine learning, generative AI, and automation create value; how to prepare data; how to evaluate outputs; and how to govern risk. Expect use cases across search, paid media, social, content, personalisation, and analytics, plus prompt design, retrieval-augmented generation (RAG), hallucination, bias, privacy, copyright, and human review.

Core rules

  • Start with the decision, not the model. A use case needs a baseline, an owner, a metric, and a failure cost. 'Use AI' is not a use case; 'reduce first-response time by 30% without lowering CSAT' is.
  • Match the technique: rules for deterministic tasks, predictive ML for scoring and forecasting, generative AI for drafting and summarisation, RAG for grounding answers in approved sources.
  • Evaluate on held-out data and business outcomes. Accuracy alone misleads on imbalanced data; use precision, recall, and cost-weighted errors.
  • Keep a human in the loop for regulated, reputational, or high-value decisions. Log prompts and outputs, red-team for prompt injection, and document data provenance.
  • Governance covers consent, retention, model access, copyright and licence checks, bias testing, and vendor due diligence.

Worked example

A lead-scoring model reviews 1,000 leads. 100 actually convert. The model flags 200 leads; 60 of those convert. Precision = 60/200 = 30%. Recall = 60/100 = 60%. The model catches most converters but wastes sales time on 140 false positives. If sales capacity is scarce, raise the threshold: fewer flags, higher precision, lower recall. If missing a converter is expensive, lower the threshold. The reasoning: choose the threshold from the cost of false positives versus false negatives, not from the model's default.

Traps

Treating fluent output as fact, testing on training data, data leakage, automating a broken process, ignoring consent and copyright, failing to label AI-generated content where required, and assuming a higher accuracy score means a better business model.