Generative AI for Marketing: A Per-Use-Case Feasibility View Most marketing teams adopting generative AI mix three very different kinds of work into one budget line: tasks the current models can already automate, tasks that look automatable but quietly require super-human capability, and questions that need investigation before anyone commits to a build. Treating the whole bundle as “GenAI for marketing” is how a campaign budget burns through a quarter with nothing defensible to show. The cleaner approach is a per-use-case feasibility pass — each candidate gets classified before development, not after. This post applies that framework to common marketing GenAI use cases. It is the partner piece to our broader argument that a GenAI feasibility assessment classifies your use cases, estimates ROI for the automatable portion, and names explicit go/no-go criteria. Here we walk the classification through real marketing workflows. What does “technically feasible” mean for a marketing use case? Feasible means the current generation of language and image models can deliver the output at the quality bar the campaign requires, on the data you actually have, within a cost envelope that doesn’t erase the saving. Three classes cover the space: Automatable — the model already performs at or near the human baseline for this narrow task. Proceed with engineering; the ROI is real. Speculative — the task requires capability beyond what current models reliably produce. Do not commit to a build; if it matters, scope a research phase first. Research question — the task is plausibly feasible but unproven on your data or in your brand context. Proceed with bounded scope — a time-boxed investigation with explicit success criteria. This is an observed pattern from our consulting engagements: marketing teams routinely classify all three as “automatable” because the demos look convincing. The demos use clean data and forgiving evaluation. Your campaign will not. The classification, applied Marketing use case Class Why Drafting blog post outlines, ad copy variants, product descriptions Automatable LLMs handle structured copy at scale; quality bar is editor-reviewable Generating banner imagery and product mockups in a defined style Automatable Image models match style guides given reference assets; production-grade tooling exists Audience segmentation and tailored email subject lines Automatable Pattern recognition on first-party CRM data; measurable lift A/B test variant generation for landing pages Automatable Narrow generative task with clear evaluation (conversion rate) Forecasting campaign demand and optimising ad spend Research question Depends on data history and channel mix; needs validation before commitment Brand-voice fidelity across long-form content without editorial review Speculative Models drift on tone over long outputs; “no human in the loop” is not yet reliable Fully autonomous social media engagement (replies, escalations) Speculative Reputational risk on edge cases exceeds current model judgement Multi-turn customer chat that books demos and qualifies leads Research question Feasible for narrow flows; needs your conversation data to validate The split matters because the automatable column compounds quickly — a marketing org can realistically cut content production time and ship more campaign variants per quarter. The speculative column does not. Treating it as automatable is where GenAI budgets evaporate. Data readiness, before anything else Before classifying a use case as automatable, check whether your data supports it. Three quick signals: Volume and labelling — for personalisation and segmentation, do you have enough first-party interaction data, and is it structured enough to train on? CRM exports with inconsistent field semantics will not produce reliable segments, regardless of model quality. Brand asset coverage — for image generation in your style, do you have a few hundred to a few thousand on-brand reference images? Style transfer with a handful of samples produces uneven output. Evaluation data — can you measure whether the AI-generated content performs as well as the human-produced baseline? If you cannot define the comparison, you cannot defend the spend later. A GenAI use case where any of these three signals is missing belongs in the research column, not the automatable one. We treat data readiness as a gating check, not a parallel workstream — in our experience, projects that defer it spend twice: once on the build, once on the post-hoc data work. Measurable outcomes you should define before development The defensibility of the spend depends on what you committed to measure before the build started. For marketing GenAI, the set is small and concrete: Time-to-publish per asset class (blog post, ad creative, email) — before and after AI integration. Cost per published asset — including model inference, tooling licences, and reviewer time. Performance parity — engagement, click-through, or conversion on AI-produced versus human-produced assets, measured on matched audiences. Brand-voice rejection rate — how often editorial review sends an AI draft back for substantive rework, not just polish. Setting these targets before development is what makes the assessment defensible afterwards. If the campaign hits them, the spend justifies itself. If it misses, the same targets show where to stop or adjust — not as a post-mortem, but as a decision the project structure already accommodated. The boundary matters because the conversations look similar from outside. “Are we ready for GenAI?” and “Is this GenAI use case feasible?” both produce a slide with green and red boxes. Only one of them is about your specific campaign. FAQ How do I judge whether a specific generative AI use case is technically feasible with current models? Classify the task into automatable, speculative, or research. Automatable means current models perform at the required quality bar on data you have. Speculative means the task requires capability beyond what current models reliably produce. Research means feasibility is plausible but unproven for your context, so a bounded investigation precedes commitment. What does a structured GenAI feasibility assessment look like, and what does it answer? It enumerates your candidate use cases, applies the three-class filter to each, checks data readiness as a gating signal, estimates ROI for the automatable subset, and names explicit go/no-go criteria. The output is a defensible artifact that either justifies the spend or prevents the waste. Which use cases should we classify as automatable, speculative, or research — and why? Narrow, well-evaluated tasks with clear baselines (copy drafting, image generation in a defined style, segmentation, A/B variants) are typically automatable. Tasks that require sustained judgement, long-form brand fidelity without review, or autonomous handling of edge cases are typically speculative. Tasks that are plausibly feasible but unproven on your data (multi-turn chat flows, demand forecasting on your channel mix) are research questions and should be time-boxed before commitment. How do I assess data readiness before committing to a GenAI build? Check three signals: volume and labelling of first-party data, coverage of brand assets for visual or style work, and the existence of evaluation data that lets you compare AI-produced output against a human baseline. A gap in any of the three moves the use case from automatable to research. What measurable outcomes should we define before development starts so the spend is defensible later? Time-to-publish per asset class, cost per published asset, performance parity against the human-produced baseline on matched audiences, and the brand-voice rejection rate from editorial review. Define these before development so the assessment doubles as the post-hoc evidence. How does per-use-case feasibility relate to (and depend on) organisational AI readiness covered in TK5-CCU-07? Organisational readiness is the prerequisite — it decides whether to run an AI project at all. Per-use-case feasibility is the filter that decides which candidates inside that project to pursue. Readiness gates the project; feasibility gates the candidates. Both checks have to pass for the spend to be defensible. Closing The interesting question is not whether generative AI changes marketing — it does, and the automatable column is large enough to justify the investment on its own. The interesting question is whether your team can tell, before development starts, which use cases belong in which column. That is the assessment worth running. Image credits: Freepik.
