Most AI-in-marketing initiatives start with a tool rather than a problem: a vendor demonstration, a competitor's announcement, or a board member's question about what "we're doing on AI" produces a procurement decision before anyone has written down what the organisation is actually trying to achieve. Davenport and Ronanki (2018) found that the AI projects which paid off were the ones matched deliberately to a specific objective type — process automation, cognitive insight from data too large or too fast for people to analyse unaided, or cognitive engagement with customers or staff — rather than a single "AI strategy" applied indiscriminately. An objective stated as "use AI in marketing" is not yet a usable objective; "identify which SME customer segments respond to which message, faster and more precisely than manual analysis allows" is.

That objective discipline is the first of five stages a responsible adoption moves through in order: objective, data readiness, tool or model choice, a scoped pilot, and governed scale-up. Each stage exists because skipping it creates a problem the next stage cannot fix — choosing a tool before checking data readiness means discovering the data problem only after money is committed, and scaling before a pilot has tested the real assumption means discovering a flaw at maximum cost rather than minimum. Fountaine, McCarthy and Saleh (2019) describe this as building the organisational foundations — data, talent, process, and clear accountability — that most companies underinvest in relative to the technology itself; an organisation that funds the model and starves the foundation gets a technically competent tool that nobody trusts or uses. The roadmap's value is less in any single stage than in the discipline of not skipping ahead to the one that feels most like progress.

The second roadmap stage is where most AI-in-marketing projects actually stall, and it has little to do with the model. Data governance and data management are often used interchangeably, but Khatri and Brown (2010) treat them as distinct: data management is the operational work of storing, moving and processing data; data governance is the decision rights that sit above it — who decides what a field means, who owns its accuracy, and who is accountable when two systems disagree about the same fact. A team can be excellent at data management, in the sense that queries run fast and dashboards load, while having no governance at all, in the sense that nobody can say with confidence whose definition of a key term is the right one.

Before trusting a model built on a dataset, that dataset has to be checked against specific quality dimensions rather than a vague sense that it "looks fine." Wang and Strong's (1996) widely used framework includes, among others, accuracy (does the value reflect reality), completeness (is anything missing that the model needs), consistent representation (strictly, whether data is held in a consistent format; this module stretches it, as "consistency" for short, to cover definitions of the same fact that disagree across systems), and timeliness (is the data current enough for the decision it's informing) as distinct dimensions that can each fail independently — a dataset can be perfectly accurate and still useless if it's six months stale, or complete and current and still untrustworthy if three systems define the same field three different ways. Redman (2016), reporting an IBM estimate, put the economic cost of exactly this kind of unmanaged data quality at around $3 trillion a year across the US economy, and the same pattern — "the AI got it wrong" turning out, on inspection, to be "the data was wrong and the AI faithfully learned it" — is a common reason a promising pilot never reaches scale.

A system can produce a systematically unfair outcome without ever being given a protected characteristic to work with, which is precisely what makes algorithmic bias harder to catch than deliberate discrimination. Obermeyer et al. (2019) demonstrated this in a widely cited study of a US healthcare algorithm that used historical healthcare spending as a proxy for healthcare need — a reasonable-seeming choice, since sicker patients generally cost more to treat — but because Black patients had historically received less care for the same level of underlying need, the algorithm learned to systematically underestimate their needs too, reproducing a historical inequity through a variable that never mentioned race. The lesson generalises well beyond healthcare: any feature correlated with a protected or disadvantaged characteristic can smuggle that characteristic's effect into a model that was never told to look for it.

Historical data compounds the risk, because a model trained on past decisions learns whatever pattern of past unfairness those decisions contained, and then a feedback loop can entrench it further: biased outputs change future behaviour, which produces more biased data, which trains the next version of the model to be more confident in the same error. Barocas and Selbst (2016) make the legal and practical case that this can produce real disparate impact even where no individual decision-maker intended discrimination at any point — intent is not a precondition for harm, and "the model is neutral because it doesn't use the protected characteristic" is not a defence once a proxy effect is demonstrated. A fairness audit, done proactively rather than only once a complaint arrives, means deliberately checking a system's outcomes across the groups it affects, not just its overall accuracy, and asking specifically what a suspiciously strong predictor might be standing in for.

A fairness audit finds a problem; governance is what determines whether that finding leads anywhere. AI governance names who is accountable for a model's decisions before it is deployed, not after something goes wrong and everyone discovers the accountability was never assigned — Mittelstadt et al. (2016), mapping the ethics-of-algorithms literature, note that many of the harms attributed to "the algorithm" are better understood as failures of the human process around it: nobody owned monitoring it, nobody had the authority to pause it, or nobody had agreed in advance what counted as an acceptable outcome. Good governance assigns a named owner to a model's decisions, sets out what oversight that owner is expected to exercise, and specifies in advance what happens when the model is found to be wrong, so the response to a fairness audit's finding is a known procedure rather than an improvised one.

Explainability means something different depending on who is asking. A customer needs a plain reason for a decision that affected them, delivered in terms they can act on; an auditor needs something closer to being able to inspect how the model actually reaches its outputs across many cases, not just one. Rudin (2019) argues, against a common assumption, that a simpler, genuinely interpretable model is often available for high-stakes decisions and preferable to bolting an after-the-fact explanation onto an opaque one purely for the sake of using a more impressive-sounding technique; the two kinds of explainability are not interchangeable, and a governance framework that only satisfies one of them has not actually solved the problem.

Two debates sit above any single AI project, and both are more contested than confident public commentary suggests. The first is when, or whether, artificial general intelligence — a system with human-like generality rather than narrow task competence — might arrive, and the range of expert opinion is wide enough that citing one voice as settled fact misrepresents the field. Anthropic's chief executive, Dario Amodei, told the World Economic Forum in Davos in January 2026 that he expects something close to AGI within a few years, possibly by 2027; Google DeepMind's Demis Hassabis has given a more cautious estimate, putting the odds at roughly 50% by 2030; and AI researcher Yann LeCun, who left Meta in late 2025 to found his own venture, has argued that "AGI" as a single threshold is poorly defined, and that current large-language-model architectures are very likely the wrong foundation for whatever comes next regardless of timeline. The disagreement is not noise around a true number; it reflects different views of what intelligence is and how current systems relate to it.

The second debate concerns safety and oversight as capability grows. The International AI Safety Report, chaired by Yoshua Bengio and now in its second full edition (Bengio et al., 2026), concludes that AI capability is advancing faster than the governance frameworks meant to manage it, and that gap is widening rather than closing — a conclusion worth taking seriously without treating it as proof of any specific catastrophic outcome. Regulation is trying to catch up on two different tracks: the EU's AI "Digital Omnibus", agreed in May 2026 and formally approved in June 2026, pushes back the AI Act's timeline for high-risk-system obligations (reported as December 2027 for stand-alone high-risk systems and August 2028 for AI embedded in regulated products), while the UK has taken a lighter, sector-by-sector approach with no dedicated AI statute, regulating mainly through existing bodies — both positions current as of this course's build date and likely to move again. A third concern, AI's energy cost, is real but its figures vary enormously by methodology: Google's own disclosure for its Gemini model (August 2025) put the median text prompt at around 0.24 watt-hours, while OpenAI's Sam Altman separately disclosed roughly 0.34 watt-hours per ChatGPT query (June 2025) — different models, different measurement boundaries, both genuine. The International Energy Agency's Energy and AI report (April 2025) put global data-centre electricity demand at around 415 terawatt-hours in 2024 (roughly 1.5% of the world total), projected to more than double to around 945 terawatt-hours by 2030 — a serious and growing cost, but one that should be cited with its range, source and date attached, not as a single fixed number.