Marketing is becoming dependent on AI, but it should not be ruled by it

Marketing is already dependent on artificial intelligence in a practical sense. Advertising platforms use machine learning to bid, target, predict conversions and moderate content. CRM systems score leads and recommend next actions. Analytics tools classify visitors, forecast outcomes and detect anomalies. Search engines use AI to interpret queries and construct answer experiences. Generative tools now draft copy, reshape campaign assets, create variants, summarise research, translate material, produce audience hypotheses and answer customer questions.

The more useful question is not whether AI has entered marketing. It has. The harder question is whether marketing can retain its judgment once AI becomes the default layer between a brand and almost every decision it makes.

That distinction matters because using AI and becoming dependent on AI are not the same condition. A company uses AI when it applies the technology to defined tasks and can still understand, challenge, measure and replace the output. It becomes dependent when campaign work, customer knowledge, creative production, media allocation or commercial decisions cannot continue at an acceptable standard without a particular model, platform or automated workflow. At that point, the organisation has not merely added a tool. It has transferred part of its operating capacity to systems it may not fully understand or control.

The pressure to transfer that capacity is real. McKinsey’s 2025 global survey found that 88 percent of respondents said their organisations regularly used AI in at least one business function, while marketing and sales remained among the functions with the most reported use. Yet only around one-third said their organisation had begun scaling AI across the enterprise, and only 39 percent reported any enterprise-level EBIT impact from AI. The gap between widespread use and proven company-wide value is the central fact that marketing leaders should keep in view.

AI can reduce the cost of producing a campaign. It can also reduce the cost of producing a bad campaign, a misleading claim, a repetitive creative idea or a flawed personalisation programme. Lower cost does not turn weak work into strong work. It only allows weak work to travel further and appear faster.

Dependence has already arrived through workflow

The first form of AI dependence is not dramatic. It appears as convenience. A marketer asks a model to create the first version of a brief, then asks it to produce social variants, email subject lines, display copy, landing-page sections and performance summaries. A creative team uses generative image tools to prepare concepts before a shoot. A growth team relies on automated campaign recommendations because the advertising platform claims it can find a better audience than manual targeting. A sales-marketing operation lets a scoring model decide which leads deserve attention.

Each decision can look harmless in isolation. The problem emerges when the workflow no longer contains a meaningful human checkpoint. The marketer becomes a reviewer of machine-generated options rather than the person who defines the commercial problem, the customer tension and the acceptable trade-off. The creative director sees outputs but does not see the training data, the implicit visual conventions or the cultural shortcuts that produced them. The performance specialist trusts a dashboard without questioning whether the metric is measuring incremental value or merely capturing people who were likely to convert anyway.

Dependence begins when a team confuses speed of output with quality of thinking.

This is why AI adoption should be treated as a workflow redesign project rather than a software purchase. McKinsey’s survey makes the point clearly: the organisations reporting the strongest AI value are much more likely to redesign workflows, define when human validation is required and set growth or innovation goals alongside efficiency. They do not simply place a model on top of an old process.

Marketing is particularly vulnerable to superficial adoption because so much of its visible output is linguistic or visual. If a model can create a reasonable-looking email in seconds, it is easy to assume the hard part has been completed. It has not. The hard part is deciding whether the email says something the customer needs to hear, whether it is distinctive against competing messages, whether it is true, whether it fits the brand, whether it reaches a person at an appropriate moment and whether it creates value beyond a short-term click.

AI is good at producing plausible forms. Marketing has always been judged by whether those forms create a meaningful response in the market.

Marketing’s old operating model has been rewired

For decades, marketing systems were built around a division of labour. Research created inputs. Strategists translated those inputs into positioning and direction. Creative teams produced messages and assets. Media teams bought reach and frequency. Analysts measured response. Agencies and internal departments coordinated each stage, usually with visible handovers and enough time for disagreement.

AI compresses those stages. A single marketer can now move from a prompt to an audience segment, a draft campaign, dozens of asset variations, a translated version and an early performance report in the space of an afternoon. This changes the economics of campaign development. It also changes the organisational temptation. If one person can apparently do the work of several people, leadership may assume that several people are no longer needed.

That assumption is often wrong because the old roles did not exist only to produce artefacts. They existed to introduce different forms of judgment. Research challenged assumptions about audiences. Strategy clarified choices. Creative work made a message memorable. Media planning balanced reach, context, cost and repetition. Analytics tested whether an apparent result was real. Removing those checks can make a marketing operation cheaper while making it less reliable.

A faster production line is not automatically a better decision system.

The change is most visible in performance marketing. Platform automation can handle bidding, budget pacing, audience expansion and creative rotation at a scale no manual team can match. The platform also sees more data than any individual advertiser. That creates a reasonable case for automation. It does not create a reason to hand over the definition of success. A platform is designed to improve outcomes within the objective it receives. If the objective is weak, narrow or commercially misleading, the automation can become highly efficient at producing the wrong result.

Consider a business that optimises campaigns for cheap leads. An AI-driven media system may find people who readily complete a form, enter a competition or download a low-value asset. The dashboard improves. Sales quality falls. The platform has not failed. It has done precisely what it was asked to do.

The same pattern appears in content. A team can use AI to increase publishing volume while reducing the time spent interviewing customers, testing product claims or developing a point of view. The content calendar fills. Search visibility may even rise in the short term. Yet the brand becomes easier to ignore because its material resembles everything else produced from the same public language patterns.

The capability trap

AI creates a capability trap because its first successes are often genuine. A team saves time on transcription, first drafts, localisation, campaign reporting or routine customer questions. Those gains build confidence. The organisation then expands usage into higher-stakes tasks without increasing governance, evaluation or senior accountability.

The move from assistance to decision-making is where the risk changes. Drafting a first version of a newsletter is not equivalent to approving a regulated product claim. Summarising customer-support themes is not equivalent to deciding which customers are eligible for an offer. Generating product images for internal concept work is not equivalent to publishing synthetic imagery that consumers may mistake for real product photography. Recommending budget changes is not equivalent to setting the commercial priorities that determine where the budget should go.

The technology often obscures these transitions because the interface looks identical. A chat window can produce a harmless summary and a dangerous assertion with the same confident tone. An agentic workflow can retrieve internal documents, make a recommendation, trigger an email sequence and update a CRM record. The number of actions it can take does not tell a company whether it should be allowed to take them.

McKinsey reported that 62 percent of respondents said their organisations were at least experimenting with AI agents, although only 23 percent said they were scaling an agentic system somewhere in the enterprise. In any individual function, no more than 10 percent reported scaling AI agents. Those figures suggest that experimentation is moving faster than operational maturity.

Marketing should not treat agentic capability as permission for agentic authority.

A useful distinction is between reversible and irreversible actions. An AI tool may safely prepare a draft audience analysis that a human can challenge. It should face much stricter controls before it automatically excludes customers from an offer, changes pricing messages, purchases inventory, makes public claims or sends high-volume communications under a brand name. The more difficult an action is to reverse, the more visible human responsibility should be.

This is not an argument for keeping every process manual. It is an argument for designing escalation. AI can prepare, classify, recommend and execute bounded tasks. Human owners should define the boundaries, approve sensitive changes and review the evidence when an automated system affects customers, reputation or revenue.

The strongest use cases are operational, not theatrical

Much public discussion of AI marketing focuses on spectacular outputs: image generation, synthetic presenters, hyper-personalised advertising or autonomous campaign agents. Those use cases draw attention because they are easy to demonstrate. Many of the most useful applications are less visible.

AI can help teams find patterns in customer-service conversations, identify repeated objections in sales calls, organise qualitative feedback, compare product documentation, create first-pass translations, summarise campaign histories and detect inconsistent information across websites. These tasks are valuable because they reduce administrative friction and give specialists more time to investigate what matters.

A strong implementation starts with a narrow question. Can the tool shorten the time required to locate recurring reasons for churn? Can it help a team compare hundreds of product reviews without losing the ability to inspect the underlying examples? Can it help sales and marketing identify content gaps that delay a purchase? Can it assist a localisation team by producing a first draft that a native speaker can improve? Can it find broken references between a campaign claim and a product specification?

Those are meaningful uses because the answer can be tested. A company can compare the old and new process, measure error rates, calculate time saved and ask whether the output improved a decision. The work remains attached to an actual business problem.

By contrast, a vague objective such as “use AI in marketing” creates activity without discipline. It encourages teams to buy overlapping tools, generate unnecessary output and measure adoption instead of value. The resulting dependency is expensive because it is not linked to a specific improvement.

The most defensible AI use cases remove friction from work that already has a clear owner and a measurable standard.

This is also where marketers can avoid a common mistake: treating generative output as the main value. In many cases, the bigger gain comes from better retrieval, better classification or better preparation. A model that helps a strategist find the relevant customer evidence may be more valuable than a model that writes the final campaign line. The first improves the quality of the decision. The second can only improve the speed of expression.

Content is cheaper but not automatically more valuable

Generative AI has pushed the marginal cost of marketing content down sharply. Copy variations, product descriptions, social posts, campaign concepts and routine articles can now be produced in volumes that were previously impractical. This has changed the content market. It has not changed the reader’s ability to recognise repetition, weak claims or empty language.

The danger is not that every AI-assisted piece of content will be poor. The danger is that organisations will decide that quantity is proof of relevance. It is not. Most customers do not need more generic explanations of a category. They need credible answers, useful evidence, clear comparisons, honest limitations and material that reflects an actual understanding of their problem.

A brand that uses AI to produce fifty versions of the same bland claim does not create fifty opportunities. It creates fifty chances to be ignored.

Search is part of this pressure. Marketers can see AI-generated summaries and answer experiences reducing the visibility of conventional search results for some queries. That creates an understandable impulse to publish more pages, more FAQs and more “answer-ready” copy. Google’s own guidance is more restrained. It says existing SEO foundations remain relevant for AI Overviews and AI Mode, that there are no special technical requirements for inclusion, and that site owners should focus on helpful, reliable, people-first content, crawlability, clear text, useful media and accurate structured data.

The response to AI search is not to create synthetic content at industrial scale. It is to create information that deserves to be cited, linked and remembered.

That demands firsthand material. It may include original research, product testing, expert interviews, real case studies, clear methodology, technical documentation, local knowledge, proprietary data or a point of view earned through experience. AI can help organise such material. It cannot generate the underlying experience without creating a fiction.

This is why content strategy must separate “content production” from “knowledge production.” Production is the act of turning inputs into formats. Knowledge production is the harder work of discovering something worth saying. AI can accelerate the first. It can weaken the second if teams stop collecting evidence because the model can fill the page without it.

Audience knowledge remains human work

Marketing platforms have always promised a more precise view of the customer. AI increases that promise. A model can group behavioural signals, identify likely interests, predict purchase propensity and recommend the next message. The outputs can look highly specific. The risk is that precision in a model is mistaken for understanding in the real world.

An audience segment is not a person. It is a statistical construction based on available signals, measurement choices and platform assumptions. It may reflect genuine patterns. It may also reflect past bias, incomplete data, temporary behaviour or proxies that seem useful until they are not.

A retailer might infer that a customer is price sensitive because they visit discount pages. The same customer may be researching a gift, comparing options or simply checking availability. A travel brand might assume that a user who searches frequently is close to booking. They may instead be anxious, undecided or helping someone else. A financial-services firm may classify people by inferred life stage without understanding the consequences of that classification.

AI can make segmentation more granular while making the marketer less curious about the person behind the segment.

That is the human responsibility that cannot be automated away. Good marketing research does not only look for correlations. It asks why a pattern exists, whether it will persist, who is missing from the data and what a customer might find intrusive or unfair. It uses interviews, customer-service evidence, sales conversations, observation and cultural knowledge. These sources are slower than a model output. They are also where a business discovers the motives that data alone cannot reveal.

A practical standard is to require a human-written explanation for every important audience rule. The explanation should state what the organisation believes, what evidence supports it, what uncertainty remains and what customer harm could occur if the belief is wrong. If a team cannot explain a segment without repeating platform labels, it does not understand the segment well enough to automate a high-stakes decision.

This is especially important when a model’s recommendations involve exclusion. Marketing teams often focus on who should receive a message. They should pay equal attention to who might be excluded, over-targeted or placed into a category that follows them across channels.

Personalisation exposes the data problem

Personalisation is usually presented as the commercial reward for AI. Better data, better predictions, better timing and more relevant messages should improve customer experience. That can happen. Yet personalisation often exposes the weakest parts of an organisation’s data practices.

A model cannot create trustworthy customer knowledge from fragmented records, unclear consent, inaccurate profiles and disconnected systems. It can only make faster inferences from the material it is given. If identity resolution is wrong, the personalisation is wrong. If transaction data is incomplete, the recommendation is weak. If consent records are unclear, the campaign may be difficult to defend. If the organisation cannot explain why someone received a message, it will struggle when a customer asks.

Data protection regulators repeatedly stress that AI systems raise questions about transparency, fairness, accuracy, lawfulness, accountability and bias. The UK Information Commissioner’s Office guidance explicitly treats fairness, transparency, statistical accuracy, security and data minimisation as core parts of the AI lifecycle rather than optional technical extras.

Personalisation is not a creative feature. It is a governance decision with customer consequences.

The most important data question is not “How much data do we have?” It is “Which data do we have a legitimate, clear and useful reason to use?” Marketing teams sometimes speak as if more data automatically produces more relevance. In practice, excessive or poorly governed data can make relevance feel like surveillance. A message may be technically tailored and emotionally wrong.

The gap between legal permission and customer comfort is significant. A company can satisfy a narrow internal approval process and still create distrust by using a signal that feels unexpectedly intimate. For example, a retailer may infer health, financial stress, family status or personal preference from behaviour that a customer never expected to become a targeting input. The business may call this relevance. The customer may call it intrusion.

AI makes that gap easier to widen because it enables more inferences, more combinations of data and more automated decisions. Good marketing governance should treat the customer’s reasonable expectation as a design constraint, not a public-relations issue to manage after launch.

Customer experience exposes the handoff problem

Marketing often treats the customer journey as a sequence of messages. Customers experience it as one relationship. They do not separate the advertisement, the website, the chatbot, the order confirmation, the support response and the refund process into departmental categories. The brand owns all of it.

AI can make this problem more visible. A company may use a polished AI-generated campaign to promise simplicity and speed, then hand customers to a badly configured chatbot that cannot resolve a basic issue. It may use an agent to answer pre-sale questions but require human intervention when money, eligibility or exceptions are involved. It may personalise a message using customer data but fail to recognise the same customer when they contact support.

The result is not merely operational inconsistency. It is a credibility failure. The customer sees a brand that can predict what they might buy but cannot understand why they are dissatisfied.

This is where marketing dependence on AI becomes dangerous. When leadership measures the marketing team only through acquisition, lead volume, click-through rate or cost per conversion, it can overlook the downstream experience. AI makes it easy to generate more demand. It does not ensure that the organisation can fulfil the promise.

Every AI-driven marketing claim should be tested against the service experience that follows it.

A practical method is to treat the campaign and the customer-service journey as one system. Before launch, teams should test the questions a campaign is likely to create. Can the chatbot answer them accurately? Can a customer reach a human when the issue becomes sensitive? Are product details consistent across ads, landing pages, retail staff and support channels? Can the organisation identify when an AI response has failed? Does anyone own the recovery process?

These questions may sound basic. They are often neglected because the tools appear in different departments. Marketing buys a content platform. Customer service buys an agent platform. Data teams manage identity. Legal reviews policies. The customer sees one brand and expects one level of competence.

Agentic marketing changes the control model

Generative AI produces outputs. Agentic AI can take actions. That distinction is central to the next phase of marketing automation.

An agent can potentially retrieve campaign data, read product information, generate an audience list, draft assets, submit items for approval, update a CRM field, send a message, adjust a budget or create a report. The commercial appeal is obvious. Marketing contains many repetitive tasks that depend on moving information between systems. Agentic workflows promise to compress that coordination.

The danger is that the agent inherits all of the weaknesses in the systems it connects. A flawed product feed, an outdated promotion, an ambiguous customer record or an incorrectly defined business rule can travel through several automated steps before a human notices. The more systems the agent can access, the more important permissions, logging and escalation become.

McKinsey’s survey describes agents as systems capable of acting in the real world by planning and executing multiple steps in a workflow. It also finds that use is still concentrated in a small number of functions and remains far from widespread scaling. That should encourage restraint rather than panic. The technology is powerful enough to justify preparation, but it is not mature enough to justify careless delegation.

A practical map of AI dependence in marketing

The purpose of this table is not to declare any activity safe or unsafe by itself. The same task can be low risk in one context and high risk in another. An internal draft that never leaves the company is different from a customer message that affects a purchase decision. A model that suggests a budget change is different from a model that applies it across multiple markets without review.

The correct control level follows the consequence of the action, not the novelty of the technology.

Measurement needs more than dashboards

AI has made marketing measurement more attractive and more difficult at the same time. It can process more data, identify patterns faster and create dashboards that make a campaign look legible. Yet the core problem remains: marketing results are rarely caused by one channel, one message or one platform setting.

A model may identify that customers who saw an advertisement were more likely to convert. That does not prove the advertisement caused the conversion. The people who saw it may already have been more likely to buy. A model may attribute credit to the last visible interaction because it has a clean record of that event, even though earlier brand activity created the demand. A reporting system may celebrate a high return on ad spend while overlooking margin, cancellation rates, returns, customer lifetime value or the impact of discounting.

AI can make these mistakes more persuasive because it can generate a coherent explanation around a weak causal assumption. The chart is polished. The narrative is clear. The recommendation sounds data-led. The evidence may still be thin.

A dashboard is an interface to a measurement model, not proof that the model is correct.

Marketing leaders need to preserve a small set of harder questions. What would have happened without this activity? Which customers are genuinely incremental? Does the result hold across time, geography or audience groups? Has the campaign created demand or captured people who had already decided? Did it improve profit, not merely revenue? Did it help the brand in a way that will still matter after the immediate promotion ends?

AI can assist with some of this work. It can identify anomalies, segment data, prepare test designs and surface inconsistencies. It cannot remove the need for experimental thinking. The most useful teams will use AI to make measurement more rigorous, not merely more automated.

This also means resisting false precision. A model can assign a decimal score to a lead or a probability to a conversion. That number is not automatically a fact about the customer. It is an estimate produced from a particular training process and available inputs. Teams should know the difference between a useful prediction and an authoritative explanation.

Search discovery is changing, but the foundations remain

AI is changing how people discover information. Search results increasingly include summaries, conversational responses and links selected through broader query expansion. Customers also ask questions in chat interfaces, compare products through assistants and use social platforms as search tools. Marketing teams have responded with a new vocabulary around generative engine optimisation, answer-engine visibility and AI discoverability.

The underlying requirement is less exotic than the terminology suggests. A brand needs to publish clear, useful, technically accessible information that can be understood by people and machines. It needs accurate product data, well-structured pages, visible expertise, consistent facts, internal links, relevant media and a reputation strong enough for others to cite.

Google’s documentation says AI Overviews and AI Mode use different models and techniques, may surface a wider and more diverse set of supporting links, and do not require special markup or a new “AI text file” for inclusion. Its advice remains rooted in standard search fundamentals: crawlability, useful content, accurate structured data and a good page experience.

The brands most likely to remain visible are not those that invent the most AI jargon. They are those that make their information easiest to trust.

This is a crucial point for marketing dependency. If a company sees AI discovery only as a technical distribution problem, it may produce optimised but weak material. It may write pages designed to answer every question while offering no original evidence. It may use AI to generate schema, FAQs and summaries without checking whether the content underneath is accurate.

A more durable approach starts with the information customers need before, during and after a purchase. What questions create friction? What comparisons are hard to make? Which product limitations are hidden in technical language? What evidence would help a buyer trust the company? Which claims are repeatedly misunderstood? Which local or expert details cannot be copied from public sources?

Answering those questions well creates material that can support conventional search, AI-driven discovery, sales enablement and customer service. The format may change. The underlying evidence still matters.

Creative distinctiveness becomes scarcer

AI can generate competent-looking work at extraordinary speed. That is why creative distinctiveness becomes more valuable, not less.

Most generative systems are trained to produce outputs that are recognisable as appropriate for a prompt. Ask for a premium technology campaign, a modern finance illustration, a clean B2B landing page or a luxury social image, and the tool will often generate something polished. It will also tend to generate familiar visual and linguistic patterns because familiarity is part of what makes the output seem plausible.

The result is a growing field of polished sameness. Brands may receive content that looks professional but has no proprietary character. The typography feels borrowed. The product story could belong to any competitor. The language is full of broad benefits that no one could dispute and no one will remember.

AI can produce style. It does not automatically produce a point of view.

A real creative idea does more than fit a category. It changes the way an audience sees the category, the product or the brand. It may contain a surprising observation, a precise cultural reference, an uncomfortable truth, a technical detail that competitors avoid, an unusual composition or an emotional tone that only makes sense for that company. These qualities emerge from knowledge, taste, risk and editorial judgment.

AI can contribute to the process. It can generate rough material, explore visual directions, help teams compare wording or create prototypes. It should not become the final arbiter of what is distinctive because its default tendency is toward the average of what already exists.

This has implications for brand systems. Companies need clearer rules about tone, visual identity, product truth, approved claims and cultural boundaries. Otherwise, every employee with a model subscription can create material that technically uses the brand name while gradually eroding the brand itself.

The stronger the AI capability, the more disciplined the brand stewardship must become.

The multiplication of plausible mediocrity

The main economic effect of generative AI in marketing may be an explosion of plausible mediocrity. It lowers the barrier to entry for basic copy, basic visual production, basic campaign planning and basic reporting. That will help smaller teams compete in areas where production budgets once created a major disadvantage.

It will also flood channels with material that is competent enough to publish and weak enough to forget.

This does not mean AI will eliminate quality. It means the baseline will rise while the average may become less distinctive. A decent landing page will be easier to create. A decent ad variation will be easier to test. A decent email sequence will be easier to automate. The competitive question moves upward: who can produce evidence, judgment, trust and creative memory that cannot be generated from generic instructions?

The answer often lies outside the model. It lies in product quality, customer relationships, distribution, field knowledge, founder perspective, community, service experience, proprietary research and the ability to take a position that carries risk.

When everyone can produce acceptable content, acceptability stops being an advantage.

Marketing leaders should be careful not to reward teams only for output volume. A team may produce twice as many campaigns and create half as much commercial value if the work becomes repetitive. Better metrics include the quality of customer response, qualified demand, brand recall, conversion quality, content usefulness, earned references, repeat purchase and the cost of correction when an automated output is wrong.

The same principle applies to agencies. AI may reduce time spent on production, but it should increase the value placed on diagnosis, strategy, concept development, editorial discipline and accountability. An agency that merely sells automated content volume is vulnerable to being replaced by the next tool. An agency that helps a business understand its market and make better choices remains useful even when the production tools change.

Brand safety, intellectual property and provenance

Marketing has always carried legal and reputational risk. AI adds new paths for those risks to appear.

A generated image may resemble protected work or create an implication that a brand cannot support. A model may invent a product feature, a customer testimonial, a statistic or a comparison. A synthetic voice may create uncertainty about consent. A campaign may use AI-generated material that looks authentic without being clear about its origin. A model may reproduce sensitive information from a prompt, a connected data source or an internal document.

The right response is not a blanket ban. It is to distinguish between internal experimentation, low-risk public output and high-risk commercial communication.

Internal concept work can often tolerate greater flexibility because the output remains inside the organisation. Public work needs a stronger chain of review. Claims must be checked against approved product information. Images need rights and disclosure assessment. Testimonials need to be real. Comparisons need evidence. Regulated categories need specialist review. Sensitive customer data should not be placed into unapproved systems simply because a model can process it.

NIST’s AI Risk Management Framework is voluntary, but its premise is useful for marketers: trustworthiness should be built into the design, development, use and evaluation of AI systems rather than added after a failure. NIST’s Generative AI Profile also identifies generative-AI-specific risks and proposes risk-management actions for organisations.

The marketing department should treat provenance as part of brand governance, not as a technical afterthought.

Content provenance standards such as C2PA can help record information about the origin and editing history of digital assets. They are promising, but they are not a complete solution. A 2026 independent analysis of C2PA found weaknesses that could prevent current specifications from meeting all claimed security goals, particularly in high-stakes contexts.

The practical lesson is simple. Labels, metadata and credentials can support transparency, but no technical signal removes the need for editorial accountability. A company still needs to know who approved an asset, what evidence supports it and what will happen if a customer challenges it.

Regulatory pressure reaches marketing operations

For years, many marketing teams treated AI regulation as a concern for legal departments, technology vendors or data scientists. That separation is becoming less realistic. Marketing sits directly at the point where AI outputs meet people. It is where profiling, targeting, persuasive design, automated communications, synthetic media and customer data become visible.

The European Union’s AI Act entered into force on 1 August 2024 and is scheduled to become fully applicable on 2 August 2026, with certain rules already in application. The European Commission states that transparency-related instruments include guidance and a voluntary code intended to support obligations around marking AI-generated content and disclosing the artificial nature of images, audio, including deepfakes, and text where relevant.

The exact legal implications depend on the use case, the system, the sector and the jurisdiction. Marketing leaders should not attempt to replace legal advice with a prompt. They should recognise that many marketing decisions are no longer purely creative or commercial decisions. They can involve data protection, consumer protection, discrimination, transparency and platform policy.

The European Data Protection Board has addressed data protection questions involving AI models, while the ICO guidance highlights accountability, transparency, lawfulness, statistical accuracy, fairness, bias, security and data minimisation.

AI governance becomes a marketing issue the moment an automated system influences what a person sees, receives, pays for or is told about themselves.

This should change operating habits. Campaign briefs should state whether AI is used, which data sources are involved, what claims need verification, whether any synthetic media may require disclosure, which decisions are automated and who can stop the system. The goal is not paperwork for its own sake. The goal is to make responsibility visible before a campaign reaches the public.

Procurement and vendor concentration create strategic exposure

AI dependence is also a supplier-risk issue. A marketing team may use a foundation-model provider, an advertising platform, a CRM vendor, a customer-service platform, a creative suite, a data-clean-room provider and several specialist tools. Each may promise integration. Together, they can create a dependency structure that is difficult to unwind.

The exposure is not only financial. A provider can change pricing, model behaviour, output restrictions, privacy terms, API access, data-retention policy or product roadmap. A platform can alter how it measures conversions or how it applies automated campaign recommendations. A model update can change the tone of generated content or the reliability of classification. A tool that worked well in a pilot can become unsuitable when scale introduces new data, security or legal requirements.

A marketing operation should never discover its vendor dependency during a campaign crisis.

Procurement needs to ask different questions than it asked in the pre-generative-AI era. Can the company export its data and prompts? Can it retrieve logs? Does it know where inputs are processed? Can it disable features by market or team? What happens if a model is updated? Can it test changes before production? Which employees can connect internal data sources? Can the company switch to another provider without rebuilding its entire workflow?

Vendor concentration also affects creative independence. If a brand relies on the same platform tools as every competitor, its media and content decisions may become increasingly shaped by the platform’s incentives. Platforms are indispensable distribution partners. They are not neutral custodians of a brand’s strategy.

The answer is not to avoid major platforms. It is to preserve options. Keep core customer knowledge in systems the company can govern. Maintain documented processes that can be run with more than one tool. Keep human expertise strong enough to identify when vendor outputs are wrong. Build brand assets and first-party relationships that do not disappear when a platform changes a setting.

Skills and team design matter more than tool access

AI changes the skills marketing teams need. It does not reduce the need for expertise. It changes where expertise matters most.

A junior marketer may be able to produce polished copy faster than before. That does not make them ready to decide a brand’s positioning, assess a legal claim, interpret a market signal or diagnose a customer problem. A senior marketer may be tempted to delegate routine work more aggressively. That does not remove their responsibility for the quality of the resulting system.

The skills that gain importance are often the ones that make AI use safer and more valuable: problem definition, source evaluation, research design, commercial judgment, creative direction, data literacy, experimentation, editing, risk assessment and the ability to explain a decision to someone outside the team.

Research on AI use is beginning to show the same pattern. One 2025 study of job postings from US public firms found that roles explicitly relying on generative AI had higher cognitive-skill requirements and an increase in demand for social skills after the release of ChatGPT. The evidence is still developing, but it supports a reasonable interpretation: AI does not simply substitute for thinking. It raises the value of people who can frame, evaluate and integrate complex work.

The useful marketer is not the person who knows the most prompts. It is the person who knows when the prompt is asking the wrong question.

This means training should be role-specific. A brand manager needs different AI skills from a paid-media specialist, a content editor, a customer-experience leader or a marketing analyst. Generic “AI literacy” sessions can create awareness. They do not create operational competence.

Teams should learn through real work. They should compare AI-assisted and human-only processes, inspect errors, document good prompts, define review standards and discuss difficult cases. The goal is not to create a culture of fear around AI. It is to create a culture where people are comfortable saying that an output is plausible but insufficient.

Workflows must change, not merely prompts

The difference between weak and strong AI adoption often appears in the workflow around the model.

A weak workflow begins with a broad instruction: write the campaign, create the audience, optimise the budget, answer the customers, summarise the market. The output is then accepted because it looks useful. There may be a final review, but the review is often rushed because the tool has already created a sense of completion.

A strong workflow begins earlier. It defines the decision, the acceptable evidence, the risk level, the input sources, the output format, the approval owner and the measurement plan. The model is used inside that structure, not instead of it.

Control points that keep AI use reversible

The table describes a management discipline, not a bureaucracy. A small team can apply it without creating a committee for every draft. The level of control should match the level of consequence. An internal brainstorming prompt needs little formal process. A campaign that uses customer data, public claims and automated messaging needs far more.

A good AI workflow makes it easy to move quickly on low-risk work and difficult to move carelessly on high-risk work.

This is also why prompt libraries alone are not enough. A prompt can help a team produce a better first draft. It cannot define ownership, prevent inappropriate data use or prove that a claim is correct. Process design creates those safeguards.

Evaluation discipline is the real bottleneck

Most organisations do not lack AI tools. They lack a reliable way to evaluate AI outputs.

The evaluation problem appears at several levels. Is the factual answer correct? Is the classification accurate enough for the intended use? Does the creative fit the brand? Does the recommendation improve the commercial outcome? Does the system treat customer groups fairly? Can the output be explained? Can a human identify the source of an error?

These questions require different forms of evaluation. A language model can be tested against approved product documentation. A lead-scoring model can be tested against later sales outcomes. A campaign-generation tool can be tested through controlled experiments. A customer-service assistant can be tested against known queries and escalation cases. A creative system can be reviewed by people who understand the brand and the category.

The mistake is to use a single generic score. A model may appear accurate on simple tasks and fail on rare but important cases. It may perform well in English and poorly in a local language. It may produce useful drafts for routine products but unreliable claims for technical or regulated products. It may look fair at an aggregate level while producing worse outcomes for specific groups.

McKinsey found that 51 percent of respondents from organisations using AI reported at least one negative consequence, with AI inaccuracy among the most commonly reported issues. It also found that organisations seeing stronger results were more likely to define processes for when outputs require human validation.

The most mature organisations do not assume that a good demo is evidence of reliable production performance.

Marketing leaders should establish a simple review architecture. Low-risk drafting can rely on editor review. Product claims should be checked against approved references. Campaign automation should be monitored against pre-defined performance and harm thresholds. Customer-facing agents should be sampled, audited and able to transfer to people. High-impact decisions should have a named accountable owner.

Evaluation is not glamorous. It is where much of the real value is created.

The case for human judgment is commercial, not sentimental

Arguments for human oversight are sometimes framed as if they are nostalgic. They are not. Human judgment matters because customers are not abstract data points, markets are not stable systems and brand value cannot be reduced to the sum of automated interactions.

A human can notice when a message is technically relevant but emotionally wrong. A human can recognise when a discount strategy is damaging long-term value. A human can see when a campaign is culturally insensitive before the data shows a problem. A human can decide that a customer should receive an exception even when the automated rule says no. A human can choose a creative direction because it is original, risky and worth defending.

AI may help with each of those tasks. It does not carry the responsibility for the decision.

The commercial importance of judgment grows as AI makes standard execution cheaper. When every competitor can access similar tools, the durable advantage comes from choices that are difficult to copy: a clearer product truth, a better customer experience, a stronger point of view, a more credible promise, more useful evidence and a willingness to say no to tactics that may produce short-term numbers at the expense of trust.

Human judgment is not the part of marketing left over after automation. It is the part that determines whether automation serves the business or distorts it.

This is particularly relevant for senior leadership. Boards and executives may ask whether the marketing function is “using AI enough.” The better question is whether the company has identified where AI can create value without weakening customer trust, brand distinctiveness or decision quality. Adoption for its own sake is not a strategy.

The operating principle should be AI optional and judgment essential

Marketing should aim for AI optionality rather than AI rejection. Optionality means the organisation can use AI where it creates value, but it can still operate, explain decisions and protect customers when a system fails, a vendor changes terms or a human judgment call is needed.

That principle has several practical implications.

First, keep the core of brand strategy outside the model. A model can test language against a strategy. It should not invent the strategy from generic market patterns.

Second, keep critical facts in governed sources. Product information, legal claims, pricing rules, customer permissions and brand standards should exist in systems that people can inspect, update and approve. AI tools can retrieve from those sources, but they should not become the only place where truth is stored.

Third, keep human relationships in the loop. The most valuable marketing insight often comes from customers, sales teams, service teams, distributors, communities and people who encounter the product in context. AI can organise that evidence. It cannot replace the relationship that creates it.

Fourth, maintain manual competence. A team should know how to assess a campaign, interpret a result, write a brief and respond to a customer without depending on one interface. This is not inefficient redundancy. It is organisational resilience.

Fifth, make AI use visible where visibility supports trust. Customers do not need a technical essay about every model in the marketing stack. They do deserve clarity when they are interacting with an automated agent, when synthetic media materially affects what they see or when personal data is used in ways that shape important decisions.

The European Commission’s AI Act materials describe transparency obligations and the growing policy focus on identifying artificial content in relevant contexts. Marketing teams should treat this direction of travel as a design challenge, not merely a compliance burden.

Dependence becomes a management failure when no one owns the boundaries

Marketing is becoming dependent on AI because AI is becoming embedded in the infrastructure of marketing. It influences the media bought, the audiences reached, the content generated, the search experiences encountered, the leads scored, the customer questions answered and the metrics reported.

That dependency is not automatically harmful. It can make teams more responsive, more informed and more efficient. It can reduce repetitive work and create room for better research, better service and better creativity. It can help smaller organisations compete where production cost once blocked them.

The harm begins when the organisation stops deciding what AI is allowed to do.

A company that lets AI draft a first version of a campaign while requiring human evidence and approval is using AI well. A company that lets AI create, target, publish, optimise and report on marketing without clear ownership is building a system it may not be able to explain. A company that relies on AI to understand customers while reducing contact with real customers is trading knowledge for simulation. A company that floods channels with generated content while investing less in original experience is reducing the value of its own voice.

Marketing will not be defined by whether it adopts AI. It will be defined by whether it can remain accountable, distinctive and useful after AI becomes ordinary.

The strongest marketing organisations will not be the ones with the most tools. They will be the ones that can make disciplined choices about where automation belongs, where it does not, what evidence a claim requires, which customer signals should remain off limits, when a person must intervene and how a brand remains recognisable when every competitor can generate polished work in seconds.

AI is becoming a permanent part of marketing. It should become part of a stronger marketing system, not a substitute for one.

Questions marketers are asking about AI dependence

Author:
Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency

This article is an original analysis supported by the sources cited below

The State of AI Global Survey 2025
McKinsey’s global survey on AI use, agents, business impact, workflow redesign, governance and reported negative consequences.

Artificial Intelligence Index Report 2026
Stanford HAI’s annual overview of AI capabilities, adoption, economics, governance and wider societal effects.

AI Risk Management Framework
NIST’s voluntary framework for incorporating trustworthiness considerations into AI design, deployment and evaluation.

AI RMF Generative Artificial Intelligence Profile
NIST guidance focused on generative-AI-specific risks and actions that organisations can use to manage them.

AI Act
European Commission information on the EU AI Act, its implementation timeline and transparency-related support measures.

Opinion 28/2024 on certain data protection aspects related to AI models
European Data Protection Board opinion on personal-data processing questions in the context of AI models.

Guidance on AI and data protection
UK Information Commissioner’s Office guidance on accountability, transparency, fairness, bias, accuracy, security and data minimisation in AI.

AI features and your website
Google Search Central documentation on AI Overviews, AI Mode, site eligibility and the continued relevance of standard SEO practices.

Guidance on using generative AI on your website
Google guidance on responsible use of generative AI for web content and the importance of maintaining quality and usefulness.

Creating helpful, reliable, people-first content
Google’s guidance on content quality, expertise, trust and user value.

Keep your AI claims in check
US Federal Trade Commission guidance warning businesses against exaggerated or deceptive claims about AI capabilities.

Operation AI Comply
FTC enforcement action addressing alleged deceptive uses and claims involving AI products and services.

2025 Digital video ad spend and strategy report
Interactive Advertising Bureau research on generative AI use in digital video advertising and changes in advertiser demand.

AI and digital trends 2026
Adobe research examining AI adoption, data readiness, customer expectations and agentic experiences.

The 2025 AI Agent Index
Research documenting technical, safety and transparency features of deployed agentic AI systems.

Anthropic Economic Index report
Research on how people and businesses use AI systems, including changing levels of task delegation.

Generative AI adoption and higher order skills
Research examining how generative-AI adoption affects demand for cognitive and social skills in work roles.

International AI Safety Report
International assessment of general-purpose AI capabilities, risks and risk-management approaches.

C2PA technical specification
Open technical specification for cryptographically recording provenance information in digital media.

Verifying provenance of digital media
Independent analysis of limitations and security concerns in current C2PA provenance specifications.

Global guidance on generative AI
World Federation of Advertisers guidance on responsible adoption of generative AI in marketing and advertising.

AI and advertising policy research
Research on the use of large language models in advertiser-content understanding and policy enforcement.

Responsible AI in marketing
OECD analysis of AI’s broader social, economic and governance implications relevant to organisations deploying AI systems.

The State of Content Authenticity
Industry resources on content provenance, transparency and the adoption of Content Credentials.