Marketing has spent the past few years discovering that machines can produce words, images and campaign variations at extraordinary speed. The harder discovery is that production was never the only constraint.
A communication still has to be accurate, a claim still needs evidence and a message must remain appropriate for its audience, consistent with the organisation's position and acceptable within the rules governing its use. Someone must also decide whether the work is routine, whether it presents a material risk and whether a specialist needs to intervene. Generative AI has made production dramatically easier, but it has done much less to resolve the decisions that surround it.
This distinction sits behind the most revealing finding in BCG's 2026 global survey of 300 chief marketing officers. Although 96% of CMOs described their function as undergoing significant end-to-end AI transformation, 42% acknowledged that they still use generative AI only as an assistant for individual tasks in a small number of workflows.
Both figures can be true because an organisation can use AI extensively without changing how it makes decisions. It can generate first drafts faster or produce dozens of campaign variations while leaving approval untouched and relying on the same people to check each one manually. Equally, it can introduce copilots throughout the business without giving them access to approved claims, current policies, previous decisions or the organisation's interpretation of its obligations.
In each case, the technology appears transformed while the operating model around it remains much the same. For regulated industries, that gap matters more than the number of tools in use. Ultimately, the future of communication will depend not simply on what an AI system can create, but on whether the organisation can explain why the resulting communication should be trusted.
More content can create less control
The language surrounding agentic marketing can make the future sound almost entirely autonomous: campaigns will plan themselves, creative will adapt in real time and agents will optimise customer journeys with limited human involvement. BCG's evidence points to a more gradual reality. Only around 8% of the CMOs surveyed are beginning to connect multiple agents to run certain kinds of campaign autonomously.
The leading organisations are not simply removing people; they are redesigning workflows so that agents and people perform different parts of the work, while retaining human oversight where it matters. That distinction becomes especially important when communication is regulated.
A model may recognise that a sentence resembles previously approved wording, but it cannot assume that the wording remains appropriate for a different product, audience or channel. A claim accepted in a detailed professional document may be misleading in a short social advertisement, just as a disclosure may be technically present but practically invisible. Even a personalised message that is factually correct may create an inappropriate impression for a vulnerable recipient.
Regulated communication is therefore not a binary exercise in matching content against a rulebook. It requires a sequence of contextual decisions about what is being claimed, which evidence supports it, who will receive it and what they are likely to understand. Teams must also establish which requirements apply, whether the organisation has encountered a comparable situation before and where established guidance gives way to fresh judgement.
In many organisations, the answers are distributed across policies, inboxes, document repositories and the memories of experienced people. That arrangement can survive while communication volumes remain relatively stable, but it becomes fragile as AI allows production to multiply.
If a marketing team moves from producing ten assets to producing one hundred, a review process designed for ten does not become ten times more effective; it becomes a queue. The organisation can respond by reviewing less, adding more people or allowing campaigns to wait, but none offers a durable answer. The more meaningful measure is not simply the speed at which content can be generated, but the speed at which the organisation can reach a defensible decision about it.
Compliance cannot remain the final stage
BCG reports substantial improvements from its work with leading CMOs: cost efficiencies of 20% to 30%, a threefold increase in marketing return on investment and a tenfold improvement in campaign cycle times. Although these are observations from BCG engagements rather than results from the full survey, the combination is instructive because the value comes not from generation alone, but from redesigning the entire route between an idea and an outcome.
For regulated organisations, this means moving checks closer to the point at which a communication is conceived. Requirements should inform the brief, approved evidence should be available while claims are being written and familiar issues should be identified before they reach a specialist. Reviewers, in turn, should receive the context needed to make a decision rather than spending their time reconstructing it.
Compliance performed only at the end of production will struggle in an agentic environment. By then, the creative direction may be fixed, media may have been booked and commercial expectations may have hardened. A late objection becomes expensive, making constructive challenge more difficult precisely when it is most necessary.
Bringing control forward is not merely safer; it is usually faster, and it changes the relationship between marketing and compliance. Instead of one function creating and another rejecting, both can work from the same evidence and requirements, resolving problems while ideas are still flexible and allowing specialists to concentrate on material questions rather than avoidable omissions. The objective is not to subject every draft to more scrutiny, but to apply the right scrutiny sooner.
The missing layer is organisational intelligence
BCG describes the emerging agentic marketing technology stack as having four parts: a data layer, a brand intelligence layer, an agentic layer and a unified interface for marketers. The brand intelligence layer is particularly significant. BCG defines it as the part of the system that encodes an organisation's rules, operating context, messaging guardrails, trusted sources and performance measures, turning a general-purpose copilot into something capable of acting in a company-specific way. In regulated industries, however, this layer must understand more than the brand.
It needs to connect external requirements with the organisation's interpretation of them, distinguishing between published regulation, internal policy, approved wording and previous judgement. It should recognise when the latest product documentation supersedes an older claim and when a question has moved beyond guidance that can be applied routinely into an issue requiring authorised review.
Provenance is equally important. Rather than simply producing an answer, a system should make its basis visible: which source it used, whether that source is current, which policy or precedent applies, what assumptions were made and who confirmed the final decision.
Without that structure, an agent may sound informed while operating on partial context, because fluency can disguise uncertainty remarkably well. With it, AI can retrieve relevant requirements, compare a new communication with previous decisions, identify missing substantiation and route an exception to the right person. It can help an organisation apply what it already knows without pretending that every situation has already been decided. The model remains important, but the surrounding institutional intelligence determines whether it becomes genuinely useful.
Governance should direct attention, not simply add control
BCG found that the risks occupying CMOs are now highly practical. Some 65% cited implementation challenges, with the same proportion concerned about data privacy and security. Copyright and legal issues, and protecting brand voice and reputation, were each identified by 64%. Regulatory compliance concerned 62%, while 61% highlighted inaccurate content.
These concerns are not arguments for slowing adoption; they are design requirements. Good governance should determine where human attention creates the most value, allowing familiar, lower-risk work to move efficiently while directing specialist judgement towards novel claims, unusual audiences, significant product changes and communications with greater potential for harm.
Treating every item identically can look rigorous, but it often produces the opposite effect. When everything is urgent, genuinely important work competes with routine checking. Experienced reviewers spend time finding documents and repeating settled explanations. Teams begin to regard compliance as an unpredictable delay rather than a source of better decisions.
A more mature model separates three kinds of work. Some decisions can be automated because the requirement and permitted response are unambiguous. Others can be prepared by a system but still require a person to confirm the outcome. A smaller group involves genuinely new judgement and should be escalated with the relevant evidence and context already assembled. The purpose is not to automate accountability, but to prevent scarce judgement from being consumed by tasks that do not require it.
"Human in the loop" is often used as a reassuring response to this problem, yet the phrase means little without explicit design. Organisations need to decide which actions an agent can take, which recommendations it can make and what requires approval. They also need to define what information a reviewer must see, what happens when sources conflict, who can override a recommendation and how that intervention is recorded.
Meaningful oversight depends on answering those questions before a system enters live work. Machines are well suited to retrieval, comparison, monitoring and the consistent application of defined criteria, while people remain responsible for ambiguity, proportionality, commercial context and accountability. The boundary between them should be designed deliberately rather than discovered after something goes wrong.
The future belongs to reusable judgement
Most organisations have more institutional knowledge than their systems reveal. They have answered difficult questions, interpreted ambiguous requirements, agreed acceptable formulations and developed a practical understanding of their own risk appetite, yet much of that knowledge remains trapped inside completed reviews.
The final asset may be stored while the reasoning behind it is lost, and a comment may be resolved while the reason for the change remains in an email thread. When an experienced colleague leaves, years of contextual understanding may leave with them. As a result, each new communication becomes unnecessarily expensive: teams repeatedly research the same issue, compliance reconstructs the same argument and similar cases receive different answers depending on who happens to review them.
Agentic systems create an opportunity to change this, but only if organisations preserve more than final approvals. A useful precedent should record what was decided, why it was decided, the evidence available at the time and the conditions that made the decision appropriate. It should also make it possible to recognise when those conditions have changed.
This is how judgement becomes reusable without becoming indiscriminate. Reuse does not mean turning every previous decision into a permanent rule; it means giving future decision-makers a stronger starting point. They can see how the organisation approached a comparable question, decide whether the reasoning still applies and focus their attention on what is genuinely different.
Over time, this produces consistency without requiring central control over every word. It also makes the organisation less dependent on informal access to a small number of experts.
The results should also be measured across the whole decision process. Leaders should ask how much work is returned because the brief lacked essential information, how often comparable claims receive inconsistent treatment, what proportion of reviews involve genuinely novel issues and where decisions tend to wait. They should also be able to reconstruct the evidence behind a live communication. Together, these questions reveal whether technology is reducing low-value effort or merely moving it elsewhere.
Transformation is organisational before it is autonomous
BCG divides CMOs into three maturity groups: 32% are leaders, 26% are followers and 42% are considered at risk of remaining at task-level assistance. The division is based not on ambition or expenditure, but on how deeply AI has changed workflows, technology, talent and results.
This offers a useful corrective to much of the discussion about agentic marketing. The organisations pulling ahead are not distinguished simply by access to better models; they are doing the less visible work of connecting data, encoding organisational context, establishing governance and redesigning processes from beginning to end. In regulated industries, the same work will determine whether AI increases confidence or simply increases volume.
The future of communication is therefore unlikely to be a choice between complete automation and traditional human review. It will be a system in which routine work becomes increasingly machine-assisted, complex judgement remains accountable to people and the organisation's accumulated knowledge connects the two.
As content becomes easier to make, expertise does not become less valuable. Instead, the ability to apply it consistently, at the moment it is needed, becomes the most important part of the system.
What teams need to know
What is agentic marketing?
Agentic marketing uses AI agents to perform or coordinate parts of marketing workflows, potentially spanning planning, content creation, activation, measurement and optimisation. Mature implementations combine agents with organisational context, defined controls and human oversight.
What is the missing layer in agentic marketing?
The missing layer is organisational intelligence: the approved rules, evidence, policies, previous decisions and operating context that allow AI systems and people to make dependable, company-specific decisions.
Why is agentic marketing challenging in regulated industries?
Whether a communication is acceptable depends on its claims, evidence, audience, channel and context. Faster content production can overwhelm review unless regulatory knowledge and appropriate escalation are built into the workflow.
Does agentic marketing remove the need for human review?
No. Routine retrieval, comparison and defined checks may be automated, but people remain responsible for ambiguity, proportionality, commercial context and accountability. Effective oversight requires clear decision rights, evidence and escalation routes.
How does organisational intelligence differ from brand intelligence?
Brand intelligence grounds agents in an organisation's voice, messaging rules, trusted sources and commercial context. Organisational intelligence extends this foundation to include regulatory requirements, internal policies, approved evidence, previous decisions and escalation routes. This broader context is particularly important when communications operate under regulatory or professional obligations.