Large consumer brands in media, financial services, retail and marketplaces are moving artificial intelligence from experimentation into the centre of their marketing and digital operations. Most are beyond pilots, yet few have built the integrated and governed systems that will matter over the next one to two years.
Across sectors, leaders responsible for digital platforms, data and customer growth are asking similar questions. They want to understand what their peers are building, which architectural patterns are emerging, and how to balance speed with risk during the next twelve months.
This paper synthesises active strategy and delivery work across media, financial services, marketplaces and specialty retail. It focuses on three concerns:
- What systems organisations are trying to build, beyond isolated use cases.
- How agentic workflows and an intelligence layer change the architecture of marketing.
- What a realistic twelve month roadmap looks like when strategy, technology and operating model are aligned.
1. The shared problem for digital and marketing
Leaders responsible for digital, data and platforms sit at one end of the challenge. Leaders responsible for brand, customers and commercial outcomes sit at the other. They are now dealing with a shared system rather than separate domains.
Customer expectations for relevance and coherence are rising while budgets and headcount stay tight. There is pressure to deliver more personalised and consistent journeys, to prove effectiveness and to protect the brand.
On the technology side there is pressure to rationalise platforms, protect information assets and manage risk, while still enabling faster change. Artificial intelligence is seen both as an opportunity to recast how digital channels and data platforms create value and as a new source of complexity that must be controlled.
Treating AI as a side experiment is no longer tenable. The core question has shifted from whether to act, to what system is being built, how that system will be governed and how quickly it can produce results without introducing new fragility.
2. What leading organisations are building
Across media networks, banks, wealth managers, marketplaces and specialty retailers, similar solution patterns are emerging, even when labels and vendors differ. The common direction is away from channel specific tools toward a coherent intelligence layer that sits across campaigns, content and customer interactions.
2.1 From campaign tooling to an intelligence layer
Most large organisations still run marketing through a familiar stack: campaign managers, advertising platforms, email and messaging tools, content management, a customer data platform or warehouse and a patchwork of analytics.
The strongest programmes do not attempt to replace this stack wholesale. Instead they introduce an intelligence layer above and between existing systems.
This layer reads by consuming data about customers, content, campaigns and outcomes from existing sources. In media this might include identity, on platform behaviour, inventory and product catalogues. In banking it includes transactional and engagement data, product holdings, risk and eligibility rules. In retail and marketplaces it includes browsing, purchase and returns behaviour, product taxonomy and pricing.
It decides by using combinations of rules, models and agents to propose actions - how to structure a campaign, which segments to target, which content variants to use, how to allocate budget and which journeys to trigger.
It writes by translating those decisions into concrete artefacts: briefs, media plans, content drafts, platform configurations and monitoring alerts.
Today this intelligence layer is still at an early stage in many organisations. It may appear as a small number of services and agents that sit alongside the legacy stack, integrated using workflow tooling and application interfaces. The strategic intent, however, is clear. Over the next twelve months a growing share of marketing work is expected to be initiated, structured or checked by this layer rather than handled only through manual processes.
2.2 Agentic workflows instead of linear steps
Campaign processes have historically been described as linear sequences: brief, design, produce, launch, optimise, report. In practice, work bounces between teams, tools and documents.
Agentic approaches formalise that reality. Organisations are starting to model key workflows as graphs of agents and humans, each with explicit responsibilities.
A media organisation might implement the following sequence for on platform campaigns. A strategy agent parses the business objective and historic performance data, identifies relevant audiences and placements and drafts a structured brief. A planning agent proposes a campaign structure with hypotheses, target segments, channels and initial budget allocations. A creative agent generates candidate headlines, promos or visual concepts aligned to brand guidelines and past performance, labelling each with intended segment, tone and placement. An execution agent translates approved plans into platform configurations, checks constraints such as budgets and frequency caps and submits changes through existing advertising platforms and internal tools. A monitoring agent tracks results, compares them against control groups and expectations, flags anomalies or policy breaches and suggests reallocations or tests.
Human practitioners stay in the loop at critical points, such as approving strategy and creative, confirming high impact changes and responding to alerts. Agents handle structured and repeatable tasks and record their actions and rationales in logs and reports.
Banks and wealth managers adopt similar patterns for lifecycle communication, with agents interpreting customer behaviour and product rules, proposing compliant communications and coordinating across email, applications and human channels. Retailers and marketplaces apply agentic patterns to merchandising, offer management and content personalisation.
2.3 Software and architecture choices
Underneath these workflows three groups of technology decisions dominate.
- Access to foundation models. Organisations are standardising around one or more primary sources of foundation models. Typically they consume text and image models through large cloud providers, sometimes supplemented by specialised third party models for recommendation or vision tasks and occasional fine tuned internal models for sensitive domains such as risk communication. The strategic choice is less about selecting a single model and more about choosing providers and abstractions that support flexibility, governance and cost control.
- Orchestration and workflow. To make agents practical, organisations are selecting or building orchestration layers that allow them to define, run and monitor agentic workflows. In some cases this is achieved through low code workflow platforms extended with custom actions for model calls and data access. In others it is implemented using application frameworks and message queues. The common requirement is the ability to coordinate multiple agents, tools and human steps with clear observability and control.
- Data access and feature delivery. Rather than building entirely new data platforms, most organisations are exposing governed slices of existing customer, content and campaign data to agents. This often involves combining a warehouse or lakehouse with a feature store or vector search service and standardised views for common queries such as recent customer interactions, similar content and its performance or eligible products and offers. The key decision is how to expose this data safely and consistently without fragmenting governance.
Around these choices sit services for experimentation, analytics and governance. Existing experimentation platforms are being extended to handle AI driven tests and variants. Monitoring tools that can inspect prompts, responses and decisions in flight are becoming essential, particularly in regulated sectors.
Exhibit 1. Agentic architecture
The enterprise model for agentic design can be understood as four connected layers. At the base, an AI enabled operating model provides autonomous cross functional teams, an AI control tower, agile delivery practices, deliberate talent choices and a focus on value capture. Above this sits the data and core technology layer. This includes governed enterprise data, data integration and transformation, vector and structured stores, search and retrieval, application interfaces and secure infrastructure. On top of that is the AI and machine learning platform layer. It combines industrialised model operations, observability and safety tooling, orchestration services and gateways that expose foundation models in a controlled way. The intelligence layer sits above. Here predictive models and networks of agents operate on specific decision domains such as acquisition, credit, monitoring and servicing. These agents are orchestrated in workflows that support campaign design, content production, lifecycle journeys and optimisation. The experience layer is where customers and internal users meet this capability through multi modal conversational experiences, intelligent products and omni channel journeys built around stakeholder digital twins.
3. Agentic workflows in practice
For leaders planning the next year it is helpful to make these concepts concrete. Three canonical workflows show how agentic patterns are being applied in marketing and customer engagement.
3.1 Campaign design and briefing
Turning a business objective into a campaign plan remains heavily manual in many organisations. Teams review prior activity, interrogate data, write briefs and align stakeholders in meetings and documents.
In an agentic design a strategy agent ingests the business objective, previous campaign performance, audience insights and constraints. It produces a first pass structured brief that sets out target segments, hypotheses, recommended channels, benchmarks and key questions to test. The marketing team edits and approves this brief instead of starting from a blank page.
A research agent can ingest unstructured material such as prior reports, customer feedback or competitor activity and return focused summaries. A pricing or yield agent might propose guardrails for spend based on inventory or budget limits.
The result is faster and more consistent strategy formation, with better reuse of existing knowledge. For this to work agents need access to well defined data surfaces and their outputs need to be captured in systems, not scattered across slides and documents.
3.2 Content and asset production
In content heavy environments creative teams are often overloaded with requests for variants across segments, placements and formats.
Agentic production introduces content agents configured with brand, legal and accessibility guidelines and informed by metadata about past asset performance. These agents generate first drafts of copy, creative concepts and layouts for specific use cases, labelling outputs with intended audience, tone and risk level. Designers and copywriters remain responsible for refinement and final approval, but start from structured proposals rather than blank screens.
Over time, performance feedback is fed back into the content agents. When particular visual patterns or phrasings perform better for a given audience and context the agent learns to prioritise those patterns. Creative exploration becomes both broader and more directed.
To support this, organisations require a digital asset and design system able to store structured content and metadata, along with interfaces that allow agents to read and write assets and labels. They also need clear decisions about which degrees of variation are acceptable and which must always be touched by humans.
3.3 Lifecycle communication and journeys
In banking, wealth and many subscription businesses, a large share of marketing effectiveness comes from lifecycle journeys such as onboarding, usage growth, cross sell and retention.
Agentic journeys use agents to interpret customer signals and propose next best actions that comply with product and risk rules.
A journey agent monitors behavioural and transactional events, eligibility constraints and current journey participation. When a trigger fires the agent proposes a communication or action, including channel, timing and message variants. A compliance agent checks the proposal against policy and regulatory rules. A routing agent schedules and executes the communication and records outcome data. Human teams can intervene for high value or high risk cases.
In this pattern, platform decisions around event collection, identity resolution and rules engines become central. Decisions about customer promises, tone and acceptable levels of automation are encoded as constraints that agents must respect.
Exhibit 2. Campaign Workflow Example: Agent-Led, Enterprise-Grade Agentic marketing architecture
4. Architecture patterns that matter
Across organisations four architectural choices are becoming common.
First, separation between the intelligence layer and underlying systems. Agents access models, data and tools through well defined interfaces rather than embedding logic inside each channel platform. This makes it possible to evolve channels without rebuilding the intelligence layer each time.
Second, shared components for common agent tasks. A single content agent service can support both campaign and lifecycle workflows, configured differently per domain. A single monitoring and policy service can scan outputs across email, web and in application surfaces. Scale economies appear when these shared services replace multiple bespoke implementations.
Third, a move toward event driven designs. Because agentic workflows depend on timely signals, organisations are investing in event streams or similar mechanisms to expose customer and campaign events in near real time while still respecting privacy and regulatory constraints.
Fourth, integrated observability. Logs of prompts, responses, decisions and human overrides are treated as first class data. They feed back into training, risk management and process improvement. This is a shift from traditional marketing reporting, which focuses mainly on outcomes rather than the decision paths that led to them.
These patterns imply that AI strategy cannot be treated as a series of channel specific projects. It is an architectural and operating model choice that cuts across digital, data and marketing.
5. A twelve month roadmap
Given these patterns it is possible to describe a pragmatic twelve month trajectory that is already visible in leading organisations.
Months 0 to 3: choose and model a small set of workflows
The first move is to select a small number of concrete workflows that matter commercially and can be owned by named teams. Typical choices are campaign design inside a single business unit, content production for a defined set of channels, or a specific lifecycle journey.
For each workflow, teams map current steps, identify key decisions and specify which could be supported or led by agents. They then design a minimal set of agents, define required data access, and integrate agents into the tools practitioners already use.
The objective at this stage is to create working agentic processes in live environments, with clear metrics on cycle time, practitioner effort and core performance indicators. Governance teams are engaged from the design stage so that constraints are embedded and data access is approved.
Months 3 to 6: connect workflows and consolidate platforms
The next move is to connect adjacent workflows. The output of a campaign design agent becomes structured input for content agents and activation agents, rather than being reinterpreted manually. Data integration focuses on what is necessary to support these links: shared identifiers, core behavioural and performance features, and content metadata.
At the same time, the organisation starts to consolidate the orchestration approach. Instead of each team building its own scripts or notebooks, a shared workflow and agent orchestration layer is promoted with standard patterns for calling models, logging activity and enforcing policies.
By the end of this phase at least one end to end lane should exist where a defined part of the business can move from objective to live activity through a sequence of agents and humans, with outputs and decisions recorded.
Months 6 to 12: platform extraction and scaling
In the second half of the year the focus shifts from individual lanes to reusable platform components.
Shared agents such as content suggestion, monitoring, compliance checking and experiment design are extracted from initial implementations and offered as services to other teams. Data access patterns that proved effective become formalised as governed interfaces or views. Experimentation and analytics practices are aligned so that impact from multiple workflows can be compared.
Organisationally, new roles such as AI product owners and marketing engineers are clarified. Incentives are adjusted to reward effective use and improvement of agentic workflows, not only maintenance of legacy processes.
By the twelve month point an organisation following this trajectory should be able to point to specific, measurable improvements and to describe a coherent architecture that can support further expansion rather than a collection of disconnected experiments.
6. Strategic questions for leadership
To make this real, senior digital and marketing leaders need to align on a small set of explicit decisions.
- Which one or two marketing decision areas are most suitable for AI enabled agents in the next six months, given existing data and tools. This might be a specific campaign type, a particular lifecycle journey or a content production stream.
- What foundation model access strategy provides enough flexibility and control without fragmenting across too many providers. This includes decisions about which cloud platforms to standardise on and how to abstract model access so that future changes remain possible.
- How agent responsibilities will be defined and governed. In particular, which decisions agents are allowed to make autonomously, which must always be reviewed, and which remain off limits. These boundaries need to be explicit and reflected in monitoring.
- Which data surfaces will be exposed to agents and under what governance conditions. Privacy, consent and regulatory constraints must be designed into data access rather than checked only at the end.
- Which orchestration and workflow tooling will become standard so that teams do not fork the architecture with incompatible approaches. Agreement here determines whether agents can be reused and scaled.
- How experimentation and impact measurement will be handled so that AI is judged on incremental value, not only activity. This includes decisions about control groups, metrics and reporting rhythms.
- Who will own the evolution of the intelligence layer over time. Ownership can sit in digital, in marketing, in a central data function or in a joint construct, but decision rights need to be clear.
Explicit answers to these questions are more important than elaborate long range visions. They determine whether AI remains a tactical feature inside a handful of tools or becomes a disciplined way of making and improving marketing decisions.
Conclusion
The state of AI in marketing among large consumer facing organisations is neither early nor settled. Most have already demonstrated that models can draft copy, generate images and make recommendations. The frontier has shifted to questions of how to structure decision making, how to architect the intelligence layer and how to embed agents alongside humans in ways that improve speed, quality and control.
Organisations that move ahead during the next twelve months will treat AI as a shared strategy and execution problem across digital, data and marketing. They will design a small number of agentic workflows end to end and invest in the software and data foundations that allow those workflows to be replicated and extended.
From that point the path is iterative rather than speculative. Each new workflow, each new agent and each new experiment becomes part of a shared system that learns. That system, more than any individual model, will be the source of durable advantage.
