Agentic Marketing
Part II · Process Architecture
How agents and humans
work together in practice
work together in practice
A definitive process architecture for Agentic Marketing · documenting the interaction model, upstream and downstream dependencies, decision gates, and the new operational roles that make intelligent automation safe, scalable, and audit-ready in regulated financial services.
23
Defined process touchpoints across the campaign lifecycle
8
Escalation triggers with named response paths
4
Actor layers in the human-agent interaction model
5
Maturity stages from manual to fully agentic
The new paradigm
What Agentic Marketing actually means operationally
Agentic Marketing is not automation with better marketing. It is a fundamentally different operating model where AI agents hold ongoing responsibilities · perceiving, reasoning, acting, and learning · while humans define strategy, set boundaries, review decisions, and govern outcomes. The distinction matters because the operational, organisational, and governance implications are entirely different from anything that came before.
Architecture
The five-layer agentic marketing stack
The stack defines the hierarchy of the system · what sits above and below what, and where each agent operates within it. Understanding this architecture is the prerequisite for understanding every process flow that follows.
The critical insight: The agents (Layer 3) do not replace the decisioning core (Layer 2) · they reason around it, over it, and across it. Tibco makes the product-level decision. The agents make the customer-level, cross-portfolio, and cross-channel decisions that Tibco cannot make alone. The human governance layer operates across all five layers simultaneously · defining what is permitted, monitoring what occurs, and intervening when required.
Interaction Model
How agents and humans collaborate across the campaign lifecycle
Human-Agent Interaction Model
Five phases · Four actors · Every handoff mapped
The definitive process architecture for Agentic Marketing. Click any cell to see the full detail for that interaction · what triggers it, what information flows, what decision is made, and what happens next. This is the operational blueprint for how the new ways of working function in practice.
Actor key
Marketing & Business (human)
Agent layer (AI)
Technology systems
Risk & Compliance (human)
Escalation trigger
✕
Inputs received
Outputs produced
Actor
01 · Discover & Brief
02 · Design & Model
03 · Configure & Deploy
04 · Execute & Monitor
05 · Learn & Optimise
Marketing
& Business
& Business
Define campaign objectives
↓
Create campaign brief & segment
↓
Submit for data gap review
Review A1 data gap report
↓
Approve channel mix (A3 input)
↓
Sign off audience logic
Final configuration review
↓
Approve audience size
↓
Sign deployment authorisation
Monitor A5 health dashboard
↓
Review daily performance
↓
Act on escalation alerts
Review post-campaign outcomes
↓
Approve model recalibration
↓
Update strategic brief
Agent
Layer (A1–A5)
Layer (A1–A5)
A1: Analyse data coverage vs brief
↓
A1: Surface missing decision signals
↓
A5: Set segment health baseline
A4: Check cross-portfolio conflicts
↓
A3: Propose optimal channel mix
↓
A1: Flag field gaps for this segment
A2: Pre-flight policy check
↓
A3: Configure channel sequencing
↓
A5: Set monitoring thresholds
A2: Real-time policy enforcement
↓
A3: Dynamic channel routing
↓
A4: Cross-portfolio conflict detect
↓
A5: Continuous health monitoring
A1: Update field priority scores
↓
A3: Recalibrate channel model
↓
A5: Generate post-campaign report
Technology
Systems
Systems
Campaign system: store brief
↓
DW: provide field catalogue to A1
↓
History: extract performance baseline
Tibco: build decisioning rules
↓
Data eng: map approved new fields
↓
Campaign system: audience extract
Tibco: activate campaign
↓
Event bus: prime for campaign
↓
Audit log: capture confirmed
Tibco: execute NBA decisioning
↓
Event bus: capture all contacts
↓
Orchestration: route to channels
DW: write outcome data
↓
Model store: apply agent updates
↓
Governance dashboard: refresh KPIs
Risk &
Compliance
Compliance
Review segment eligibility
↓
Confirm consent status
Validate audience vs regulations
↓
Review A4 conflict report
↓
Approve communication strategy
Confirm audit log active
↓
Verify A2 policy configuration
↓
Issue final deployment clearance
Monitor policy compliance rate
↓
Review all escalation alerts
↓
Respond within defined SLA
Review governance report
↓
Assess override patterns
↓
Countersign model recalibration
How to read this model: Vertical flows within each column show what happens within a phase, in sequence. Horizontal flows across columns show the campaign progressing through time. The critical handoffs · where one actor's output becomes another actor's input · occur between rows within the same phase column. The click-through detail for each cell shows exactly what information flows across those handoffs.
Process Architecture
Campaign lifecycle, policy enforcement, and cross-portfolio flows in detail
Process flow
Campaign lifecycle · before and after
The most important process comparison in Agentic Marketing: the same campaign lifecycle, with and without the agentic operating model. What changes is not the campaign itself · it is who does what, when, and how decisions are made.
Current state · manual, fragmented, reactive
1
Marketing
Campaign brief created manually
No data gap analysis. No channel performance data. Campaign designed based on experience and assumption.
↓
2
Data / Technology
Ad hoc data request to engineering
No prioritisation framework. Marketing requests fields; engineering decides when to deliver them. Average wait: 2–6 weeks.
↓
3
Marketing
Audience built, campaign configured in Tibco
Credit card only. No cross-portfolio view. Customers in mortgage retention may receive credit card upsell simultaneously.
↓
4
Compliance
Manual contact policy check (batch)
Previous day's count referenced. Concurrent jobs each pass the check independently. Overcommunication possible and undetected.
↓
5
Marketing
Campaign deployed · DM and telemarketing only
SMS and email managed separately. No unified view. Contact history reconciled manually by batch file. Real-time decisioning impossible.
↓
6
Analytics
Post-campaign report produced manually
3–4 weeks after campaign ends. No root cause attribution. Learning is qualitative and informal. Models don't automatically improve.
Agentic state · intelligent, unified, self-improving
1
Marketing
Campaign brief submitted · agent analysis begins immediately
A1 analyses data coverage against this brief's requirements. A5 establishes segment baseline. Both return reports before the design phase begins.
↓
2
Agent Layer
A4 validates audience · A3 proposes channel mix
Cross-portfolio conflict analysis eliminates customers in conflict states. Channel recommendation includes expected response rates per channel per segment. Marketing reviews and approves.
↓
3
Agent Layer
A2 runs pre-flight policy check across full audience
Every customer checked against rolling 7-day contact window before deployment is authorised. Zero concurrent-job loophole. Flagged customers excluded automatically.
↓
4
Risk & Compliance
Final clearance · audit log confirmed active
Human compliance sign-off. Audit log capture verified. Policy Guardian confirmed operational. Deployment authorised by named individual.
↓
5
Agent Layer
Live execution · A2, A3, A4, A5 operational simultaneously
Every dispatch request checked by A2 in real time. Every customer routed by A3 to optimal channel. A4 resolving cross-portfolio conflicts as they emerge. A5 monitoring and alerting continuously.
↓
6
Agent Layer
Real-time learning · models improve during and after campaign
A3 improves channel predictions with each response event. A1 reprioritises fields based on what proved predictive. A5 post-campaign report with root cause attribution produced within 48 hours of campaign close.
Process flow · contact policy
Real-time contact policy enforcement
The detailed process architecture for how the Policy Guardian agent intercepts, evaluates, and decides every contact dispatch request across all channels and all platforms · in real time, with zero race condition risk.
End-to-end enforcement flow
①
Any platform (Tibco / Email / SMS)
Dispatch request raised
Before any communication can be sent to a customer on any channel, the originating platform must call the Policy Guardian API with the customer token, channel, and timestamp.
↓
②
A2 · Policy Guardian
Acquire atomic lock on customer record
The lock prevents any concurrent request for the same customer from reading an intermediate state. Eliminates the race condition that created the batch-mode loophole. Lock duration: typically 20–50ms.
↓
③
A2 · Policy Guardian
Read live contact count · rolling window
Query the live contact ledger: how many contacts has this customer received in the last 7 days (rolling, not "yesterday")? Includes contacts across all channels and all platforms.
↓
④
A2 · Policy Guardian
Evaluate all applicable policy rules
Check rolling weekly limit. Check channel-specific daily limit. Check message-type frequency (e.g., promotional vs transactional). Check opt-out flags and consent status. All rules evaluated atomically before releasing the lock.
↓
All rules pass
✓
A2
Approve + increment counter
Increment contact count atomically. Release lock. Return approval. Platform proceeds to dispatch.
Any rule fails
✕
A2
Block + reason code
Return block with primary reason code. Log decision with full policy state. Release lock. Suggest alternative channel if capacity exists.
↓ ↓
⑤
Audit log (immutable)
Every decision recorded · approved and blocked
Decision ID, customer token, timestamp, channel, policy rules evaluated, input state snapshot, outcome, and pre-generated explainability text. Written to append-only store.
Volume anomaly detection sub-flow
A5
A5 · Health Monitor (parallel)
Continuous volume baseline monitoring
A5 monitors total contact volume per campaign per hour, compared to the established baseline for that campaign type. Running independently of A2 · a second detection layer.
↓
Volume within normal range
✓
A5
Continue monitoring
No action. Update rolling baseline with current observation. Log health metric.
Spike >3× baseline
!
A5
Pause campaign + escalate
Pause the suspect campaign before more contacts are sent. Generate immediate alert with diagnostic context. Notify on-call marketing ops and compliance.
↓
H
Marketing Ops (human)
Investigate within 30 minutes
Review the diagnostic context provided by A5. Identify root cause. If no legitimate cause found within 60 minutes, second line is notified and campaign remains paused. Campaign cannot resume without human approval.
Why two detection layers? A2 prevents individual customers from being overcommunicated. A5 catches runaway jobs · situations where A2 approves each individual contact correctly (each one is within policy) but the aggregate volume indicates a system malfunction. Both are required for complete coverage.
Process flow · cross-portfolio orchestration
How A4 detects and resolves cross-portfolio conflicts
The most strategically significant process in the agentic architecture. A4 holds the complete picture of a customer's relationship with the institution and uses it to prevent commercially and ethically inappropriate decisions that product-silo systems cannot detect.
The commercial and ethical value: A customer receiving mortgage retention communications who is simultaneously served an aggressive credit card upsell is not just a compliance risk · it is a customer experience failure that damages the institution's relationship value. A4 makes this coordination possible at scale, across millions of customers, in real time, without any human needing to manually cross-reference product databases.
Dependency architecture
Upstream inputs and downstream outputs · per agent
Every agent in the architecture has a defined set of upstream systems it consumes from and downstream systems it produces to. Understanding these dependencies is essential for sequencing the implementation, identifying single points of failure, and designing the monitoring layer correctly.
Upstream inputs consumed
A1 · Data Prioritisation
DW field catalogue (5,000 field metadata)
Campaign performance history by segment
Business team field request log
Model usage frequency · which fields power which models
Current mapped field set (500 fields)
A2 · Policy Guardian
Real-time contact event stream (all channels)
Live contact ledger · rolling 7-day counts
Policy rule set · frequency caps, channel limits, opt-outs
Concurrent dispatch request queue
Customer consent and opt-out registry
A3 · Channel Selection
Historical channel response rates per customer
Remaining contact capacity (from A2 API)
Customer channel preference and opt-in data
Message urgency classification from Tibco
Time-of-day and day-of-week performance data
A4 · Journey Orchestration
Real-time product events (cards, mortgage, deposits)
Customer lifecycle state per product
Relationship depth score and tenure data
Active campaign and suppression windows
Service interaction history (complaints, escalations)
A5 · Health Monitor
Contact event lag metrics (both platforms)
Decision rule coverage rate (no-action %)
Campaign volume vs historical baseline
Policy breach rate from A2
Model performance metrics from A1, A3
Agent · function performed
A1 · Data Prioritisation
Scores 4,500 unmapped fields by expected decisioning uplift
Generates prioritised field mapping backlog with justification
Validates new fields when ingested
Updates scores based on outcome evidence
A2 · Policy Guardian
Enforces contact policy in real time · atomic, cross-channel
Approves or blocks every dispatch request
Detects volume anomalies and runaway jobs
Maintains immutable decision audit trail
A3 · Channel Selection
Scores each channel for each customer per decision
Selects optimal channel and timing window
Configures fallback sequencing if no response
Recalibrates utility model with response outcomes
A4 · Journey Orchestration
Detects cross-portfolio conflicts before execution
Applies suppression and holds campaigns where conflicts exist
Maintains customer journey state machine
Escalates large-scale blocks for human review
A5 · Health Monitor
Monitors all platform metrics continuously
Detects anomalies against established baselines
Generates health alerts and pauses campaigns
Produces weekly and post-campaign reports
Downstream outputs produced
A1 outputs consumed by
→ Data engineering: field mapping backlog
→ Tibco: new fields as they become available
→ A5: new fields added to drift monitoring scope
→ Marketing ops: data gap report pre-brief
A2 outputs consumed by
→ All platforms: approve / block decision
→ A3: remaining contact capacity per channel
→ A5: policy compliance rate feed
→ Audit log: immutable decision records
→ Marketing ops: escalation alerts
A3 outputs consumed by
→ Orchestration layer: channel + timing instruction
→ A2: channel-specific pre-flight request
→ Audit log: channel routing decision record
→ Model store: updated utility model weights
A4 outputs consumed by
→ Tibco: suppression instruction + context
→ A3: journey state context with each routing request
→ Marketing ops: large-scale block escalation
→ Product teams: eligibility state change notifications
A5 outputs consumed by
→ A2: platform health → tighten thresholds if degraded
→ A3: channel health → reroute if platform degraded
→ Marketing ops: campaign pause and alert
→ Risk & Compliance: governance report
→ All agents: A1 field scope updates
Sequence the implementation by dependency: A2 and A5 have the fewest upstream dependencies and should be deployed first · they create value immediately and establish the observability foundation that makes subsequent agent deployments safer. A3 depends on A2 (for capacity data). A4 depends on cross-portfolio data availability. A1 depends on the DW catalogue and campaign history. Deploy in the order that minimises risk, not the order that sounds most exciting.
Escalation & override architecture
The complete decision gate map
Every point in the agentic architecture where a human must enter the process · either because the situation exceeds the agent's authority, because a threshold has been breached, or because the decision carries consequence that requires named accountability. These are not failure modes. They are designed features of a responsible agentic system.
Designing for the override: The override record is as important as the original agent decision. Any human who overrides an agent decision must document their name, their reasoning, and the outcome they anticipated. A pattern of overrides in the same direction by the same individual is a flag for second-line review · it may indicate that the agent logic needs adjustment, or that the governance framework is being systematically circumvented.
New ways of working
How every key role transforms
Agentic Marketing does not eliminate roles · it elevates them. Every function that currently spends significant time on execution and manual coordination gains that time back for strategy, interpretation, and governance. What follows is the definitive description of how four critical functions change.
Marketing Operations
Campaign management · Audience building · Execution
Current responsibilities · manual, execution-heavy
Build audience segments manually in Tibco using available 500 fields
Configure campaign rules per channel individually
Reconcile contact history between platforms via manual batch file
Coordinate with email and SMS teams separately for channel execution
Request data fields from engineering and wait 2–6 weeks
Manually check contact policy counts before deployment
Agentic responsibilities · strategic, supervisory
Define campaign strategic objectives and success metrics · agents translate into execution
Review and approve agent-generated audience validation and channel recommendations
Monitor the A5 daily health dashboard · intervene only when threshold alerts fire
Exercise campaign-level overrides where business context agents cannot infer
Own the agent charter · define what each agent is authorised to do autonomously
Review post-campaign learning proposals · approve or modify agent recalibrations
Data Engineering
Field mapping · Data pipelines · Technical delivery
Current responsibilities · ad hoc, unprioritised
Receive unstructured field requests from marketing with no business case
Prioritise field mapping requests based on team capacity rather than business value
Maintain manual batch file for contact history synchronisation between platforms
Investigate data quality issues reactively after campaign teams surface problems
Limited visibility into which fields actually drive decisioning outcomes
Agentic responsibilities · intelligence-directed, outcome-measured
Receive A1-generated field mapping backlog with ROI justification attached · work in priority order
Build and maintain the event streaming pipeline · eliminating the manual batch file entirely
Respond to A1 field ingestion specifications · validated output, not open-ended requests
Review field drift alerts from A5 · investigate and resolve upstream data issues proactively
Measure their own output against campaign outcome metrics that A1 now surfaces continuously
Risk & Compliance
Model risk · Contact compliance · Regulatory oversight
Current responsibilities · batch, retrospective, reactive
Audit contact policy compliance retrospectively · discovering breaches after they occur
No visibility into cross-channel contact frequency · each platform audited separately
Rely on marketing to flag policy issues · no independent detection mechanism
Produce compliance reports manually from batch extracts · days to weeks turnaround
No real-time mechanism to halt campaigns that are breaching policy
Agentic responsibilities · real-time, proactive, evidence-based
Monitor A2 policy compliance rate in real time · receive immediate alerts on any breach risk
Approve the agent governance framework · the rules within which each agent operates
Countersign model recalibrations when material · independent oversight of learning
Sample audit log entries weekly · verify decisions match approved policy at the time made
Produce regulatory response packs on demand · audit log extracts available within minutes
Chair the AI Governance Committee · monthly pack review, escalation resolution
Analytics & Decisioning
Model development · Campaign analysis · Segment strategy
Current responsibilities · retrospective, model-blind
Build propensity models on 500 available fields · no visibility into which unmapped fields would improve accuracy
Produce campaign performance reports 3–4 weeks after campaign ends
No automated feedback loop · models are static between development cycles
Channel selection determined by campaign rule, not by response evidence
No cross-portfolio view of segment composition or conflict states
Agentic responsibilities · prospective, continuously improving
Design the learning architecture · how A1, A3 update their models and what triggers recalibration
Review A1 field prioritisation recommendations · validate the ROI forecasts with analytical rigour
Design the channel utility model that powers A3 · ongoing ownership of model quality
Set the monitoring thresholds for A5 · defining what "anomalous" means for each campaign type
Interpret post-campaign health reports · translate agent findings into strategic insight for leadership
Lead the SR 11-7 model validation for A3 and any model-influenced agents · independent rigour
The governing principle of role transformation in Agentic Marketing: No role disappears. Every role moves up the value stack. The work that agents take over is the work that was preventing humans from doing the strategic work they were actually hired to do. Marketing Operations was not hired to reconcile batch files. Risk was not hired to audit spreadsheets retrospectively. Analytics was not hired to produce reports. Agentic Marketing gives these functions back to strategy · and adds governance of the agents themselves as a new and genuinely skilled responsibility.
Agentic Marketing Maturity Model
Five stages from manual to fully agentic
No institution moves from Stage 1 to Stage 5 in a single programme. The maturity model provides a framework for understanding where you are, what the next stage requires, and what business and governance capabilities must be in place before the transition. Click each stage to see the full capability profile.
Stage 1
Manual execution
Campaigns built and run by people
Stage 2
Rule automation
Fixed logic executes without humans
Stage 3
Assisted decisioning
Models score; humans decide
Stage 4
Agentic execution
Agents act; humans govern
Stage 5
Self-improving system
Agents learn and adapt continuously
Capability profile
What must be true to reach this stage
Governance posture at this stage
The most important thing the maturity model reveals: The gap between Stage 2 (where most institutions believe they are) and Stage 4 (where this architecture targets) is not primarily a technology gap. It is a data architecture gap (from 500 to 5,000 fields), a channel integration gap (from 2 to 4 channels unified), a contact policy architecture gap (from batch to real-time), and a governance readiness gap (from retrospective audit to real-time compliance). Technology is the final implementation layer. Everything else must be in place first.
Complete the transformation
This framework · the interaction model, process flows, dependency architecture, escalation design, role transformation, and maturity model · provides the complete operational blueprint for implementing Agentic Marketing in a regulated financial institution. The next step is to map your current maturity stage and design the transition programme to Stage 4.
Agentic Marketing · Process Architecture & Human-Agent Operating Framework · Part II of II
Thought Leadership · For Authorised Distribution Only