AI is no longer waiting at the edge of the marketing stack, helping someone write a subject line or summarize a dashboard. It is moving into the middle of the work itself: deciding which audience to reach, which product context to surface, which journey to trigger, which campaign signal to read, and which next action to recommend.
That shift is making an old problem harder to ignore. Many marketing organizations have invested in tools faster than they have redesigned the handoffs those tools depend on. Customer data sits apart from content production. Campaign planning sits apart from experimentation. Commerce feeds sit apart from brand governance. Measurement sits apart from the systems now making or recommending decisions.
The strongest signal across recent martech and AI marketing coverage is that vendors are responding with orchestration layers. Olyzon is pitching agentic CTV coordination across planning, activation, and measurement. OuterSignal’s Monocle acquisition links customer intelligence with autonomous lifecycle execution. Google’s Universal Cart and AI Mode ads pull discovery, explanation, shopping, and checkout into the same surface. Optimizely’s Opal metrics point to marketers building agents that execute multi-step work rather than isolated prompts.
The operator lesson is not that every team needs another AI agent. It is that AI agents expose workflow debt. If the underlying handoffs are messy, automation will not remove the mess. It will move it faster, hide it deeper, and make the resulting decisions harder to challenge.
Table of contents
Jump to section:
- The orchestration pitch is a response to broken handoffs
- Workflow debt is now a performance risk
- The new control plane needs new operating metrics
- Where marketers should audit before scaling agents
- The practical test is whether work gets easier to govern
The orchestration pitch is a response to broken handoffs
The most interesting AI marketing products are not selling generation anymore. They are selling continuity.
In CTV, the promise is continuity across fragmented inventory, buying pipes, and measurement systems. In lifecycle marketing, it is continuity between customer identity, intent, offer logic, message timing, and channel selection. In search and commerce, it is continuity between discovery, product explanation, cart behavior, payment, and merchant systems. In enterprise DXP workflows, it is continuity between content operations, experimentation, analytics, and campaign execution.
This is why the language of “orchestration” keeps appearing. It reflects a real buyer problem: the tools are there, but the work still breaks at the joints.
Optimizely says nearly 1,700 customers using Opal have built more than 4,000 AI agents and run them more than 172,000 times across experimentation, campaign execution, content production, and reporting. More than 97% of the activity comes from customer-built agents, and 32% involves multi-step tasks. Those numbers matter because they show where AI adoption is moving: from individual productivity to repeatable workflows.
Google is moving in the same direction from a commerce angle. Its Universal Cart announcement says people shop across Google more than a billion times a day and that Universal Cart will work across Search, Gemini, YouTube, and Gmail, with UCP-powered checkout expanding across more markets and verticals. For marketers, that compresses several old handoffs into one environment: product feed quality, AI interpretation, cart behavior, payment, loyalty, and post-click attribution.
Those developments are different in category, but similar in strategic direction. Platforms are trying to become the connective tissue between signals and actions. That can be useful. It can also make weak internal workflows look like platform problems once the system begins operating at speed.
Workflow debt is now a performance risk
Workflow debt is the accumulated cost of unresolved handoffs. It shows up when teams know what should happen, but the process required to make it happen is manual, unclear, slow, or politically fragile.
In a pre-agentic environment, workflow debt was annoying. A campaign brief took too long. A test queue stalled. Reporting required spreadsheet stitching. Product data needed cleanup before launch. A customer segment had to be rebuilt across platforms.
In an agentic environment, the same debt becomes a performance risk because the system can act on flawed assumptions before anyone has inspected the handoff.
Gartner’s 2026 CMO Spend Survey puts numbers behind that tension. CMOs are allocating an average of 15.3% of marketing budgets to AI initiatives, but only 30% report mature or fully developed AI readiness capabilities. The same survey found that 70% of CMOs consider becoming an AI leader a critical 2026 goal, while 70% also acknowledge that their internal marketing processes are not mature enough to implement and scale AI effectively.
That is the workflow debt problem in plain language. The ambition is ahead of the operating system.
Salesforce’s 2026 State of Marketing research points to a similar gap from the customer data side. Its survey of 4,450 marketing decision makers found that 75% of marketers have adopted AI, but 69% still struggle to promptly respond to customers. Salesforce also found that marketers with unified customer data are 42% more likely to regularly respond to customers and 60% more likely to use AI agents to scale their efforts.
This is not a contradiction. It is the difference between owning AI tools and having workflows that can support AI decisions. A response agent cannot personalize responsibly if service data, sales data, commerce data, permissions, and message governance are all maintained as separate realities.
The new control plane needs new operating metrics
Marketing teams are used to measuring campaigns. They are less practiced at measuring the health of the workflows that create, approve, activate, and learn from campaigns.
That gap becomes more important as orchestration layers become the control plane. A platform may report more tests completed, more content variants produced, more journeys optimized, or more conversions attributed. Those are useful signals, but they do not answer the operator’s harder question: did the workflow itself become more reliable?
IAB’s 2026 Outlook Study found that five of the top six buyer focus areas are AI-related, with two-thirds of buyers centered on agentic AI and 96% awareness of agentic ad buying and campaign execution. The same study identifies AI-driven consumer behavior shifts, agentic buying complexity, and GenAI campaign application as major challenges.
That mix of interest and complexity is exactly why campaign KPIs are not enough. If agentic systems begin influencing buying, creative, audience logic, and measurement interpretation, teams need operating metrics that sit beneath the campaign scorecard.
Useful workflow metrics include cycle time from insight to activation, approval latency by workflow stage, rework rate after AI output, data exception rate, feed correction time, experiment completion rate, and the share of recommendations that can be explained to a business owner. These are not glamorous metrics. They are the metrics that reveal whether automation is increasing capacity or just producing more work for downstream teams.
McKinsey’s 2025 State of AI survey reinforces the point. It found that 88% of organizations regularly use AI in at least one business function, but only about one-third have begun scaling AI programs. It also found that AI high performers are nearly three times as likely as others to have fundamentally redesigned individual workflows, and that workflow redesign is one of the strongest contributors to meaningful business impact.
The lesson for marketing leaders is uncomfortable but useful. If a team cannot measure the workflow, it is not ready to automate the workflow at scale.
Where marketers should audit before scaling agents
The practical answer is not to pause AI adoption until every process is perfect. That would be unrealistic and, in many categories, commercially expensive. The better answer is to audit the workflows most likely to be automated before those workflows become hidden inside agents or vendor dashboards.
Start with signal intake. What data enters the workflow, who validates it, and where does it lose context? In lifecycle marketing, this may mean customer attributes, purchase intent, discount sensitivity, suppression rules, and consent signals. In commerce, it may mean product feeds, inventory, pricing, return terms, availability, and brand-safe product descriptions. In media, it may mean audience definitions, creative metadata, measurement windows, and conversion events.
Then inspect decision points. Every workflow has moments where someone decides what is good enough to move forward. AI changes those moments because it can draft, recommend, route, optimize, or approve at speed. Teams need to know which decisions can be delegated, which require review, and which should remain human-owned because the brand, legal, financial, or customer-experience risk is too high.
Next, map approval latency. If an agent can produce ten campaign variants in minutes, but legal, brand, or product review still takes a week, the bottleneck has not disappeared. It has become easier to flood. The fix may be modular content rules, pre-approved claims, stricter offer libraries, or clearer escalation paths, not another generation tool.
Finally, audit learning loops. A workflow is only agent-ready if the system can learn from outcomes that the business trusts. This means the team should know how experiment results, sales outcomes, customer replies, unsubscribes, assisted conversions, and qualitative feedback return to the system. Otherwise, the agent will optimize against whatever metric is easiest to observe, not necessarily the metric that matters.
This is where the recent Optimizely and Deloitte Digital collaboration is directionally important. The notable part is not simply another AI partnership. It is the emphasis on operating model redesign, content supply chain changes, experimentation discipline, and measurable outcomes. That is where many AI programs succeed or fail after the software is live.
The practical test is whether work gets easier to govern
AI workflow value should not be judged only by whether work gets faster. Speed is useful, but speed without governance often produces a larger review burden, more duplicated assets, more attribution disputes, and less confidence in automated decisions.
The better test is whether work gets easier to govern at higher volume.
Can the team see which agent changed which output? Can a marketer explain why a journey selected one offer over another? Can a media lead trace a CTV optimization back to the inputs that shaped it? Can a commerce operator correct a product claim before it appears inside an AI-mediated shopping surface? Can analytics reconcile platform-reported outcomes with the company’s own incrementality logic?
These questions are not anti-automation. They are how automation becomes durable.
The next phase of AI marketing will favor teams that treat agents as workflow infrastructure rather than novelty software. That means the real implementation work will sit in places that are easy to underfund: data hygiene, content modularity, permissions, process design, measurement architecture, approval rules, and operating metrics.
Platforms will keep promising smoother orchestration, and some of those promises will be useful. But the advantage will not go to the team with the most agents. It will go to the team whose workflows are clean enough, measured enough, and governed enough for agents to improve them without taking control of the business logic by accident.
Leave a Reply