AI agents in the marketing workflow: why the productivity story is only half true

The conversation around AI agentsAI agentsAgentic AI refers to AI systems that pursue goals autonomously by planning, taking actions through tools, and adapting based on results, with minimal step-by-step human direction.View full definition → in marketing has reached a kind of breathless momentum. Vendors are shipping agentic features into every layer of the stack. Salesforce's Agentforce, launched in late 2024 and now embedded across enterprise deployments, lets marketing teams build autonomous agents that handle lead qualification, content personalisation, and campaign pacing without human sign-off on every step. Microsoft's Copilot Studio allows similar orchestration across Dynamics and Teams. The pitch is consistent: agents do the repetitive work, humans do the strategic work, and the whole machine runs faster.

That pitch is not wrong. But it is incomplete in ways that matter at the CMO level.

The consensus view

The mainstream argument goes something like this. Marketing operations have always been bottlenecked by repetitive, low-judgment tasks: pulling performance reports, resizing creative assets, segmentingsegmentingDividing a market into distinct groups of customers who share similar needs, characteristics or behaviours, so each group can be served with a tailored approach.View full definition → lists, A/B testingA/B testingA/B testing is a controlled experiment that compares two versions of something (A and B) by splitting traffic randomly to learn which performs better on a chosen metric.View full definition → subject lines, routing leads to the right sequence. Generative AI handled content at scale; agents handle workflow at scale. The combination means a small team can produce the output that previously required a much larger one.

Forrester's research through 2025 tracked strong enterprise interest in agentic AIagentic AIAgentic AI refers to AI systems that pursue goals autonomously by planning, taking actions through tools, and adapting based on results, with minimal step-by-step human direction.View full definition →, with marketing and sales cited as the two functions most likely to see early deployment. The intuition behind this is sound: marketing workflows are relatively well-documented, the data inputs are often structured, and the failure modes are less catastrophic than, say, autonomous agents in financial compliance or medical triage. A bad email subject line costs less than a bad loan decision.

Advocates also point to the compounding effect. An agent that runs nightly competitive monitoring feeds into a brief that another agent drafts, which a human then edits in twenty minutes rather than writing from scratch in two hours. The leverage is real, and organisations that have structured their data cleanly are already seeing it.

Where the story gets complicated

The productivity case is being used to justify something more consequential: a structural thinning of mid-level marketing expertise. This is the part that deserves more scrutiny.

When an agent handles audience segmentationsegmentationDividing a market into distinct groups of customers who share similar needs, characteristics or behaviours, so each group can be served with a tailored approach.View full definition → continuously, the humans who used to do that work stop developing intuition about audience behaviour. The skill atrophies. For a CMO, this creates a dependency that is easy to miss because the outputs look fine, sometimes better than fine, right up until they do not. The agent optimises for the signals it can measure. It does not know that a competitor just changed positioningpositioningThe mental space you want your brand to occupy in your target customer's mind relative to alternatives.View full definition →, that a distribution partner relationship is under strain, or that the CFO is about to announce a guidance revision that will make the current campaign tone a liability. Those are judgment calls that require people who are paying attention and who have context the agent cannot access.

There is also a data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.View full definition → problem that the vendor narrative tends to skip past. Agents are only as coherent as the systems they read from. Salesforce (a vendor with a direct commercial interest in agentic adoption) has published figures suggesting significant productivity lifts from Agentforce deployments, but these figures should be treated as directional at best and checked against independent audits before being used to justify headcount decisions. The enterprise deployments that have struggled tend to share a common feature: CRMCRMCustomer Relationship Management: software and strategy to manage and analyse customer interactions throughout their lifecycle.View full definition → and CDPCDPA Customer Data Platform unifies customer data from all sources into persistent, actionable profiles that other systems can use.View full definition → data that was already fragmented, inconsistently tagged, or poorly governed. Agents running on dirty data do not just underperform, they produce confident-looking errors, which are harder to catch than obvious failures.

The second-order effect that almost nobody is talking about seriously is accountability diffusion. When a campaign underperforms, the post-mortem increasingly involves the question of which agent made which decision and why. Most current agentic frameworks do not produce audit trails that are interpretable by a non-technical marketer. A CMO who cannot explain to a CEO or board why a campaign went sideways, because the decision chain ran through three interconnected agents, is in a structurally weak position regardless of how good the average performance looks.

There is also a competitive dynamic worth noting. If every enterprise is using the same underlying models, fine-tuned on similar commercial data, running through the same vendor platforms, the personalisation and creative differentiation being generated is drawing from a shrinking pool of genuine variation. Homogenisation is a real risk. CocaCocaCustomer Acquisition Cost: total sales and marketing spend divided by the number of new customers acquired over the same period.View full definition →-Cola, Unilever, and a direct-to-consumer challenger brand could, in principle, be generating campaign copy through systems that share 80% of their underlying logic.

What a sharp operator should actually do

The right frame is not "adopt agents" versus "wait." It is deciding which workflows benefit from agentic automation and which ones need to stay human-intensive precisely because they build the institutional capability that makes the rest of the strategy coherent.

Routine reporting, asset resizing, list hygiene, basic performance monitoring: automate these. The cognitive load is high, the strategic value is low, and the failure modes are recoverable. Free up your team from these tasks and hold them accountable for using that time on things agents cannot do, which means market sensing, relationship management, positioning work, and creative judgment.

For higher-stakes workflows, run agents in a copilot mode rather than autopilot mode. The agent drafts the audience segmentation logic, a senior analyst reviews and approves it. The agent flags anomalies in campaign performance, a human decides what to do about it. This keeps skills alive and keeps accountability visible.

On the data infrastructure question: before expanding agentic coverage, audit what the agents will actually be reading. A two-week data quality review before an agentic deployment will save months of downstream confusion. This is unglamorous work, but it is the difference between an agent that accelerates good decisions and one that amplifies bad data at speed.

Finally, push vendors on interpretability. Ask Salesforce, Adobe, or whoever is in your stack to show you what an audit trail looks like for a specific agent decision. If the answer is vague or requires a professional services engagement to decode, that is information worth having before you move further.

The productivity gains from agentic marketing are real and the organisations ignoring them will fall behind. But the CMOs who will look smart in three years are the ones who treated agent deployment as an organisational design question, not a technology procurement question.