Why AI agents need marketers trained to review their work

AI use is spreading faster than organizations are learning to manage it

Marketing teams can automate the entry-level work they later rely on to develop experienced reviewers. As agents, AI systems that perform successive task steps, increase output, organizations still need people who can assess correctness, strategic fit and risk, then approve the result. Cutting the jobs where people develop that judgment can create a capacity problem now and a succession problem later.

The Marketing AI Institute’s 2026 State of AI for Business Report puts that staffing problem within a broader gap between individual adoption and organizational preparation. Among 2,109 professionals surveyed, more than half said they were incorporating AI into their workflows or changing how they worked. About 25% of responding organizations were expanding AI use under a strategy. These measures describe different populations, but they show why leaders need to distinguish employees’ use of AI from an organizational plan for managing it.

Managing that use was a central concern at MAICON 2026 in Cleveland. A participant who has attended since the inaugural gathering in 2019 recalled about 300 people then and reported nearly 2,000 attendees from 48 states and 25 countries at the latest event, held at the Huntington Convention Center of Cleveland. Earlier years focused on awareness and terminology, while the 2026 program assumed existing use and emphasized building and running AI systems. Introductory education remained available before the main conference to accommodate uneven levels of adoption.

Mike Kaput, SmarterX’s chief content officer, explained the program change in a pre-event interview on The Agile Brand: “We don’t have to convince anyone this matters anymore,” he said. Paul Roetzer, founder and CEO of SmarterX and Marketing AI Institute, delivered both the opening and closing keynotes. Their organizations have a commercial interest in demand for AI adoption guidance, which readers should keep in view when considering their recommendations. At the close, Roetzer acknowledged how difficult even a 12-month forecast had become: “It’s going to be dynamic for all of us,” he said.

Uncertainty makes a departmental operating model, agreed responsibilities, standards and approval processes around AI use, more important. It gives teams a way to manage changing assignments while keeping accountability for results clear. Leaders also have to answer a staffing question: who can judge the resulting work, and how will the organization develop more people who can?

Agents can increase review demand while weakening the pipeline that supplies reviewers

Review demand changes because agents reduce the production work separating junior and senior roles. Roetzer described agents handling substantial drafting, research and production, allowing experienced employees to complete assignments previously delegated to juniors. Junior employees can also attempt assignments that previously required greater experience. His illustration put either group at roughly 90% completion with agent assistance.

Completing most of an assignment still leaves someone responsible for reading the result, finding errors, correcting weaknesses and deciding whether to approve it. In marketing, that includes judging whether technically acceptable work serves the intended strategy. A junior marketer may produce a sophisticated deliverable without knowing which assumptions to challenge, while a senior employee may produce more independently and retain responsibility for evaluation. Production ability and evaluation ability therefore need separate assessment.

Kaput raised the same distinction through the relationship between tool knowledge and marketing expertise. In the pre-event conversation, he was asked when knowledge of AI begins to matter more than knowledge of marketing. He said there was no satisfactory answer yet. Tool proficiency can help someone complete an assignment, but leaders still need to establish whether that person understands why the result is appropriate.

The need to assess that understanding grows with output. Roetzer offered a tenfold-output scenario: if AI training enables a team to produce ten times as much work, the review workload also increases tenfold. Each item still has a final portion requiring qualified reading, correction and approval. Under that scenario, faster production increases demand for the people qualified to approve it.

A 2027 agent roadmap therefore needs to identify review owners and reserve their time before output expands. Without those decisions, drafts can accumulate faster than the organization can approve them, corrections can become inconsistent, and work can advance without a qualified decision. MAICON’s emphasis on moving beyond pilots and proofs of concept, small tests of whether an approach works, makes this an immediate question for expansion. Shared standards should define acceptable work, while review gates should identify where a person must approve it before it proceeds.

These controls address current output, but staffing decisions also determine who will review work in the future. Entry-level positions have traditionally given people repeated exposure to assignments, corrections and consequences through which they develop judgment. Removing those positions also removes employment opportunities and the developmental work attached to them. Organizations can therefore shrink their future supply of reviewers while increasing their present need for review.

Roetzer’s succession concern extends five to 10 years ahead, when organizations will still need people capable of making management and executive decisions. Pressure to reduce entry-level headcount can arrive before leaders have redesigned how they prepare those candidates. Roetzer made the need explicit: “The traditional framework is broken, and we need to start thinking now to train the next generation of workers,” he said. His opening keynote had raised the reviewer-training question the day before the closing discussion.

Redesign can change that outcome because juniors can use AI to attempt more demanding work and learn from it. Apprenticeships can preserve development, while simpler automation can reduce verification demand. A reviewer shortage or a missing generation of senior marketers is therefore avoidable. The conflict arises when companies expand production and cut the development process without replacing it.

Redesign entry-level work to teach judgment

Preserving development requires making learning an explicit part of entry-level work. Kaput proposes an apprenticeship model in which experienced practitioners teach how they reason, evaluate and decide. He attributes his own ability to assess work to a decade and a half of agency and business-writing experience. His proposal organizes junior work around acquiring that ability.

An apprenticeship would make the reasoning behind corrections visible. A junior employee would learn the underlying marketing task, observe how an experienced colleague evaluates agent output and examine why particular changes matter. Supervision would explain what makes a result strategically useful, what creates unacceptable risk and what evidence justifies approval. The agent-generated draft would give the apprentice material to assess, while the teaching would connect those assessments to their consequences.

Roetzer’s emphasis on the “apprentice” role also places learning within agent-assisted work. Entry-level employees would learn the work itself and how to coordinate agents performing parts of it, while human responsibilities could concentrate on the judgment-intensive remainder. He connected that human contribution to improving technology: “your ability to drive your own creativity and innovation within your organization keeps compounding”. Developing that ability becomes part of the investment in AI operations.

To decide what apprentices must learn, organizations need to distinguish analysis from decisions about action. Descriptive work establishes current conditions and likely developments, and AI can perform much of that analysis. Normative work determines what the organization ought to do, which remains a human responsibility. Automating analysis therefore leaves people accountable for deciding how to use it.

That accountability requires supervision to vary by task and risk. The longtime conference attendee described the degree of supervision as a “dial”: teams can require closer checking for consequential work and allow more independence where the consequences are limited. Learning to set that level of oversight is part of learning to supervise agents. It requires recognizing when an apparently complete result still needs expert attention.

The consequences of a mistake also depend on the organization’s setting. In one of two fireside conversations hosted by the same attendee, Amanda Todorovich of Cleveland Clinic discussed the transition from a handful of AI pilots toward organization-wide integration. Healthcare was presented as a field where AI will remain explicitly identified for a long time because errors can carry substantial consequences. Moving AI into routine operations therefore brings continuing obligations around supervision.

Those obligations affect employees as well as the work they approve. The other fireside conversation, with Kate Dobbs, focused on employees’ experience of adoption and the leadership responsibilities that accompany enthusiasm about technology and efficiency. Decisions about assignments, supervision and development shape that experience directly. An operating model should therefore include how people learn and how their work changes as AI use expands.

Choose the least complex automation that meets the task’s accuracy needs

Training develops reviewers, while system design can limit how much verification an assignment requires. Before assigning an agent to a task, a team should define the accuracy needed and choose a mechanism that can meet it. A simple rule may handle a narrowly defined task adequately. More complex mechanisms place additional demands on computing resources and the people checking their results.

The progression discussed at MAICON begins with a rule, which applies a predefined condition or instruction. A predictive model estimates an outcome or assigns a classification, while a generative model produces content. An agent operating repeatedly takes actions, examines intermediate results and decides what to do next. Each step changes the behavior the team must understand and verify.

Moving toward more complex mechanisms therefore creates two cost categories: compute, meaning the computing resources used to run the system, and verification. The conference attendee’s description of paying “twice” referred to those costs. In that framing, the heaviest mechanism in the progression also creates the greatest review burden. The implementation decision is whether a mechanism meets the task’s accuracy requirement with an acceptable verification effort.

Verification effort makes architecture a staffing decision as well as a technical one. A system that needs extensive checking consumes the time of people qualified to perform it, so choosing a sufficient, less complex mechanism can reduce demand before the organization responds through hiring or training. Some conference sessions accordingly emphasized fit for the task over selecting the newest technology. That choice matters more as employees take a larger role in building systems.

The building-focused presentations showed several forms that participation can take. Dan Slagen of Zapier discussed helping teams progress from using AI to building with it; Zapier has a commercial interest in teams building automated workflows. Holly Chen described broad participation in agent construction at Samsara, giving an account from within a company investing in that adoption. Both presentations brought organizational perspectives to the question of who builds.

Other presentations made construction concrete at the level of a marketing assignment. Andy Crestodina showed how an assignment could become a reusable tool, while Sandy Carter discussed lessons from building “one marketing agent three times.” As more marketers participate in system design, they also help choose required accuracy, acceptable behavior and review responsibility. Shared decision rules help those builders make consistent choices about when an agent is worth its verification cost.

Measure capacity after review, and what the recovered time enables

A suitable mechanism can still appear more efficient than it is if measurement stops at production. Kaput identified hours saved as an easy starting metric for a pilot, but a weak final objective. A team can record faster completion while overlooking the work required to make the result usable. That omission can influence an expansion decision before the full operating cost is visible.

The conference attendee called the omitted effort the “review tax”: the human work required to check and repair AI output. A few hours saved during initial production can be consumed by cleaning up agent work, bringing the net saving to zero. The calculation must therefore include reading, correction and approval alongside the time removed from drafting or research. Only then can leaders see how much usable capacity the team has recovered.

Review accounting also needs to show who bears the cost, because a saving for one employee can create additional work for an experienced colleague. That colleague may already have other responsibilities requiring judgment, making their time a constraint on the whole workflow. Counting this transfer gives leaders a firmer basis for deciding whether a pilot can continue and expand. It also reveals whether the proposed reviewers have time for the work assigned to them.

Positive net savings still leave managers with a choice about how to use the recovered time. Kaput described teams filling it with more of the same work and connected that behavior to burnout. A pilot can therefore improve task efficiency while creating a workload employees struggle to sustain. The measurement plan needs to capture what happens after the initial saving.

Kaput framed that choice through Roetzer’s distinction between making work “10% better” through productivity and making it “10 times better” through innovation. The comparison asks leaders to consider what becomes possible once routine work takes less time. Metrics can encourage teams to repeat existing production or create room to develop new capabilities. The choice affects the pilot’s value beyond its immediate time saving.

Developing future reviewers is one use of that capacity. Experienced practitioners need time to explain decisions, supervise apprentices and assess readiness for greater responsibility, but allocating every saved hour to additional output leaves that teaching unfunded in practice. Tracking development time makes the staffing intention visible in the operating plan. Before expansion, leaders can then establish who owns review, whether they have capacity, who is learning to take it on and who will teach them.

Customer agents extend the approval problem beyond the team

Internal review is also a starting point for managing interactions with agents outside the organization. Meta launched its consumer-side Muse agent on Sept. 8, weeks before the event; it and other recent competitive developments received limited conference-stage attention. The longtime attendee expected more discussion at next year’s event about interactions between customer agents and brand agents. Those exchanges extend the question of responsibility beyond work produced inside a marketing team.

When a customer’s agent interacts with a brand’s agent, marketing leaders need to map where people participate and which decisions require their approval. The map can identify where an automated exchange needs human attention and who has authority to intervene. As teams prepare for those interactions, they will need to assign responsibility across both the systems they control and the exchanges those systems enter.

Key executive takeaways

• Review capacity limits AI expansion: Marketing leaders need named reviewers and reserved review time before increasing agent output. Cutting entry-level roles can also weaken the pipeline of people qualified to approve future work.
• Apprenticeships develop marketing judgment: Experienced marketers need dedicated time to explain corrections, assess junior employees’ decisions and teach risk-based supervision. Redesign entry-level assignments around these learning goals.
• Simpler automation can reduce verification costs: System builders need to define each task’s accuracy requirements before choosing rules, predictive models, generative models or agents. Use the least complex mechanism that meets those requirements.
• Net capacity is the useful productivity measure: Pilot owners need to count reading, correction and approval time alongside production savings, including work transferred to senior colleagues. Allocate recovered capacity explicitly across output, innovation and training.
• Customer agents extend approval responsibilities: Marketing leaders need to map interactions between customer agents and brand agents, identifying decisions that require human approval and assigning authority to intervene.