If you are a marketing professional, CMO or business leader, you may recognise the following situation.
You receive a campaign plan, a market analysis or a performance report that appears complete. It has clear headings, a confident summary, perhaps a chart or two.
On closer reading, however, it does not answer the question it was meant to address. Now you have two jobs instead of one. You have to work out what is missing, and then decide whether to send it back or fix it yourself.
Last year, researchers at Stanford and BetterUp Labs gave this phenomenon a name. In a survey of 1,150 US full-time employees published in the Harvard Business Review, Niederhoffer and colleagues (2025) called it workslop, which they define as "AI generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task".
Around 40% of respondents said they had received workslop in the previous month, and each instance cost its recipient nearly two hours (1 hour 56 minutes on average). In the authors' words, workslop "transfers the effort from creator to receiver".
Marketing teams have their own version. Everyone talks about content slop, the flood of generic posts and articles online. The version that is less discussed is internal. Briefs, reports, slides, analyses, dashboards and plans are now produced faster than anyone can read them. Call it marketing slop.
What interests me is what this reveals about how workload shifts within marketing teams. Generative AI makes it possible to produce more in less time, but responsibility for quality remains with the people delivering the work. If output grows faster than their capacity to interpret, check and correct it, a productivity gain becomes an information overload problem. Fortunately, research on information overload long predates generative AI, and it offers a useful map for understanding when faster production becomes a burden.
The information overload equation and the inverted U-curve
Information overload may sound like a social media problem, but researchers in accounting, organisation science, information systems and marketing have studied it as a management problem since the 1970s.
The standard review of that work is by Eppler and Mengis (2004), published in The Information Society. Their definition is simple and, for our purposes, very useful. Overload occurs when the information processing requirements (IPR) of a task exceed the information processing capacity (IPC) of the person doing it. Both can be measured in time, that is, how much information has to be dealt with in a given period and how much a person can actually deal with in that period.
Information overload occurs when requirements exceed capacity (IPR > IPC).
The relationship between information load and decision quality is not linear. Decision quality improves as more information becomes available, up to a point, then declines as processing demands exceed capacity. This is the inverted U-curve shown in Figure 1. Beyond the peak, additional information makes it harder to distinguish relevant signals, set priorities and integrate what is already known.
Figure 1. The information overload curve (adapted from Eppler and Mengis, 2004, p. 326)
Eppler and Mengis group the causes of overload into five families.
- the information itself
- the person
- the task
- the design of the organisation
- information technology
On the last one, they wrote a sentence in 2004 that could have been written today. Information technology "can thus potentially increase the individual's IPC while at the same time increasing the IPR" (p. 331). In other words, technology can increase both how much we can process and how much we have to process.
The key insight here is that the equation does not care where the information comes from. A dashboard, a colleague, a newsletter or a language model all land on the same side of the equation.
Fifteen years later, Roetzel (2019) reviewed more than 1,500 papers on the topic for Business Research and identified a gap. Machine learning and artificial intelligence had not yet been linked to the information overload literature covered by his review (p. 507). The map existed, but AI's place on it remained underexplored.
Why AI shifts the work from production to evaluation
To understand how AI might contribute to overload, we can draw on an older field, human factors research on automation. Its central question is relevant here. When a machine takes over part of a task, what work remains for the person?
In 1983, Lisanne Bainbridge examined this question in "Ironies of automation", published in Automatica. Writing about industrial control systems, she observed that the designer "still leaves the operator to do the tasks which the designer cannot think how to automate". As she put it, "the more advanced a control system is, so the more crucial may be the contribution of the human operator."
Simkute and colleagues (2025) apply this reasoning to generative AI. In the International Journal of Human-Computer Interaction, they argue that one reason AI does not always deliver the productivity it promises is a shift in the user's role "from production to evaluation". Generating an output becomes easier, but assessing its quality and deciding how to use it remain human responsibilities. Bainbridge's problem reappears in a new setting, since automating production does not remove the need for human judgement.
Lee and colleagues (2025) show what this changed role involves in practice. Drawing on 936 examples from 319 knowledge workers, their survey identifies three activities towards which critical thinking shifts (verifying information, integrating responses into existing work, and overseeing the task). It also exposes a constraint. Respondents cited lack of time as a barrier to critical thinking, and greater confidence in AI was associated with less critical thinking. As one sales representative put it, "I use AI to save time and don't have much room to ponder over the result."
This matters for marketers because evaluation is itself information processing. AI-generated analyses, briefs and recommendations still need to be verified, integrated and assessed against the decision at hand. As their volume grows, so can the processing requirements in Eppler and Mengis's equation.
In my exchanges with clients and students, I see the same shift from producing to orchestrating, coordinating and validating. It does not necessarily mean less work. It means different work, wit
Does AI save time or create more work?
None of this means that AI inevitably creates overload. Faster production can also bring genuine relief. The question is whether those gains outweigh the work that remains or is newly created.
Dillon and colleagues (2025) examined this in a six-month randomised field experiment involving 7,137 knowledge workers across 66 firms. With an AI assistant embedded in their email, meetings and documents, active users spent about two fewer hours a week on email and worked less outside regular hours. The study is a working paper, not yet peer-reviewed, but it offers behavioural evidence that AI can save time in routine work.
The researchers also found no detectable change in the composition of participants' tasks. Individual access to AI made parts of the existing workflow faster without changing what people worked on. Making a task more efficient is not the same as redesigning the work.
A systematic review of 30 studies by Su and colleagues (2026), published in Acta Psychologica and focused on higher education, broadens the picture. It describes AI's effect on workload as a transformation rather than a simple increase or decrease. Routine activities become faster, but new responsibilities such as reviewing AI-generated output and exercising ethical judgement also emerge. Whether the balance brings relief or additional burden depends on organisational conditions, including clear policies, training and guidance.
These findings bring us back to workslop. When AI produces material that can be checked with limited effort, faster production frees time for other tasks. When the output needs substantial interpretation, correction or reworking, that saving is absorbed by evaluation. If the evaluation happens before the output is shared, the workload stays with its creator; if it does not, unfinished work travels to colleagues as workslop. The relevant question is therefore not how quickly AI produces a report, but how much human effort is needed to make it useful.
When polished AI output is not finished work
In my teaching, I regularly see AI-generated reports, plans and papers that look finished before the thinking is complete. The structure is clear, the language is fluent and the presentation is convincing, yet closer reading shows that the work does not answer the question set or develop the analysis far enough. Its apparent completeness is not matched by its substance.
Research offers some clues as to why the checking stops too early. Lee and colleagues (2025) found that lack of time is a barrier to critical thinking, and Vasconcelos and colleagues (2023) showed that reliance on AI responds to the cost of checking its recommendations. These were not studies of my students, but they help explain how generating an answer can become separated from evaluating it.
My conversations with clients reveal a related difficulty. Some describe output that needs so much checking and rewriting that the promised time saving is hard to find. Others struggle to decide where AI belongs because their processes and quality criteria are not clear enough. Both expose the gap between producing something quickly and making it useful.
I did not find a peer-reviewed study of AI-related overload among marketers specifically, but industry surveys echo these concerns. In October 2026, Optimizely surveyed 100 UK marketers. 65% said their teams felt more stretched since AI took on work, and 54% said AI often adds more work than it removes. A small vendor survey is an indication, not proof, but it suggests that what my clients describe goes beyond individual cases.
This is the question I am exploring in my own research. When does AI support the thinking behind a decision, and when does it become a veneer that makes unfinished thinking look complete? The literature offers plausible mechanisms, but not yet evidence of how widely they operate in marketing teams.
For practitioners, the immediate question is what we can change. How do we organise AI-assisted work so that faster production leads to usable output, rather than unfinished work passed to someone else?
What AI does to the information overload curve
To bring the findings together, I have adapted Eppler and Mengis's curve (Figure 2). It draws on my experience and on the research above, and it is a conceptual extension, not a tested model.
Figure 2
The detail that matters is the horizontal axis. In this adaptation, it represents how much processing work is required relative to the time and capacity available, rather than simply the amount of information received. Before the peak, additional relevant information improves decisions. Beyond it, processing demands become too great and decision quality declines.
AI can move a team in either direction. When AI output meets your standard, it reduces the human effort needed to produce usable work, whether through faster production with proportionate checking or through help with filtering and summarising. The team moves left, towards a manageable position. When AI generates more material than the team can adequately assess, it increases processing demands and moves the team right.
If the volume of work and the deadlines stay the same, the team then faces a difficult trade-off. It can spend more time checking, which absorbs the time saved or extends working hours, or it can check less, which lets unfinished work circulate as if it were complete. Both often happen at once, which helps explain how teams can feel stretched while also receiving workslop.
The dashed line shows why checking less is tempting. The report is well formatted and the summary is confident, so the work looks right even when nobody has had the capacity to verify it. The dashed line represents this apparent quality, and the solid line the actual quality of the decision. The gap between the two is what I call the veneer.
There is also a possible feedback loop. If unchecked output is repeatedly accepted, it can gradually redefine what counts as finished work, and the team starts managing to the appearance of quality rather than to the standard the decision requires. This is why I see the problem as organisational rather than a matter of individual discipline.
Define the process before you automate it
So what decides which way a team moves?
In my experience with clients, it depends largely on how the work is organised. A process describes how the work gets done and a standard defines what acceptable work must achieve.
The difficulty is rarely choosing a tool. It is making the process behind the deliverable explicit enough to decide where AI can help and how its contribution will be assessed. The chain in Figure 3 brings those decisions together.
Figure 3
The first three steps are an investment of human judgement, revisited as the work changes. Their value is that the team does not have to reinvent its criteria for every output.
This also changes what "human in the loop" means. Adding checkpoints is not enough. Review helps when the reviewer knows what to check, has the expertise to judge it and has time allocated to do so. Clear criteria reduce avoidable checking effort, but they do not remove the need for judgement.
This sequence is my practical recommendation, not a tested finding. Eppler and Mengis list organisational design and technology among five families of countermeasures without ranking them, but the Su review gives organisational conditions a clear role in whether AI brings relief or additional burden. Documenting the process is not the goal in itself. It is what makes it possible to decide where AI belongs and what usable output looks like.
How a small marketing team can put this into practice
You do not need to redesign the whole department. Next Monday morning, start with one recurring process and work through it.
- Pick one process. Choose something manageable, such as the monthly performance report. State which question it must answer and who will use it.
- Write down each step. Note what happens, what information is needed and what an acceptable result looks like. Make the decisions and checks visible, not just the production tasks.
- Test AI on one step at a time. Use real examples, and keep the test small enough to see where AI helps and where it creates extra work.
- Measure total effort, not speed. Assess the result against your criteria, and include the time spent prompting, checking and correcting. Does it produce usable work with less total effort?
- Automate what works, and keep reviewing it. Connect the steps that reliably meet the standard and keep human approval where errors could materially affect the decision. Use errors and feedback to refine the instructions and checks, and if a step keeps failing, revisit the process rather than generating another answer.
The aim is not to automate every step. It is to build a workflow that reliably delivers the quality you need, with a checking burden your team can sustain.
One last thing
AI can save time, but a faster first draft is not the same as finished work. The gain becomes real when the output answers the right question, meets the required standard and takes less total effort to deliver.
Start with one process. Make its steps and standards explicit, test where AI helps, automate what proves reliable, and keep improving it.
That is how I would put the overload equation into practice. The point is not to ask the team to process ever more output, but to make sure the output deserves their attention.
If you find this kind of evidence-based content useful, subscribe to my newsletter, where I break down marketing research and translate it into frameworks you can actually use.
