How does AI change the distribution of work and skills in marketing teams?

Task exposure, heterogeneous assistance and the boundaries of organisational redesign

This retrospective review considers source versions available by 17 June 2026. The cutoff defines the evidence included; preparation and source checking took place on 21 October 2026.

Keywords: tasks; skills; marketing teams; labour; human AI collaboration

1. Tasks, jobs and organisational design

Marketing roles bundle different activities. A brand manager may interpret evidence, coordinate suppliers, approve a claim and negotiate a budget within the same week. A content specialist may research, write, adapt and check material for several audiences. Assessing whether AI can perform one activity therefore leaves much of the role unresolved. The job is a bundle of tasks and relationships, not an indivisible unit of production.

This review asks how task-based labour research and recent AI field studies inform the distribution of work and skills in marketing organisations. It examines the difference between technical exposure, observed assistance effects and changes in employment. These concepts frequently appear together in discussion, but the inference from one to another requires additional evidence.

The question also concerns coordination. Removing a task from one role can alter what other people need to know, review or decide. A faster draft may increase the volume awaiting approval; a broader recommendation may cross an existing functional boundary. Organisational consequences therefore depend on how tasks connect, as well as on the capability of the model performing them. The reviewed research offers distinctions for examining those connections without prescribing one universal organisation chart.

2. Review method and version boundaries

Targeted searches on 21 October 2026 located twelve primary research sources on task allocation, technological exposure, assisted performance, teamwork, adoption and labour-market outcomes. Author or publisher abstracts, institution-hosted research summaries and bibliographic records were consulted. This is a selective narrative synthesis, with no independent screening, full-text statistical appraisal or meta-analysis. Quantitative statements are retained only where they appear in the inspected primary materials.

This snapshot uses the March 2025 working-paper account of The Cybernetic Teammate, involving 776 professionals [7]. Its sample and reported outcomes are kept together. Findings from a subsequent journal account are not incorporated into this earlier evidence base. The customer-support study is discussed through its identified working-paper summary [4]. Version discipline matters because a project title can refer to reports with different samples, analyses and publication status.

Historical task-based models are used to clarify mechanisms, rather than to classify current marketing jobs mechanically. Generative AI changes the capabilities available for some language and reasoning tasks, so an old routine/nonroutine distinction cannot be imported as a complete exposure map. Likewise, an exposure assessment describes potential effects under its rubric; it cannot provide a measured forecast of redundancies, wages or hiring.

3. The task perspective on technological change

Autor, Levy and Murnane study computerisation through changes in task requirements, distinguishing routine and nonroutine work [1]. Their framework shifts attention from occupation labels to the content of production. It explains why the same technology can replace some activities while complementing others within an occupation. Its empirical setting concerns earlier computerisation and does not estimate the performance of present language models.

Acemoglu and Autor distinguish worker skills from the tasks to which those skills are assigned [11]. A skill is a capability held by a person; a task is an activity contributing to output. Technology can change the allocation between them. This distinction prevents a task from being treated as permanently attached to one job title and prevents a change in allocation from being mistaken for the disappearance of the underlying capability.

For marketing, the analytical unit might be source gathering, summarisation, claim checking or selecting a creative route. These examples are task descriptions rather than estimates of automability. Their resource requirements and acceptance criteria differ. One may be largely constrained by access to information, another by evaluating uncertain alternatives, and another by authority to approve. A common model interface does not make those production problems equivalent.

A task inventory also needs dependencies. Writing a campaign recommendation depends on what counts as sufficient evidence; approving it may depend on a commercial objective and knowledge of constraints outside the draft. If the upstream task changes, downstream work can change too. Measuring each activity in isolation can miss the way that a new division of labour redistributes coordination and verification.

4. Exposure assessments describe potential

Eloundou and colleagues assess US occupations using a rubric combining human expertise and GPT-4 classifications. Their author abstract distinguishes an LLM alone from software built around it and explicitly avoids predicting an adoption timeline [3]. This makes exposure an assessment of potential task impact under specified criteria. It is not an observed employment effect.

The distinction has practical consequences for interpreting marketing forecasts. A role with many language tasks may appear exposed while its actual use is constrained by incomplete records, review requirements or organisational decisions. Another role may benefit indirectly because colleagues produce better inputs. A capability map can identify where to investigate, but the resulting work allocation depends on adoption and production conditions.

Exposure also need not imply replacement. If a task becomes cheaper, an organisation can reduce expenditure, increase output or change quality. Demand for related work can rise or fall depending on how those responses interact. The classification of a task as technically affected does not select one economic response. A claim about headcount requires observations or a model of those additional choices.

Aggregating tasks to occupations introduces weights. Counting the number of affected tasks assumes that all tasks contribute equally, while weighting by time or value answers different questions. Some infrequent tasks determine whether the role can function at all. A task that occupies little time may require authority or expertise that prevents an entire role from being automated. The aggregation rule should therefore accompany any occupational exposure percentage.

5. Assistance effects vary across people and assignments

The customer-support study reports heterogeneous productivity benefits, with larger gains for less-experienced or lower-performing workers [4]. Noy and Zhang's writing experiment reports improvements in task speed and assessed quality [5]. These findings provide evidence that assistance can change performance distributions in defined settings. They leave open how a marketing team's particular task mix responds.

The consultant experiment shows that performance can worsen on a task outside the assessed capability frontier even while improving on other tasks [6]. This matters for task allocation because a role can contain both kinds. Assigning the whole role according to its average apparent suitability can expose the organisation to failures on a small but consequential component.

A distributional gain is also different from skill acquisition. Improved assisted output can reflect access to useful suggestions, faster drafting or a new search process. Establishing that the worker has learned a transferable capability requires measuring later performance under conditions that test that capability. The operational value of assistance can be substantial without resolving the learning question.

Observed differences by experience should be interpreted within the study. Years in an occupation, baseline task score and familiarity with a particular workflow are related but different measures. A result that varies by one of them cannot automatically be described as evidence that all junior workers benefit more or that all experts benefit less. Marketing organisations often combine general seniority with highly task-specific experience.

6. Teamwork and expertise integration

The March 2025 working-paper account describes 776 professionals at Procter & Gamble addressing product-innovation challenges in individual or team conditions, with or without AI. Individuals using AI achieved proposal quality comparable to teams without it. AI use also altered the balance between technical and commercial orientations [7]. The comparison concerns that task and sample. It does not establish an optimal team size, replace every collaborative function or measure a complete marketing organisation.

This is unusually relevant to marketing because the study crosses commercial and research-and-development expertise in a consumer-goods company. Its direct evidence concerns the assigned innovation tasks and assessment. Comparable output in that setting does not mean that all benefits of a colleague, specialist function or ongoing relationship are replaceable. Teams also operate across time, hold responsibility and coordinate implementation.

The result identifies a question about where expertise enters work. If a model supplies technical considerations to a commercial professional, the output may cover a broader set of issues. Whether the professional can verify those considerations is a further matter. Broader coverage and deeper accountable expertise should therefore be assessed separately. A recommendation that mentions both functions can remain weak in one of them.

Team-level evaluation also needs to distinguish idea generation from selection. A system can improve the available options while a human still determines which option fits the assignment. The contribution of each stage can be obscured if only the final proposal is rated. Process observations can reveal where the gain occurs, but identifying the causal effect of each component requires a design that varies those components.

7. Adoption changes who receives the potential benefit

Humlum and Vestergaard study ChatGPT adoption using Danish survey and administrative data. Their published research identifies substantial differences in use, including a gender gap, with employer restrictions and perceived training needs among the barriers discussed in the institutional summary [9]. Potential capability and actual access to its benefits can therefore differ across workers.

Bick, Blandin and Deming's adoption research also distinguishes use of generative AI from the tasks and time to which it is applied [12]. A headline adoption rate can hide considerable differences in intensity. For a team, the relevant distribution includes who uses the system, what work they use it for and whether the results enter accepted output.

An organisational redesign based only on highly active users can mistake a selected group's experience for the whole team's. Volunteers may have more suitable assignments, greater confidence or better support. Non-users may face restrictions that the redesign can change. Evaluating those differences requires observing the conditions of use, rather than inferring motivation from the adoption label alone.

Training can address some barriers while leaving others intact. Familiarity with a tool cannot repair missing source data or grant authority to approve a decision. Conversely, access to an enterprise account does not establish competence in using it. An assessment of skill and opportunity needs to separate these conditions if it is to explain why use and performance vary.

8. From task substitution to labour-market outcomes

Acemoglu and Restrepo distinguish displacement through automation from the creation of new tasks that can reinstate labour demand [2]. Their framework allows productivity and labour demand to move differently. The economic outcome depends on which tasks change, what new activities emerge and how output demand responds. It supplies no fixed conversion from an automatable task share to a job-loss percentage.

The Danish labour-market research links AI adoption to administrative outcomes and finds limited early changes in measured earnings and hours despite widespread adoption and reported workplace changes [8]. The observation is consistent with adjustment occurring through tasks before it appears in those aggregate outcomes. It does not determine the future path of employment or exclude benefits outside the recorded measures.

At the organisational level, a task can be removed without a role disappearing. Remaining tasks may expand, more work may be demanded or responsibility may remain indivisible. A role can also disappear for reasons unrelated to AI while some of its work continues elsewhere. Attributing a staffing change requires an appropriate comparison and a record of the alternative that would have occurred.

A reduction in external agency work is another allocation outcome. It may shift tasks to internal staff, to software or to a different supplier. Its effect on the advertiser's productivity depends on the total resource change and accepted result. Observing a lower agency invoice alone leaves internal review and coordination costs unmeasured. The boundary of the organisation should not determine whether work is counted.

9. Expertise, verification and decision rights

Historical firm-level research connects information technology, workplace organisation and demand for skills [10]. For current marketing work, the relevant inference is that technology changes can interact with how decisions and knowledge are distributed. The historical evidence cannot select the correct degree of centralisation for every company or the job titles a contemporary team needs.

Decision rights are distinct from the ability to produce an answer. A system can suggest a price promotion, but the person authorised to commit expenditure may remain elsewhere. Faster production of recommendations can increase the burden on that decision point. Redesigning tasks without examining the approval constraint may move a bottleneck rather than remove it.

Verification is itself work requiring knowledge. A reviewer needs to identify which claims are consequential, what sources support them and whether a recommendation fits the objective. If the organisation reduces the time allocated to review while increasing output volume, the quality of the complete process can change. That is a measurable workflow effect, not a necessary property of AI assistance.

The distribution of expertise also has resilience implications. A process that depends on one person recognising subtle failures is vulnerable when that person is unavailable. A process that distributes every judgement across a committee can incur coordination costs. The trade-off depends on task consequences, information and workload. The reviewed evidence supports examining those constraints; it does not establish a universal ratio of reviewers to generated outputs.

10. Evaluating a changed division of labour

A before-and-after organisation chart supplies little evidence about whether the new division of labour improves work. Evaluation needs the tasks, resource use, accepted quality and consequences of failure. A comparison can then ask whether work has become faster, more accurate or less expensive at an equivalent standard. The answer may differ across the portfolio.

The unit of randomisation or comparison should match the intervention. Individual access can be studied at the worker level, while a changed approval process may affect an entire team. Colleagues can also share outputs or techniques, creating spillovers. An individual comparison that ignores those interactions can estimate a different treatment from the organisation-wide change of interest.

Learning and staffing outcomes require time. A short task experiment can reveal immediate assistance effects while leaving the development of expertise unresolved. Hiring and career progression respond to demand, turnover and managerial expectations as well as tool capability. Claims about long-term skill formation therefore need longitudinal evidence with a clear comparison, not only a current performance snapshot.

Performance distributions should remain visible. An unchanged mean can accompany better performance among some workers and worse performance among others. A higher mean can coexist with a wider gap in access or a larger tail of failures. The organisational decision may depend on those distributions, especially where a particular role holds an important approval responsibility. A single average cannot answer every distributional question.

11. Skill formation and the scope of current evidence

An entry-level task can be both productive work and a route to learning. If AI changes who performs it, the training pathway may change even when immediate output improves. Establishing the effect requires evidence about what novices practise, what feedback they receive and what they can subsequently do. The presence of a better assisted output alone leaves those mechanisms unobserved.

The task perspective also separates durable domain knowledge from interface-specific familiarity. Knowing how to operate a tool can improve current performance, while knowledge of consumers, measurement and business constraints supports evaluation across tools. Both may be useful. Their relative importance depends on the task and can change with the technology. A general assertion that prompting replaces marketing expertise exceeds the selected evidence.

A team can learn from model output, but it can also accept it without acquiring the underlying capability. Distinguishing those outcomes requires an assessment designed to test understanding or independent performance. Organisational records of completed work are useful operational data, yet they are not automatically a measure of learning. The criterion should be chosen to match the claim.

11.1 Coordination and the interpretation of a smaller team

A smaller team can reflect higher productivity, a narrower remit, lower demand or work transferred elsewhere. Headcount alone cannot distinguish these explanations. An assessment of organisational redesign therefore needs a stable description of the work covered, the quality required and the resources used beyond the team's formal boundary. If an external specialist or central function absorbs exceptions, its contribution belongs in the account.

The same applies to broader roles. Giving one person responsibility for more channels can reduce handoffs, but it can also increase the range of judgements that person must make. The result depends on whether the system supplies useful support, whether the person can evaluate it and whether decision volume remains manageable. A job description records the intended allocation; observations of accepted work establish what the arrangement actually produces.

A comparison should also distinguish vacancy management from task removal. An organisation can choose not to replace a departing worker while distributing the remaining work across colleagues. That staffing outcome may be associated with AI use, but the causal contribution requires evidence about workload, alternative hiring and the support provided. A statement that AI replaced the role is stronger than an observation that headcount fell during adoption. Task-level records and an appropriate comparator make the difference assessable.

12. Conclusions and limitations

The reviewed research supports analysing AI through tasks and their allocation, with observed effects varying across people and assignments. Exposure, assistance, reallocation and employment remain separate outcomes. Teamwork evidence shows that AI can alter the combination of expertise in a defined innovation setting, while adoption research shows that access to potential gains is uneven.

This synthesis is selective and largely based on primary summaries. It does not predict marketing headcount, prescribe a universal structure or estimate long-term career effects. Historical models and current experiments have different scopes, and publication versions must remain separate. Direct evidence on a particular marketing organisation is needed to establish how its own task distribution changes.

A defensible organisational assessment follows the work beyond initial production. It records who verifies, selects and authorises, and measures the complete result. That makes it possible to distinguish a changed interface from a changed division of labour, and a changed division of labour from a demonstrated improvement in organisational performance.

Source access and appraisal register

The 12 references below were located through targeted title and topic searches on 21 October 2026. Author or publisher abstracts, official research summaries and publication records were consulted. The discussion uses the claims available in those materials; full manuscripts, supplementary analyses and underlying data were not independently appraised. Institutional guidance is treated as guidance, rather than a controlled estimate of effectiveness. Where a working paper is cited, its provisional status remains part of the interpretation. The evidence map is selective and does not estimate the prevalence of results in the complete literature.

References