Optimised to a Farm: On AI Adoption in a Creative Function
The creative function is the first to hollow. It is rarely the last.
Some changes are sowed unnoticed, brief by brief, until they settle as a muted norm. Each decision is rational in isolation and defensible on a slide. Then one day the creative team is still showing up and hitting deadlines, but the work offends nobody, surprises nobody, and is remembered by nobody.
The best people are professionally present and creatively absent. Some have left. Others have learned to stop bringing ideas nobody will act on. Meanwhile, the numbers improve: more output, shorter cycles, lower costs.
No one decided against originality, just no one decided for it either. This is “the farm”: a creative function optimised towards interchangeable output, with its capacity for differentiation quietly spent along the way.
Preventing it takes two lines. The first defines where human judgment must lead the work. The second protects the time and resources needed to exercise that judgment. A boundary inside the workflow needs a commitment in the governance structure, or it lasts only until the next urgent brief.
The mechanism
AI can improve an individual piece of work while making the wider body of work more alike. In a short-story experiment, Doshi and Hauser found that access to AI-generated ideas improved evaluations of individual stories, particularly for less creative writers, while reducing diversity across the stories produced.¹
For a creative function, that tension matters. A stronger average output can coexist with a narrower range of ideas. The brief gets a suggested structure, the concept a reference bank, the headline ten ranked variations. Each intervention may be useful. The risk emerges when the team loses the space to question the structure, depart from the references, or reject the options altogether.
The effect also reaches the person doing the work. Task identity—the experience of completing a whole, identifiable piece of work—is one of the job characteristics linked to meaningfulness and motivation in Hackman and Oldham’s model.² Fragment the work into reviewing and refining outputs, and a person can remain responsible for delivery while losing their sense of authorship. Task volume stays. Investment withdraws.
The economic loop reinforces this. Automation frees capacity; the capacity is absorbed into savings or more delivery; the efficiency metrics rise; management sees evidence that the model works. Business-as-usual expands to fill the margin that might have supported original thought. Nothing in the sequence needs to look like a mistake.
The manager presenting those results may have earned their pride. Everyone in the room is responding to what the room rewards. But the numbers reveal little about the ideas no longer proposed, the assumptions no longer challenged, or the judgment no longer exercised.
One place to look is detail literacy: the ability to explain what produced a result, what it rested on, and what was tried and rejected. Research on AI use and critical thinking gives reason to watch this closely: a survey of knowledge workers associated greater confidence in AI with less reported critical-thinking effort.³ It does not establish inevitable skill loss, but it raises a practical question: is the team still doing enough of the thinking to understand the work it approves?
A longer-term warning comes from an accounting firm studied by Rinta-Kahila and colleagues. After seven years of automation, withdrawing a fixed-asset management system exposed how much relevant expertise the accountants had lost.⁴ Creative work can offer a less decisive reckoning. Weak results have many plausible explanations, allowing the loss of judgment to remain obscured until an unfamiliar problem makes it necessary.
The resource signal
An allocation weighted towards AI also sends a message to the people carrying the remaining human work. They may be asked to maintain the standard with less investment, coordinate the automated output, and watch their distinctive contribution become less central to the function’s priorities.
Research on effort-reward imbalance describes the strain of sustained high effort under low reward.⁵ In a creative team, one possible response is withdrawal: people continue meeting the brief while reducing the thinking they once volunteered beyond it. The organisation receives the work it specifies and loses the work it never knew to request.
This process need not follow a steady slope. It can come in tides. A funded adoption push rewards standardisation: processes built, workflows automated, capacity released. Once the push recedes, the remaining task is to judge whether the output is any good. That requires capabilities the push may have given little room to practise.
Then the team recovers some breathing space. Enough work improves that the pressure seems temporary. Another push follows. Each phase can appear manageable while the cumulative loss continues. The tide keeps receding before the damage becomes undeniable.
The prior mistake: AI as a problem-flattener
Before deployment, there is often a failure of diagnosis. A campaign keeps missing. A brief lands too narrowly. A team stops pushing back. The cause might be a trust deficit, a cross-departmental conflict, an unclear decision, or a structural problem at the top. Each calls for a different intervention.
When AI has already been chosen as the organisation’s modernisation strategy, these problems can all be routed towards the same answer. A communication failure becomes a personalisation problem. A leadership dynamic suppressing initiative becomes a workflow problem. A disagreement about direction becomes a testing problem.
The original problem is translated into a shape the tool can process. More versions, faster execution, and additional tests may follow while the cause remains untouched. Successful implementation can even delay recognition of the failure: the organisation now has evidence that something changed, without evidence that the right thing changed.
The tell is simple. When AI is the first word in the conversation, before the problem has been named, the diagnosis has been skipped. Defining what is wrong must remain a leadership responsibility, including when the answer turns out to be a problem AI can help solve.
What to read
By the time turnover and output metrics move, the intellectual capital has already left. The early signals are often behavioural. Five are worth watching.
Watch how initiatives end. A killed idea had a decision behind it. An archived one simply stopped having a champion. When the archive grows faster than the kill list, ask whether the organisation is still making choices or merely allowing uncomfortable ideas to expire.
Notice the split inside the team. Some people feel enabled by AI; others feel worked around by it. Treat the second response as information about how the work has been designed. It may reveal a contribution being displaced without anyone deciding whether it still matters.
Test whether the numbers come with details. Ask what produced the result, what assumption it rested on, and what was tried and failed. A team that can answer retains a relationship with its work. Repeated inability to answer locates where that relationship has weakened.
Read turnover as a capital account. A departure can remove relationships, accumulated judgment, and a perspective the team relied on without documenting it. Filling the role does not immediately restore those assets. Examine what leaves alongside how many leave.
Listen for AI named before the problem. A prescribed tool can make a discussion feel decisive while leaving the actual diagnosis undone.
Holding both horizons
Near-term efficiency is a legitimate business need. A creative function has to deliver, and standardisation can make delivery faster, more consistent, and more affordable. The danger is allowing that need to consume the capacity through which the function develops something distinctive.
Lean too far towards the near future and the work becomes easier to substitute. Lean too far towards the far future and the organisation can exhaust its resources before the experiments pay. Leadership has to hold both horizons: delivery that funds survival, and exploration that develops future value.
This means carrying two ledgers. One records measurable output. The other tracks the health of the function producing it: the quality of its judgment, the trust within the team, and the real chance of an uncomfortable idea being heard. The second ledger requires observation and explicit discussion, even where a clean metric is unavailable.
The balance tips when performance on the first becomes an alibi for neglecting the second. Someone must be accountable for both, with the authority to make the trade-offs visible.
For anyone who cares about brand, the question is: are we holding both horizons, or have we quietly let the near future spend the far one?
Two architectures
Individual conviction can protect this balance for a while. A leader draws a boundary, defends an experiment, or absorbs the KPI cost of preserving capacity. But that defence is vulnerable to a change of manager, a tighter quarter, or a reorganisation.
The commitment needs to survive the person making it. Two architectures show the difference. Their suitability depends on what the business expects from creative: efficient communication of an existing proposition, or a contribution to developing the next one.
Model A: creative as a service house
In Model A, Product/Service owns the narrative and creative translates it into communications. The function is organised around delivery. This can be a rational design where consistency and cost matter most, or where responsibility for exploration sits elsewhere.
The risk appears when the business also expects this function to generate differentiation. Management promises that automation will free creatives for higher-value work, but assigns no capacity to that work. The savings become a lower cost base or a higher delivery target. Improved efficiency then justifies repeating the process.
Meanwhile, people have fewer opportunities to exercise the capabilities that attracted them to the work. Some leave; some reduce their investment. The function becomes less able to originate the work it is nominally expected to produce.
The cost is difficult to book. An opportunity never explored leaves no failed project to review. A narrowing range of ideas can coexist with reliable delivery. Where creative differentiation matters, the service house can become the farm while still meeting its targets.
Model B: creative beyond the service house
Model B keeps the delivery engine and adds a standing exploration branch, run by the same squad.
Branch 1 is the service house: semi-automated, fast, and consistent. It handles business-as-usual and supports the near horizon.
Branch 2 is a regular path from observation to experiment to launch. Its sessions begin without a predetermined campaign brief or a fixed deliverable to fulfil. They have a strategic direction, but permission to discover what is worth pursuing within it.
The branch produces two kinds of value. Observations feed improvements back into the service house. More ambitious ideas enter a small in-house launchpad, where the team can test a new product, format, brand proposition, or piece of owned IP. What proves useful feeds back into Product/Service. This is the far horizon.
One squad circulates between the branches. Participation and weighting vary with the work, while protected exploration capacity remains. Frontline experience in Branch 1 supplies observations for Branch 2; exploration brings new judgment and possibilities back into delivery. Keeping those experiences connected also avoids dividing the team into permanent classes of executors and originators.
The distinction is in how the work starts and how it is governed. Branch 2 has accountability for learning, testing, and advancing worthwhile ideas. It retains the freedom to discover an opportunity before being asked to execute a solution.
The second line, defined
The first line assigns human responsibility for framing the problem, setting the creative direction, and judging the result. AI may inform those decisions. The team must still be able to explain and defend them.
The second line protects the conditions in which that responsibility can be exercised. Model B gives substance to the promise of “freeing creatives for higher-value work” through three commitments.
A protected cadence. Branch 2 runs on a fixed rhythm, with the standing of a business review. Its sessions and development time are planned capacity. “When there is time” is how the far horizon disappears.
An empowered owner. A named person is accountable for the branch and authorised to defend its allocation. Any decision to reclaim that capacity must be explicit, with an agreed route for resolving the conflict. The role needs backing from the level that sets delivery expectations; a title alone cannot withstand a contradictory mandate.
An efficiency-dividend rule. A declared share of the capacity automation releases goes to Branch 2. The allocation is agreed when the automation is deployed and reviewed against the capacity actually freed. Exploration receives a defined claim before every gain becomes another delivery expectation.
These commitments make the trade-off visible. A business may still decide to reduce exploration during a crisis. It should have to name that decision, its cost, and when it will be reviewed. Otherwise a temporary exception can quietly become the operating model.
The two, side by side
Model A is simple, legible, and economical to run. It fits where reliable delivery is the creative function’s primary contribution, and differentiation matters less commercially or is developed elsewhere. Its characteristic risk is extraction: capacity is repeatedly reclaimed for throughput, and the ability to originate new work drains away while delivery metrics improve.
Model B earns its governance overhead where differentiation has commercial value and creative is expected to help generate it. Its returns take longer, and its characteristic risk is theatre: sessions run, but ideas never reach a meaningful test, or delivery pressure reclaims the protected capacity. The choice turns on what the business expects creative to contribute. If that includes developing future value, the time, authority, and resources to do so must be part of the design.
The minimum viable version
Model B can begin without a reorganisation. Start with a named owner, protected time, a small seed budget, and an efficiency-dividend rule. Aim to bring one idea to a meaningful pilot each quarter, with scope matched to the resources available.
The pipeline is simple: capture, develop, pilot, then scale or kill. Record what each test taught the team and why it advanced or stopped. A justified kill is a useful outcome; an idea left indefinitely in the archive is an unresolved decision.
Give seeds room to develop, then bring Product/Service into the pilot and scale decisions. The branch needs a route into the business as well as protection from its daily demands.
Defend one property from the start: exploration must be able to originate its own questions. Once the branch exists only to fulfil incoming campaign briefs, its purpose has been absorbed back into the service house.
Coda
The farm forms through a chain of reasonable decisions that gradually removes the conditions for original work. Its alternative must be built into the decisions that allocate time, authority, and resources.
Draw the first line around the judgment people remain responsible for. Make the second line a commitment the organisation must answer for. Then the capacity AI releases has somewhere deliberate to go, and the creative function has room to build what the next brief does not yet know to ask for.
Notes
¹ Doshi, A.R. & Hauser, O.P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290.
² Hackman, J.R. & Oldham, G.R. (1976). Motivation through the design of work: Test of a theory. Organizational Behavior and Human Performance, 16(2), 250–279.
³ Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R. & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. ACM, Article 1121, 1–22.
⁴ Rinta-Kahila, T., Penttinen, E., Salovaara, A., Soliman, W. & Ruissalo, J. (2023). The vicious circles of skill erosion: A case study of cognitive automation. Journal of the Association for Information Systems, 24(5), 1378–1412.
⁵ Siegrist, J. (1996). Adverse health effects of high-effort/low-reward conditions. Journal of Occupational Health Psychology, 1(1), 27–41.
Theory anchors. The two-branch design draws on the tension between exploration and exploitation described by March, and on organisational ambidexterity as discussed by O’Reilly and Tushman: sustaining present delivery while developing future possibilities. The rotating squad and efficiency-dividend rule are this article’s proposed applications of those ideas.
March, J.G. (1991). Exploration and exploitation in organizational learning. Organization Science, 2(1), 71–87.
O’Reilly, C.A. & Tushman, M.L. (2013). Organizational ambidexterity: Past, present, and future. Academy of Management Perspectives, 27(4), 324–338.
