Generative AI has streamlined the creative process, but it also leaves a huge hidden pitfall.
Generative AI has drastically lowered the threshold for producing creative concepts, allowing users to effortlessly generate visually stunning, highly immersive renderings. However, many of these images are nothing more than the product of algorithmic imagination, failing to conform to the objective conditions of real-world shooting and production. Those polished reference images will quietly raise clients' psychological expectations. When the project moves to the implementation phase, the gap between reality and AI-generated concepts will translate into a large amount of extra communication, adjustments, and rework. How to preserve the creativity itself and prevent the exquisite illusions generated by AI from hijacking real project execution has become a realistic proposition that creative teams must face.
Imagine this scenario: a photographer is reviewing a set of concept art for an advertising campaign that has already been approved by the client. He spots a problem: the man in the image is wearing a baseball cap, the brim of which should cast a shadow over his face, yet his eyes are clearly lit, staring sharply directly at the camera. From his professional perspective, the photographer can tell at a glance that this kind of lighting is impossible to achieve in reality, and no on-site shooting can produce such an effect. This image was created by a senior creative professional using generative AI, and the client is very satisfied with it. Now the production team has been tasked with using real cameras and live actors to recreate an image that cannot be realized in the real world.
This is just a real case that reflects a more universal issue, which is also the theme of our year-long research: How generative AI tools are reshaping creative workflows. For decades, one of the major pain points in creative work has been persuading clients to imagine the final look of the finished product through rough drafts. Generative AI has completely turned this problem on its head: today, as long as a practitioner has AI tools at hand, they can produce nearly complete-looking solutions that can be directly used for presentations. At this point, however, the expert team responsible for implementation has not fully polished and refined the idea.
We refer to this phenomenon as workflow collapse. Research shows that it triggers a series of subsequent problems: clients develop unrealistic expectations, and time and resources are wasted. But many companies are neither aware that their workflows have changed, nor have they established measures to deal with the various consequences brought about by this change.
Although the subjects of this study are mainly practitioners in the advertising industry, the findings are equally applicable to managers in charge of creative work in many fields: when product teams make polished prototypes, consultants write detailed client plans, architects use AI to generate drawings, and software teams use "vibe coding" to quickly build demo versions of client websites, all of these scenarios face similar problems.
Based on the findings of this study, we help managers identify which links AI has compressed or even skipped key steps, so that enterprises can not only enjoy the convenience brought by the tools, but also avoid various subsequent derived problems as much as possible.
The Creative Workflow Has Undergone Fundamental Changes
For this study, we partnered with RTL Nederland, one of the largest media companies in the Netherlands. Creative practitioners in the company's advertising department have been using generative AI on a regular basis since mid-2023, especially the text-to-image tool Midjourney. From March 2024 to March 2025, we conducted research on multiple teams that produce advertisements for streaming media and TV programs, observing their workflows and conducting interviews. To restore the changes that have taken place in their work, we also collected presentation materials from before the introduction of generative AI, and asked creative practitioners to compare the working modes before and after the arrival of AI.
During the research period, creative staff followed RTL's internal usage policies, which clearly specified permitted tools, as well as relevant requirements for client approval, copyright and data protection, verification of records retained for AI-generated content, and the principle that humans should maintain core creative leadership over AI outputs.
Before the popularization of generative AI, at least two senior creative professionals would brainstorm campaign ideas together. Then they would send requirements to graphic designers, who would produce several visual solutions for several core concepts. The senior creative professionals would then present these visual materials to the client. The presentation materials would be deliberately made relatively rough, using only sketches, mood boards, and Photoshop collage drafts to convey the core creativity. When the client signed off on the approval, they could only perceive the general direction of the campaign, and ultimately had to trust the creative team to further refine and implement the advertising solution in the subsequent production phase.
After the introduction of generative AI tools, the entire workflow has become completely different. Take this case as an example: let's call one of the senior creatives Ben. Sitting at his workstation, he directly operated Midjourney to brainstorm, generate visual images, and iterate repeatedly for a dating reality show advertisement. The whole process involved only him using prompts, with no other senior creatives participating, and no graphic designers involved.
The images returned by Midjourney often brought him surprises, both pleasant and troublesome. Ben once was very troubled by the weird output: he entered a prompt asking for "an old hotel key card with a red lipstick kiss mark on it, suede green background, photo texture". The generated result was completely off track: instead of a lipstick mark on a square key card, the image showed an old-fashioned metal key, with a pair of disproportionately large, photorealistic red lips floating above it, looking very strange.
But we also witnessed many moments of inspiration, where a peculiar detail output by AI ended up shaping the core idea of the entire advertisement. Another senior creative asked Midjourney for runway images for a model reality show advertisement, and the returned result was a coastal boardwalk covered with red carpets, surrounded by flashing camera lights. This image inspired him, and the core concept of the entire advertisement was thus determined: paparazzi photographers holding flashes, snapping candid photos of the show's hosts.
With the help of generative AI, the time spent producing presentation materials has been greatly reduced. The original multi-round communication and handover processes, including brainstorming with other senior creatives, submitting requirements to graphic designers, and coordinating with photographers, are now all compressed into one person continuously interacting with AI.
This brand-new presentation creation model has its advantages. A senior creative said: "Now AI, especially Midjourney, can directly visualize my ideas, making it much easier to persuade clients." Many senior creatives share the same view: the images produced by generative AI are polished enough to impress clients more easily than rough sketches. Clients no longer need to imagine what the final creative implementation will look like on their own, and can directly see the effect with rich details.
However, in the downstream implementation phase, we found that AI-generated concept art will cause clients to lock in the details of the solution prematurely, bringing a series of troubles to the production team.
Various Derived Implementation Challenges
If the advertisement ultimately uses live-action shooting instead of AI-generated images, the situation where clients lock in the solution prematurely will cause particularly prominent troubles.
One creative practitioner admitted that managing client expectations has become extremely tricky. The AI concept art is so polished that to satisfy the client, "you basically have to recreate the exact image you sold earlier", leaving much less room for the production team to further refine the creativity. Creative staff mentioned that clients often assume that all the details they have finalized are standards that must be strictly followed in subsequent implementation, but these details were originally only conceptual references, not mandatory requirements for execution, which makes expectation management even more difficult.
We followed an advertising project like this: the creative team used AI to generate a concept art of Norway's rugged coast. After finalizing the plan with the client, an entire production crew, including photographers, lighting technicians, costume staff, makeup artists, and actors, all traveled to Norway to try their best to restore this concept art. After the on-site shooting was completed and entered post-production, everyone found that the photos taken on location in Norway could not achieve the same immersive atmosphere as the AI concept art. The team finally decided to directly use the AI-generated image as the background of the advertisement, and overlay the on-site shooting footage of the show's celebrities. Compared with the AI concept originally sold to the client, the real Norway was ironically "not Norwegian enough".
The time and energy saved in the ideation and presentation phases often translate into what we call alignment work: production and post-production personnel have to work hard to match the effects depicted in the AI concepts with the results that can be achieved in reality. AI has not made graphic designers redundant; instead, it has added a large amount of new alignment work. In some cases, the extra time spent on this part even exceeds the time that would be consumed if the designer had been involved from the very beginning of the project.
Actions That Managers Can Take
Generative AI can indeed speed up part of the creative ideation process, and provide creative professionals with a brand-new way to explore and visualize ideas. Repko van den Berg, head of the video and design team at RTL Nederland, compares generative AI to "having a brand-new palette to paint with". The key is to seize the efficiency dividend while avoiding paying high new costs.
The suggestions below are applicable to companies that have already implemented generative AI or are planning to introduce such tools. They help enterprises build or restructure their workflows in a more purposeful way: iterate ideas faster without prematurely solidifying the plan, while retaining the personnel and review nodes that can optimize the solution.
1. Identify the Right Position for AI to Intervene in the Workflow
Only when managers can see clearly how creative work actually operates can they have a better chance of converting short-term speed improvements into long-term benefits. Identifying these changes is not simple. Managers can first clarify who is responsible for writing AI prompts, to prevent this work from being concentrated in the hands of a few people. They should create opportunities for personnel who would otherwise be involved throughout the entire production process to continue participating in the project, ensuring that there is always discussion and critical feedback during the evolution of creativity. Require the team to proactively expose various preset assumptions and real-world constraints that are prematurely embedded by AI concepts. Set review nodes together with clients, reserve space to question the plan, and proactively carry out expectation management.
2. Let Downstream Experts Participate in the Early Stage of Work in Advance
Facing the phenomenon of workflow collapse, if you want to preserve the efficiency brought by AI, you must avoid the losses caused by rework and out-of-control expectations. Just like the case at the beginning where the photographer found that the lighting in the image violated reality brought inspiration to RTL: creative directors, graphic designers, photographers, and video directors should all join in at the earliest stage of creativity. With experts intervening in advance to complete collaboration and feasibility verification before presenting the plan to clients, enterprises can use generative AI in a more responsible and more efficient manner.
3. Emphasize the "Tentative Nature" of the Solution
AI causes clients to lock in the plan prematurely, which changes the way creative professionals manage client expectations. At RTL, real-world feasibility verification has become a core part of presentations. When creative staff show AI drafts to clients, they no longer just render beautiful visions, but proactively manage expectations, clearly explaining which elements can be implemented in reality and which cannot. In the past, clients needed to imagine the effects that had not yet been presented; now, on the contrary, they need to be reminded that many of the polished images they see cannot be delivered as they are in reality.
AI creates the illusion that the plan is already finished, so it is necessary to protect the tentative nature of the concept draft — it is only a temporary reference for the complete creativity. The team can operate in this way: present multiple creative directions at the same time, instead of only showing one polished final draft; clearly distinguish which elements are deliberately designed and which are only randomly generated and filled by AI; leave some decision points unresolved directly.
During review discussions, the focus should first be on the underlying creativity itself, discussing what is feasible, what is insufficient, and what can be adjusted, before delving into the execution-level details. The goal is: even if AI has generated a seemingly complete image, try to ensure that this concept retains room for interpretation and can continue to be iterated and improved.
You can also add a disclaimer on the early concept draft to inform the client that the solution is generated by AI and does not represent the effect of the final finished product. Personnel operating AI should further explain which details are core demands and which are only accidentally generated by AI. Pull the focus of communication back to the underlying creativity, and promote further polishing of the solution.
4. Set Up Client Review Nodes to Make Up for the Disappeared Handover Links
Traditional workflows are built on repeated client confirmation nodes. When the traditional workflow collapses, the collaboration model with clients also needs to evolve accordingly. Especially after generative AI realizes rapid concept generation, it is necessary to add regular client review nodes to replace the original disappeared handover steps.
Marco Heesen, head of the in-house advertising agency Ad Alliance under RTL Nederland, describes the production model in the AI era as requiring creative teams to align with clients much more frequently: "The biggest difference is that many problems in traditional projects are left to be solved in post-production. Now a lot of work has to be moved forward to the early preparation stage." He mentioned that teams need to invite clients and relevant stakeholders to participate more frequently, and advance the work in an iterative mode, blurring the originally clear boundaries between the early preparation phase and the formal production phase. "While building the plan, keep asking the client, 'Is this part okay?' 'Yes.' 'Then let's move to the next step, we will make the animation, are you satisfied with this version of the animation?' At each node, the team needs to distinguish which content can still be modified and which is finalized. 'Okay, record the confirmation, this part will not be changed anymore.'"
The review nodes where clients and teams jointly finalize details are not only applicable to the creative industry. For example, when product teams use AI prototypes, they also need to clarify with clients when the prototype will exit the exploration phase, and when a certain detail will be formally set as the implementation standard.
There are also several points worth noting here. If the team is small, already uses an iterative working mode, and has flexible boundaries for job responsibilities, the impact brought by workflow collapse will be relatively smaller. The level of risk depends on the gap between the presented concept and the final finished product; when a polished AI concept needs to be recreated under real-world constraints in subsequent steps, the risk reaches its highest level.
These suggestions themselves will also bring extra workload. Adding review nodes and bringing experts in earlier are easier to implement in long-term stable client cooperation projects. But the scenario of competitive bidding presentations is different: a seemingly complete and polished concept draft itself is more persuasive. In such scenarios, managers need to accept that some details will be locked in prematurely, and plan for the subsequent alignment work that needs to be carried out in advance.
Finally, the specific links skipped and the derived problems vary from industry to industry. But all industries have one thing in common: the formal process on the surface looks no different from the past, but the actual underlying working mode has been completely changed.
Enterprises that truly make good use of generative AI understand that even if the output of AI seems to be a final draft, they still need to leave room for diverse interpretations of creative ideas. The ultimate goal is to leave the right of choice to human judgment, collective discussion, and repeated revision.
Jana Retkowsky, Ella Hafermalz, Marleen Huysman, Daan Odijk | Text
Jana Retkowsky is an assistant professor at Rotterdam School of Management, Erasmus University. Ella Hafermalz is an associate professor at the KIN Center for Digital Innovation, Vrije Universiteit Amsterdam. Marleen Huysman is the founding director of the KIN Center for Digital Innovation, Vrije Universiteit Amsterdam. Daan Odijk is head of the Data & AI department at RTL Nederland.
This article is from the WeChat official account "Harvard Business Review" (ID: hbrchinese), author: HBR-China, 36Kr published it with authorization.