A few weeks ago, while developing a system for designing presentations with artificial intelligence, I encountered a problem that initially appeared too small to deserve much attention.
An AI presentation tool had added a short explanatory section to a slide. The addition was coherent, visually appropriate, and consistent with the general subject. Under ordinary circumstances, it might have been considered an improvement.
But I had not asked for it.
The presentation concerned the possibility of introducing a four-day workweek. Its purpose was exploratory: to organize the questions management needed to consider, without recommending a decision or presenting unsupported conclusions. The additional section introduced an interpretation that had not been approved in the preceding design process.
The problem was not whether the sentence sounded reasonable. It was whether the tool had the authority to introduce it.
This distinction became increasingly important as the development continued.
We often evaluate AI-generated presentations by examining the final slides. Are they attractive? Is the information organized? Does the narrative appear coherent? Yet these questions tell us relatively little about the decisions made during generation.
A presentation may look professionally designed while containing interpretations, priorities, or assumptions that its author never intended.
I began to think that the difficulty was not primarily a matter of improving the prompt. It concerned the architecture of the process itself.
The problem with allowing generation and decision-making to become one operation
When someone asks an AI tool to prepare a presentation, the instruction is usually incomplete. It may identify the subject, audience, objective, and preferred style, but it rarely specifies every decision required to produce the final slides.
The system must therefore make choices.
It determines which subjects deserve separate slides, how information should be grouped, which statements should be emphasized, whether a diagram is more appropriate than text, and how the visual arrangement should direct the audience's attention.
Some of these decisions are necessary for implementation. Others can alter the meaning of the material.
This is where I believe an important boundary exists.
Choosing the position of a text block is not equivalent to introducing a new argument. Adjusting spacing is not equivalent to changing the importance of a conclusion. Selecting a visual representation is not necessarily a neutral operation when that representation can suggest certainty, causality, or comparison.
If we treat all these actions as ordinary design decisions, we may overlook the point at which the system moves from executing the user's intentions to interpreting them on the user's behalf.
The distinction is particularly important in professional presentations, where the audience may interpret visual emphasis as an indication of evidential strength or organizational priority.
A polished slide can make an uncertain statement appear settled.
The question, then, is not how to prevent AI from making decisions altogether. That would make generative design impractical. The question is how to distinguish decisions that the system may make independently from those that require explicit authorization.
Designing a system around decision boundaries
This question shaped the development of AI Presentation Designer — Professional.
The project began as an attempt to improve AI-assisted presentation design through structured instructions. As development progressed, however, the work increasingly concerned the relationships between different design responsibilities.
I found that presentation structure, information hierarchy, layout, visual assets, and final composition could not be treated as independent activities.
A change in structure could affect the argument. A change in hierarchy could alter the audience's interpretation. A visual asset could introduce meaning that was absent from the source material.
The challenge was therefore to coordinate these responsibilities without allowing one to assume the authority of another.
The system was developed as a coordinated prompt-based architecture rather than a conventional software application. Its components are implemented through structured instructions, each assigned a specific responsibility within the presentation-design process. The instructions do not operate as isolated requests for generating slides; they form a connected workflow in which the output of one stage becomes a controlled input to another.
What distinguishes this approach from a simple sequence of prompts is the attention given to decision authority and the preservation of approved requirements. Each stage is expected to work within its assigned responsibility, while subsequent stages must not silently reinterpret earlier decisions. These boundaries are expressed through the instruction system and examined through review, rather than enforced by a deterministic software engine.
The resulting architecture organizes the design process into eleven specialized modules, supported by an execution workflow and a review stage.
These modules are not intended to operate as eleven unrelated prompts. Their value lies in the boundaries established between them and in the decisions transferred from one stage to the next.
For example, the component responsible for slide structure may organize authorized information, but it should not manufacture additional content to achieve a preferred slide count.
Similarly, the component responsible for layout may determine how content is presented, but it should not treat a familiar design pattern as evidence that a particular conclusion belongs in the presentation.
This distinction may appear restrictive. In practice, it addresses a recurring problem in generative workflows: an implementation choice gradually becomes an assumed requirement, even though no one explicitly authorized it.
Why incomplete information must remain incomplete
One of the more interesting problems during development concerned missing information.
Generative models are particularly effective at producing coherent text. When information is incomplete, they can often construct a plausible continuation.
In some creative contexts, this is desirable. In a professional decision-making context, it may be precisely what we need to prevent.
Consider a presentation evaluating a four-day workweek.
If the organization has supplied no productivity measurements, the presentation can identify productivity as a relevant consideration. It cannot legitimately report that productivity will improve, decline, or remain unchanged.
The absence of evidence is not an invitation to complete the argument.
The same principle applies to design.
A chart may be visually attractive, but its presence can imply that measurable evidence exists. A prominent concluding statement may suggest that the analysis has reached a recommendation, even when the original purpose was only to explore alternatives.
During development, this led to a broader requirement: uncertainty must be preserved not only in the text but also in the way information is structured and displayed.
A system that avoids inventing numbers but visually presents unresolved questions as established findings has not fully preserved the meaning of its source material.
What an end-to-end test revealed
To examine the workflow, I used an exploratory management presentation concerning a four-day workweek.
The initial brief specified the purpose and audience but did not provide organizational productivity data or customer-coverage measurements.
The design process produced a sequence of decisions about structure, visual language, layout, information hierarchy, and other presentation elements. These decisions were then used to guide implementation through an AI presentation tool.
The test exposed two useful problems.
The first appeared during structural development. An approximate target of eight slides encouraged unnecessary expansion, including material that existed primarily to satisfy the expected length.
This revealed a conflict between a numerical preference and the actual informational requirements of the presentation.
The workflow was revised so that an approximate slide count would not be treated as a requirement to manufacture content. The final structure contained six substantive slides.
The second problem appeared after implementation.
The presentation tool introduced additional explanatory material, including a section labeled “Starting Point.” The content was plausible, but it had not been authorized by the preceding design decisions.
The review identified the deviation, and the presentation was corrected.
What interested me was that the additional material did not look like an obvious error. It appeared to improve the presentation.
Had the review focused only on visual quality or narrative coherence, the deviation might have remained unnoticed.
The test was limited to an architectural development scenario. It was not a controlled study comparing different presentation systems, and it cannot establish a general performance advantage.
Its value was more specific: it exposed where decision boundaries could fail and allowed the workflow to be revised accordingly.
The importance of separating execution from verification
A further architectural decision concerned the role of the presentation-generation tool itself.
I did not want the system to depend entirely on one presentation platform, nor did I want the host tool's existing preferences to determine the meaning of the presentation.
The architecture was not developed as an autonomous AI agent, nor was it implemented as a software engine tied to a particular presentation platform. It was designed as a coordinated system of structured prompts that can guide compatible AI models through a sequence of defined design responsibilities.
During development, ChatGPT was used to process the decision-oriented stages, while Gamma was used to implement the resulting presentation. This arrangement allowed the decision workflow and the presentation-generation tool to be examined as distinct components.
Although the architecture is intended to be adaptable across compatible models and tools, broader interoperability remains a subject for further testing.
This separation does not mean that the host tool becomes passive. It still performs essential work, including rendering slides, arranging elements, and producing the presentation artifact.
But implementation should not automatically confer authority to change the approved requirements.
The resulting presentation must also be examined as an artifact in its own right.
A correct instruction does not prove that it was correctly implemented.
This is why the workflow includes verification after execution. The review is concerned not only with whether the slides appear consistent, but also with whether their content, hierarchy, and composition remain faithful to the decisions established earlier.
It is possible for a presentation to be internally consistent and still depart from the original brief.
That possibility makes verification a separate engineering responsibility rather than a final cosmetic inspection.
A question extending beyond presentation design
I do not consider this architecture a complete solution to the broader problem of controlling generative AI.
Its current implementation has been developed and examined within a specific presentation-design workflow. Broader claims would require additional testing across tasks, tools, and users.
Nevertheless, the development raised a question that extends beyond presentations.
As generative systems become capable of producing increasingly complete artifacts, how do we preserve the distinction between what the user intended, what the system inferred, and what the system actually implemented?
This question is relevant wherever AI is entrusted with work that involves interpretation rather than mechanical execution.
It is tempting to assume that greater generative capability will reduce the need for explicit design decisions. My experience during this project suggests a different possibility: the more work a system can complete independently, the more important it becomes to understand the boundaries of that independence.
I am interested in AI systems that can do more than produce convincing outputs. I am interested in systems whose decisions can be examined, whose assumptions can be questioned, and whose implementation can be evaluated against the intentions that authorized it.
The development of this presentation-design system has been one practical exploration of that problem.
And it has left me with a question that I believe deserves further engineering attention:
When an AI-generated artifact appears complete, how can we determine whether it is complete according to the user's requirements—or merely according to the system's own interpretation of them?
This article draws on the development of AI Presentation Designer — Professional, an independent project exploring decision authority, structured prompt architectures, and verification in AI-assisted presentation creation.
AI Assistance Disclosure
This article is based on my original development work, architectural decisions, and practical testing experience. I used ChatGPT to assist with English drafting, technical discussion, and editorial refinement. I reviewed and approved the final manuscript and take responsibility for its content, accuracy, and conclusions.