Business Case Study Flagged as AI? SWOT, Five Forces and Two Kinds of Sameness

When a whole cohort hands in documents that look alike, the first worry is usually plagiarism. That reflex points at the wrong report. You are probably holding two numbers, and they are not measuring the same thing.

A framework does two things to your writing. It hands you a fixed vocabulary in a fixed order, which is why your paragraphs look like your classmates' paragraphs. And it turns your report into a set of parallel slots, which is why your own five paragraphs look like each other.

Only the second one is what an AI writing percentage is built to react to. Turnitin's published description of AI scoring is a per-submission process: the file is cut into overlapping segments, each segment gets a probability, sentence scores are pooled and aggregated. Nothing in it compares your document with anyone else's. The comparison against other work is the other number, the similarity score, which is defined against "Turnitin's comprehensive collection of content", and the vendor states the two are "completely independent and do not influence each other".

There is a third thing worth knowing before you touch anything, and it is the one that surprises people: in a framework-heavy case study, most of what you wrote is not being scored at all. The percentage is calculated over prose sentences. By that definition a SWOT grid, a bulleted PESTEL and a list of five-force headings are not prose sentences.

Worth separating these before deciding anything, because students routinely bring the first worry to the second report.

Two details on the similarity side that are specific to framework language. First, framework labels are short. Turnitin's exclusion-filter guide documents a small-match setting: "By default, the threshold is 8 words, meaning that only matches 8 words or longer appear in the Similarity Report. You can increase this number, but 8 words is the minimum allowed value." Count your own framework strings against that. "Threat of new entrants" is four words. "Bargaining power of buyers" is four. "Strengths, weaknesses, opportunities and threats" is five. The setting is not available in the classic version of the report, and it is not yours to switch either way. Eight words is the default once somebody turns the filter on, not a floor that applies whether or not anyone did — but it still tells you what length of match the vendor treats as noise.

Second, on the AI side, the model is not looking for shared phrases at all. The FAQ says the classifiers "are trained to detect these differences in word probability and are adept at the particular word probability sequences of human writers", and elsewhere that the model "is not explicitly programmed to evaluate specific signals such as 'burstiness,' 'perplexity,' or other individual metrics sometimes referenced in public discussions". Turnitin AI Writing Report vs Similarity Report has the full split if you have both numbers in front of you.

The FAQ is explicit about what gets analysed:

"This qualifying text includes only prose sentences, meaning that we only analyze blocks of text that are written in standard grammatical sentences and do not include other types of writing such as lists, bullet points (short non-sentence structures), or other non-sentence structures."

And about what the percentage is a percentage of:

"As noted previously, this percentage is not necessarily the percentage of the entire submission. If text within the submission is not considered long-form prose text, it will not be included."

Now walk through a normal MBA case report and mark what survives that definition.

That table has two consequences, and the number on your screen shows you neither.

Your percentage and your classmate's are not comparable if you laid the same analysis out differently. Same PESTEL content, one of you in bullets and one of you in full sentences inside a table, and you are being scored over different amounts of text.

And the qualifying-prose remainder in a grid-heavy report can be small. That matters because of what the FAQ says about small amounts of text: "In shorter documents where there are only a few hundred words, the prediction will be mostly 'all or nothing' because we're predicting on a single segment without the opportunity to overlap." Below 300 words of prose text, an AI writing report is not produced at all. If your 3,000-word case study is mostly exhibits and bullets, the percentage you are looking at may have been computed over the linking commentary and little else. What Content Does Turnitin AI Skip? covers the exclusions in general; the point here is that framework layouts change the denominator more than almost any other assignment format.

Turnitin publishes a description of the text its false positives land on:

"Sometimes false positives (incorrectly flagging human-written text as AI-generated), can include content without a lot of structural variation, text that literally repeats itself, or text that has been paraphrased without developing new ideas."

Framework prose reaches that description honestly, and for reasons that have nothing to do with the writer.

The vocabulary is fixed by the source. Harvard Business School's Institute for Strategy and Competitiveness publishes the five forces under exactly these names: Threat of New Entrants, Bargaining Power of Suppliers, Bargaining Power of Buyers, Threat of Substitute Products or Services, Rivalry Among Existing Competitors. Michael Porter first set them out in a 1979 Harvard Business Review article. Those strings are not yours to vary, and no marker wants you to.

The order is convention rather than doctrine. The ISC page itself lists them substitutes first and new entrants fourth, which is not the order in the paragraph above. What is fixed is the sentence that ends each block. The characteristic move is a verdict in a fixed frame: force, rating, one clause of justification. Do that five times and you have written five sentences with identical structure and swapped nouns, which is a fair description of "content without a lot of structural variation".

I also checked whether any of this is documented as a special case.

What I counted. On 27 August 2026 I read the rendered body text of the three Turnitin pages that define the AI Writing Report: Using the AI Writing Report (7,218 characters), the AI writing detection capabilities FAQ (31,712), and the AI writing detection model release notes (9,957). Searched across all three: `SWOT` 0 · `PESTEL` 0 · `PESTLE` 0 · `Porter` 0 · `case study` 0 · `business` 0 · `marketing` 0 · `discipline` 0 · `genre` 0. The single hit for `framework` is the vendor describing its own testing framework. Sensitivity check on the same text: `prose` 13 · `qualifying` 15 · `bibliograph` 4. `subject area` returns three times, all of them on the FAQ page and all of them describing the training sample: once as "academic writing across geographies and subject areas", then twice in a repeated list of "less common subject areas such as anthropology, geology, sociology, and others".

So there is no business-school carve-out, in either direction. The model does not know a case study from a lab report, and the documentation does not claim it does.

The framework part of a case study is supposed to be interchangeable. That is the whole point of a framework. The analysis is supposed to be the part only you could have written, and in most reports the two are sitting in the same paragraph, in the same register, which is what makes the document uniform.

Concrete moves that separate them:

1. Put the verdict and the reasoning in different sentence shapes. If the rating sentence is a fixed frame, let it be one, and then do not write the next sentence in a fixed frame too. A short flat rating line followed by a longer argumentative paragraph reads differently from five identical hybrids.
2. Name the number, the date and the source. A sentence that pins supplier power to the two named contract manufacturers in your own Exhibit 4, and to the year the second one was acquired, carries information that "supplier power is high due to industry concentration" does not. The second sentence can be written about any industry by anyone who has read nothing.
3. Say where the framework fails this company. A paragraph explaining that the substitutes force is close to meaningless in a regulated market with a statutory monopoly is analysis. Rating it "low" and moving on is form-filling.
4. Drop a box and justify the omission in one sentence. Uniform-length treatment of forces that are not equally important is exactly the flat structure the vendor describes, and it spends the same wordage on a force worth a sentence as on one worth a page.
5. Write the recommendation before the frameworks, then check whether they support it. Reports assembled framework-first tend to end with a conclusion that restates the boxes. A 2023 Turnitin release note observed false positives concentrated in the first and last sentences of documents, "written in a generic way", and the detection logic was changed to reduce that; the underlying writing habit did not go anywhere. Why Intros and Conclusions Get Flagged as AI has the release note in full.

I am not telling you what a percentage does afterwards, and nobody honest will. What the list above describes is which paragraphs in a case study are yours and which belong to the framework, and that line is worth drawing whether or not a detector is involved.

Assume the report is back and something has to change. The order matters, because a case study is one of the easiest documents to damage by rewriting.

1. Read the highlights, not the percentage. Where the highlighting falls tells you whether the model reacted to your linking commentary or to your recommendation. The number does not.
2. Leave the framework labels alone. They are terms of art. A marker checking whether you applied Porter correctly wants to see the five names.
3. Leave the exhibits, ratios and figures alone. They are not qualifying prose and they are the part of the document where an error is fatal.
4. Start with the analysis paragraphs. They are the part a rewrite can legitimately touch, and if the report came out uniform they are also the part that was thin to begin with.
5. Re-proofread everything you touched. Every paragraph that changes is a paragraph you now have to check against your own exhibit numbers. That cost, not the rewriting, is what makes whole-document passes expensive.

Point five is the reason scope matters more than anything else in a tool. HumanPen works on the file rather than on a pasted excerpt: a DOCX goes up and an editable DOCX comes back, with the structure, the terminology, the citations, the layout and the styles left as they were. For a report full of tables and exhibits that is the difference between one job and an afternoon of re-formatting. Either mark the passages yourself, or point it at the AI report you already have from Turnitin or iThenticate and let it pick out the flagged ones; either way it rewrites those and leaves everything else alone. One paragraph is the smallest thing it will touch, and the expanded selection is put in front of you to approve first, so you know exactly how many paragraphs you have signed up to re-read. When eligible, you can keep reducing AI for free.

Thirty of us wrote the same SWOT. Will that show up as AI? Not through that mechanism. The AI writing percentage is described as a per-submission classification of your own text into overlapping segments, not a comparison with other submissions, and the FAQ states there is "no separate repository for AI writing detection". Text shared with other work is the similarity side, which is a different number with its own filters.

Does the SWOT grid itself get scored? Bullets and other non-sentence structures are documented as outside the analysis. Full prose sentences inside a table are a different case: an August 2023 release note says long-form prose in tables is now processed, with existing submissions needing resubmission to be reprocessed. So the same PESTEL can be in or out of the denominator depending on how you laid it out.

My case study came back at 100%. It is only 900 words of actual prose. That combination is worth reading against the FAQ's own caveat about short text: "In shorter documents where there are only a few hundred words, the prediction will be mostly 'all or nothing' because we're predicting on a single segment without the opportunity to overlap." A grid-heavy report can have far less qualifying prose than its total word count suggests. Short Documents and the All-or-Nothing Problem in Turnitin explains the segment mechanics.

Should I stop using frameworks? No, and your marking criteria probably require them. What you can change is the ratio: less space spent labelling the boxes, more spent on the judgement the boxes were supposed to support. That is a better report as well as a less uniform one.

Can I just take the percentage to my tutor and argue it is a false positive? You can bring it up, and the vendor's own wording helps. Its guide to using the AI writing report says the indicator "may not always be accurate" and "should not be used as the sole basis for adverse actions against a student". Its detection FAQ, in the passage about false positives, advises taking structurally uniform text "into consideration when looking at the percentage indicated". None of that is proof of who wrote the document. What a 1% false positive rate means when a university submits 75,000 papers is the honest version of the arithmetic.