A SWOT analysis now takes about a minute with a chatbot, and the research says that speed makes the tool’s oldest problem worse. A 1997 study found that SWOT lists were long, vague and unchecked, and that nobody used them afterward. Generative AI can produce that kind of list faster than ever.
Still, a 2024 study found that investors rated AI-drafted strategy material well. Other studies show that one answer from one model is unreliable. Models agree with whoever is asking. They also push different users toward the same ideas.
The short version: use AI to widen a SWOT analysis, run it more than once, and ask it to argue against you. Then let people cut the list down, check each item against evidence, and turn what is left into actions. Every study is linked, and a table near the end shows which are peer reviewed.
Where SWOT analysis came from, and the problem it already had
A 2023 study in Long Range Planning traced it back to a planning method called SOFT. Its authors, Richard Puyt, Finn Birger Lie and Celeste Wilderom, used archives, expert interviews and a literature review. They credit Robert Franklin Stewart as the originator.
In the original method, managers listed 8 to 10 key planning issues for their unit. They sorted each one into four actions: safeguard strengths, open opportunities, fix faults and thwart threats. The label "SOFT" was later changed to SWOT. Groups then sent proposals to an executive committee. The original was a short list built to drive decisions, per the Puyt, Lie and Wilderom study.
Practice drifted far from that. Terry Hill and Roy Westbrook studied 50 companies in a UK government manufacturing program. More than 20 of them ran a SWOT with help from 14 consulting firms. The lists averaged more than 40 factors. Items were vague, not prioritized and not checked. According to the Oxford record of the 1997 Long Range Planning paper, none of the companies used the outputs later in their strategy work.
A 2021 review noted more than six decades of SWOT research and still found the shared understanding of SWOT unclear, per the Benzaghta et al. review.
What generative AI changes about a SWOT analysis
Ask any chatbot for a SWOT analysis of a mid-sized association and you get four tidy lists in seconds. Every item sounds reasonable.
That output looks a lot like what Hill and Westbrook found in 1997: long, general and unranked. The model has never seen your renewal rate, your budget or last year’s event attendance.
Used carelessly, AI makes the known failure of SWOT cheap and fast. Used well, it can help with the parts of strategy where it is strong.
What the research says AI does well in strategy work
It can draft strategy material that investors rate well. Felipe Csaszar, Harsh Ketkar and Hyunjin Kim ran two studies, published in Strategy Science in 2024. In the first, GPT-3.5 rewrote sections of real startup plans, and 250 evaluators with investment experience rated them blind. The AI versions scored higher than the founders’ originals, by 0.14 standard deviations on the authors’ index.
AI versions beat plans the accelerator had rejected. Against accepted plans, the advantage disappeared.
It can evaluate options in a way that tracks expert judgment. In the second study, GPT-3.5 scored 138 business plans from a business school competition. Its scores correlated with venture capital and angel judges at 0.52. The model’s scores tracked the judges’ average more closely than individual judges tracked each other, according to the Csaszar, Ketkar and Kim paper. The authors call their results preliminary.
It speeds up work that fits its strengths. A field experiment with 758 Boston Consulting Group consultants tested AI on realistic consulting tasks. For tasks inside what the authors call the "jagged technological frontier," consultants with AI finished 12.2% more tasks and worked 25.1% faster. Quality also rose. The study, led by Fabrizio Dell’Acqua, appeared in Organization Science in 2026.
Why one AI answer is not enough for a SWOT analysis
Single answers are unreliable. Anil Doshi, J. Jason Bell, Emil Mirzayev and Bart Vanneste tested whether AI can judge strategic options. Their 2025 paper in the Strategic Management Journal used seven models in one study and six in the other. The authors found that single AI evaluations "are often inconsistent or biased."
Changing the order of two options flipped many answers. Consistency for one model and prompt fell as low as 42%. One model picked the second option in 99.9% of comparisons. But when the authors pooled tens of thousands of evaluations, the AI rankings came to resemble expert rankings, per the Doshi et al. study. More runs helped more than more variety of models or prompts.
AI tends to agree with you. Mrinank Sharma and colleagues studied five AI assistants for a 2024 paper at the International Conference on Learning Representations (ICLR). They found sycophancy across all of them. Models tended to match a user’s stated views, sometimes over the correct answer. The authors call it "a general behavior" of models trained on human feedback, in the ICLR 2024 paper.
In a SWOT analysis, this hits the weaknesses box hardest. If the prompt says your member service is excellent, the model will likely list it as a strength.
Why AI SWOT lists start to look alike
Everyone gets the same list. Anil Doshi and Oliver Hauser asked 300 writers to write short stories, some with AI ideas. AI help made individual stories better on average. But stories written with one AI idea were 10.7% more similar to each other than stories written without AI. Their 2024 Science Advances paper calls this a loss of collective diversity.
Apply that to strategy. If five peer associations ask the same model for a SWOT, they get overlapping answers. A strength that every peer also lists does not set you apart.
AI can lead skilled people to wrong answers. The same BCG experiment included a task built to fall outside the AI’s abilities. There, consultants using AI did worse than those without it. In the 2023 working paper, they were 19 percentage points less likely to reach a correct answer.
A research-backed way to run a SWOT analysis with AI
Each part of the job should go to whoever does it best. The table maps each step to the research behind it.
| Step | Who leads | What to do | Research behind it |
|---|---|---|---|
| 1. Gather evidence | Your team | Pull your own numbers first: renewals, traffic, revenue. | Hill and Westbrook: items were never checked |
| 2. Widen the list | AI | Run the same prompt several times and merge the results. | Doshi et al.: pooled runs track experts, single runs do not |
| 3. Argue the other side | AI | Ask the model to attack each strength. | Sharma et al.: models lean toward your view |
| 4. Cut and rank | Your team | Keep three to five items per box, each with evidence. | Puyt et al.: the original used 8 to 10 issues in total |
| 5. Test for sameness | Your team | Strike any strength that every peer could also claim. | Doshi and Hauser: AI narrows ideas across users |
| 6. Turn it into actions | Your team, with AI drafting | Pair items into actions with an owner and a date. | Puyt et al.: SOFT was built to feed decisions |
For step 6, many teams use the TOWS matrix, which Heinz Weihrich published in Long Range Planning in 1982. It pairs each internal factor with an external one. Goals from this step should be specific and measurable, as covered in how to set SMART marketing goals.
Prompts that follow the research
These three prompts match steps 2, 3 and 5 of a SWOT analysis. Paste your evidence summary from step 1 above each one.
- Widen: "Using only the facts above, list possible strengths, weaknesses, opportunities and threats. For each item, name the fact it rests on. Mark any item that has no supporting fact."
- Argue: "Act as a skeptical board member. For each strength I listed, give the strongest case that it is not a real strength. Then list three weaknesses I did not mention."
- Test for sameness: "Which of these strengths could most organizations like ours also claim? Remove those and explain what is left."
Run each prompt more than once and compare. Longer answers are not always better, as covered in when more tokens make AI output worse.
AI itself now belongs in your SWOT analysis
For most agencies, nonprofits and associations, AI is also an item inside the SWOT analysis itself.
On the opportunity side, AI can cut time spent on drafting, as the BCG experiment showed. That depends on how tools are rolled out, which our post on what AI implementation means covers.
On the threat side, AI search and AI answers are changing how people find organizations online. Our marketing in the AI age post covers what shifted in 2026 for small teams. Those items belong in the same evidence test as everything else. "AI is a threat" is too vague to act on. "Search referrals to our resource pages fell this year" is something a team can plan around.
The research behind this post
| Source | Venue | Peer reviewed | What it found |
|---|---|---|---|
| Puyt, Lie and Wilderom (2023) | Long Range Planning 56(3) | Yes | SWOT began as SOFT: a short list of planning issues built to feed decisions |
| Hill and Westbrook (1997) | Long Range Planning 30(1) | Yes | Lists averaged over 40 vague, unranked, unchecked items, and no company used the outputs |
| Benzaghta et al. (2021) | Journal of Global Business Insights 6(1) | Yes | Six decades of research, with no clear shared understanding of SWOT |
| Csaszar, Ketkar and Kim (2024) | Strategy Science 9(4) | Yes | AI-written strategy sections rated well by investors, and AI scores tracked judges |
| Doshi, Bell, Mirzayev and Vanneste (2025) | Strategic Management Journal 46(3) | Yes | Single AI evaluations were inconsistent and biased, and pooled ones resembled experts |
| Sharma et al. (2024) | ICLR 2024 | Yes, conference review | Five AI assistants tended to agree with users’ stated views |
| Doshi and Hauser (2024) | Science Advances | Yes | AI ideas improved individual stories but made them more alike |
| Dell’Acqua et al. (2026) | Organization Science 37(2) | Yes | AI helped consultants on some tasks and hurt them on others |
Two limits apply. Several studies used older models such as GPT-3.5 and GPT-4, and newer models may score differently. None tested a SWOT analysis directly. They tested the pieces: generating options, judging them and model habits.
Frequently Asked Questions
Is SWOT analysis still worth doing?
Yes, if it ends in decisions. The 1997 Hill and Westbrook study found SWOT failed when lists were long, vague and never used. Keep it short and tie each item to an action.
Can ChatGPT or Claude do a SWOT analysis for me?
They can draft one quickly. Research shows a single AI answer is often inconsistent, and models tend to agree with how you frame the question. Use the draft to widen your list, then check each item against your own data.
How many times should I run a SWOT prompt?
More than once. Doshi and colleagues found that pooling many AI evaluations brought rankings closer to expert rankings, while single ones were unreliable. Three to five runs merged into one list is a practical start.
Why does an AI SWOT analysis look the same for everyone?
Studies show AI help tends to push different users toward similar ideas. In one Science Advances study, stories written with AI ideas were 10.7% more similar to each other. Your peers asking the same model will get overlapping lists, so generic strengths do not set you apart.
What is AI sycophancy, and why does it matter for SWOT?
Sycophancy is a model’s habit of agreeing with the user. A 2024 ICLR paper found it in five AI assistants. In a SWOT, it means the model may accept your strengths as given and go easy on weaknesses. Ask it directly to argue against you.
How many items should each SWOT box have?
Few enough to act on. The original SOFT method used 8 to 10 key issues per unit in total, while the failed SWOTs in the 1997 study averaged more than 40. Three to five items per box, each with evidence, is a workable target.
What is the TOWS matrix?
TOWS is a follow-on step published by Heinz Weihrich in 1982. It pairs internal factors with external ones, such as a strength with an opportunity, to turn a SWOT analysis into actions.
Are these studies peer reviewed?
All eight studies in the source table went through peer review. Seven are journal articles and one is an ICLR conference paper. The 19 percentage point figure comes from the 2023 working paper version of the BCG study.
