Entrepreneurship instructors consistently report the same change this year: business plans look more polished, and market analysis sections have become longer and smoother, yet many of them convey very little.
ChatGPT is present in the classroom and has been since last winter. Banning it has not worked where it was attempted, and ignoring its presence is even less effective.
We teach the ASP Lancement d'une entreprise at the CFP Calixa-Lavallée. The program runs 12 weeks online, with two days per week of virtual classes and three days of coaching, and each student develops their own business project. A large share of the content consists of precisely what these tools produce most readily: business plans, market descriptions, mission statements and pitch outlines. This has required more reflection than anticipated. The approach described below reflects our current position, which will likely continue to evolve.
Why do ChatGPT and business plans make a difficult combination?
Because the value of a business plan lies in the thinking, not the document. The purpose of the assignment is the reasoning required to write it: Who is the customer? Why would they pay? What does it cost? What happens if the assumptions prove wrong?
ChatGPT can produce the document without any of that reasoning. Asked for a business plan for a café in a small town, it delivers a well-organized plan in under a minute, complete with an executive summary, a SWOT analysis and a marketing plan. It reads like a real plan and even includes figures, although no one, including the tool, knows where they came from.
This is the central risk in a course of this type. The output resembles competence. A learner can submit it, receive reasonable feedback on its presentation and never have considered whether anyone in that town wants another café.
The opposite is also true. These tools are genuinely effective for part of what entrepreneurs need: generating name ideas, rewriting an awkward paragraph, explaining the break-even point for the fifth time without impatience, or playing the role of a sceptical banker. Discarding all of this because of concerns about cheating would be a missed opportunity.
How should AI be used in developing a business plan?
To question the plan rather than to write it. The most sensible shift, in our view, is to change the AI's role: instead of "write my business plan," the instruction becomes "challenge my business plan."
Several prompts are well suited to this purpose:
- "Here is my business idea. Act as a loans officer at a bank and ask me the ten hardest questions you'd ask before lending me money."
- "Here is my Business Model Canvas. Which box looks weakest, and why?"
- "List the assumptions in this plan that, if wrong, would sink the business."
- "Who are three types of customers who would never buy this, and what would it take to change their mind?"
The learner must then provide the answers. The tool performs the role of a good mentor or a demanding investor, while the thinking remains with the learner. This approach is also much harder to circumvent: answers to questions about one's own idea cannot be supplied without understanding that idea.
The Business Model Canvas is well suited to this method. It is short, fits on one page, and each box requires a specific claim. Nine claims can easily be examined one at a time; a forty-page generated plan cannot.
Why must every source be verified?
Because these tools invent figures. When a plan states that "the local market for organic pet food is worth $4 million," the first question is where that figure came from. When it came from ChatGPT, there is usually no genuine answer: the number was generated because a number fit the sentence.
This has been well documented since the spring. These tools invent references, studies and statistics that appear entirely credible. We recommend a firm rule in any business course: every figure must have a source that a person can open and read, such as Statistics Canada, the Institut de la statistique du Québec, an industry association report, a municipal document or a survey conducted by the learner. Any figure that cannot be traced must be removed.
This is a sound practice for business owners as well as learners. We addressed the use of these tools in our Beginner's Guide to ChatGPT earlier this year, and verification is the point we emphasize most. The tool is a starting point for research, never a substitute for it.
What should be evaluated: the reasoning or the presentation?
The reasoning. This is the change we consider necessary on the evaluation side.
A clean, well-formatted plan was once at least a modest indicator of effort. It no longer is, since polish now costs nothing. Continuing to grade polish amounts to grading the tool.
The following elements still demonstrate genuine work:
- Explained choices. Why this price rather than another? Why this location?
- Evidence gathered in the field: interviews, a count of competitors on the main street, supplier quotes.
- Acknowledged weaknesses. A plan that states "this is the part I am least certain about, and here is how I would test it" shows more maturity than one in which everything appears flawless.
- The ability to discuss it. A ten-minute conversation about a plan reveals more than the plan itself.
Oral presentations and pitches now carry greater weight. The purpose is not to catch anyone out, but to develop a skill every entrepreneur needs: explaining and defending an idea aloud. Bankers do not read a plan and then transfer funds; they ask questions.
AI detectors do not offer a solution either. OpenAI discontinued its own AI text classifier in July because of its low accuracy. If the company that built ChatGPT cannot reliably distinguish AI-generated text, a free website claiming to do so warrants little confidence, and it should not serve as the basis for accusing anyone.
What guidance is the education sector providing?
Last month, UNESCO published its guidance for generative AI in education and research. It recommends a minimum age of 13 for the use of these tools in classrooms and places strong emphasis on protecting human agency, meaning that the learner continues to make the decisions. The guidance aligns closely with the approach we had been developing in business education.
Human agency is an apt summary of the purpose of an entrepreneurship course: training people to make decisions under uncertainty and to take responsibility for them. A tool that makes decisions on their behalf undermines that purpose. A tool that challenges their decisions and requires them to defend those decisions supports it.
No one has yet fully resolved these questions. Teachers are learning alongside everyone else, often a step behind. Businesses face the same situation with their own staff, which is why we anticipate strong demand for basic AI training for small teams.
In our assessment, within a few years the question will no longer be whether AI was used to prepare a business plan, but whether the person behind it can answer for it.