AI can now solve the hardest problems in mathematics. Can it come up with a good business plan?
Business is harder in one respect: a computer can check a proof, but no procedure can run a business plan and report whether it will succeed. So how would AI know its ideas are getting better?
My new paper, forthcoming in the Strategic Entrepreneurship Journal, proposes a method for this, which I call conceptual search: AI generates venture concepts, an AI judge compares them, the most promising get refined, and the loop repeats.
It builds on two earlier findings:
- AI can write business plans that experienced investors rate at least as favorably as the entrepreneurs’ own (with Harsh Ketkar and Hyunjin Kim).
- AI can predict which ventures will raise more money better than human experts, as shown in a prospective tournament with 30 live Kickstarter ventures (with Aticus Peterson and Daniel Wilde).
The judge in the loop is one of the models from that tournament. It isn’t the market, but it has a track record of predicting it.
In a prototype using those same 30 ventures:
- Local refinement (propose five small changes, keep the best, repeat) moved a mid-ranked venture from 16th to 3rd in the judge’s ranking. The winning changes sharpened the plan: a narrower launch market, a small user test, a proposed price. A fidelity check kept the AI from inventing customers or traction.
- Global search (a genetic algorithm that recombines parts of different ventures) reached concepts far from any starting point. After refinement, the best outranked 24 of the 30 originals, again in the judge’s eyes.
The whole run took three hours and cost under $20.
More important than any single result: the process of imagining ventures can now be designed, run, and audited, with every candidate and decision on record.
The paper also develops two broader ideas:
- A “generative view of entrepreneurial imagination”: the ideas people pursue depend on how they search for ideas before acting.
- A way to keep studying search in strategy, long modeled with abstract NK landscapes, using real businesses instead of bit strings.
So, can AI come up with good business plans? For entrepreneurs, “good” means at least two things: promising and new. Both can now be measured: promise by a judge validated against real outcomes, novelty by distance from existing ideas (in the paper, semantic distance from a pool of “seed” ideas). Run a program like this one and check whether its plans clear both bars. That turns a debate into an experiment.
Everything is open, so others can rerun it, change the design, and push it further.
—