Nell'era AI il tuo prodotto non è così facile da copiare come credi

Salvatore Sanfilippo argues that AI-assisted programming makes implementation cheaper without automatically supplying a good product idea. Using his computer-aided design (CAD) prototype as an example, he distinguishes building useful software from copying an already successful product: the first still depends on judgment, while the second may become cheaper.

This English summary is editorially translated from the video's Italian auto-generated captions, without an independent audio or visual review. The prototype results are his account; market opportunities and future capabilities are proposals or opinions, not independently verified findings.

Key points

• A fast prototype is not a finished product. Sanfilippo describes spending two or three exploratory days on a CAD proof of concept, but says turning it into a distinctive product would take months of development and careful refinement of its interactions. He proposes a niche between Tinkercad and Fusion, with a permanently usable purchased version and paid later upgrades rather than mandatory ongoing subscription access. The prices and periods of included updates he mentions are illustrative, and he explicitly says pricing needs research. 00:01–04:02(opens in a new tab)

• Differentiation comes from accumulated choices, not just the ability to generate code. His examples include clearly labeled controls, responsiveness, mathematical trade-offs, exported mesh quality, supported file formats, and ways to express geometric constraints without reproducing a full commercial CAD system. Together, these choices determine how the product feels and what it helps users do. He argues that someone who has used CAD extensively and knows what they wish were different can supply the direction that access to the same coding agent does not supply by itself. 04:02–07:56(opens in a new tab)

• Lower implementation costs do weaken one barrier to imitation. Previously, a competitor had to pay substantial engineering costs to reproduce an established product's interactions while competing with something already known. Sanfilippo argues that a sufficiently capable agent could reduce that cost, including the cost of copying interactions. His distinction is that choosing a worthwhile product direction remains difficult even if reproducing the result becomes easier. He expects copying to matter once a product has gained some recognition; that is his competitive assessment, not a guarantee of protection. His accompanying copyright discussion is tentative and should not be treated as an established legal constraint on competitors. 07:56–11:06(opens in a new tab)

• AI inside a product can still require deliberate interface design. He considers a separate challenge: whether people would need CAD at all if AI designed everything. Rather than assume complete delegation, he proposes an integrated workflow in which users select a particular part and specify how it should change, while the product provides the agent with tools suited to inspecting the object. This differs from asking a generic agent to produce an entire part: the product would organize human direction and local edits. He presents this as an opportunity for product design, not a demonstrated system or proof that generic agents cannot perform such edits. 11:06–12:08(opens in a new tab)

• Technical refinement still matters when an agent writes the code. His closing argument is not that algorithms become irrelevant: the step-by-step choice and improvement of algorithms and interactions still shape the result. AI assistance may let someone realize a product idea without already knowing every relevant programming specialty, just as strong programming skills alone never guaranteed an idea worth turning into a product. 12:08–13:10(opens in a new tab)

Full video: https://www.youtube.com/watch?v=AZCMN2_OCfE(opens in a new tab)