AI Transformed Software Development by Fixing Its Data Problem First. E-Commerce Is Next.
Every AI tool e-commerce operators are excited about right now, a shopping assistant, a recommendation engine, an agentic commerce integration, depends on something that has nothing to do with the tool itself: whether your product data is clean enough to feed it. I spent years in software development watching AI change how code gets written, and the lesson from that shift is not about speed. It is about precedent. AI only performs as well as the structured product data input sitting underneath it, and in e-commerce, that input is your product data.
What Actually Happened When Software Development Adopted AI
McKinsey's research on generative AI and developer productivity found developers documenting code for maintainability and writing new code in half the time, and refactoring existing code in two-thirds the time. McKinsey's own framing is that with the right upskilling and enterprise support, these speed gains can turn into productivity increases that outperform past advances in engineering tooling and process.
But the same research draws a hard line around those numbers.
- Gains on high-complexity work dropped to less than 10%, far below than previous figures.
- Developers with under a year of experience sometimes took 7 to 10% longer with AI assistance.
- The lift showed up on routine, well-scoped tasks: autocompleting functions, updating existing code, starting from a clear spec.
- On ambiguous, context-heavy problems, the model had nothing solid to work from, and the advantage shrank or disappeared.
A separate and more sobering data point comes from METR's randomized controlled trial with experienced open-source developers. Given real tasks on codebases they already knew, developers using AI tools took 19% longer to finish, not faster. Afterward, those same developers estimated that AI had sped them up by about 20%. The perception and the result pointed in opposite directions.
Put those two findings together and one pattern holds across both: AI's payoff in software development was never a flat multiplier. It scaled with how structured, well-scoped, and well-documented the underlying work already was.
Good Data Is the Precedent, Not a Nice-to-Have
This is the part of the software story that maps onto e-commerce directly, and it is easy to miss if you only look at the speed numbers.
Every AI tool that helped software teams sat on top of something else: clean repositories, defined interfaces, documented dependencies. The AI did not create that foundation. It required it already be there. Teams that had it got tasks done in half the time or less. Teams that did not, got the METR result: slower work and a false sense of progress.
E-commerce is walking into the same structure, one layer removed.
- An AI shopping assistant only recommends products it can understand. It cannot understand a product with a missing product attribute or a vague title.
- A recommendation engine only surfaces relevant products when the underlying attributes and categories are consistent across the catalog.
- Agentic commerce integrations, the kind now live on Shopify, Amazon, and Google, only convert at their advertised rate when they run on AI ready product data.
- A customer-facing chatbot answering product questions is only as accurate as the data behind its answers.
None of these tools fix product data quality as a side effect of doing their job. They assume it is already fixed. Good product data is not an input among many for these systems. It is the precedent that has to exist before any of them deliver what they promise.
Why Product Data Is E-Commerce's Version of a Messy Codebase
Every operator managing a large catalog already knows this problem by feel, even without the software analogy. It comes down to four structural issues.
- No shared standards. Every supplier, brand, and merchant defines product data differently: which attributes to use, which are required versus optional, how titles and descriptions should read, which image counts as the main one.
- Too many sources, too many formats. Data arrives from suppliers, distributors, brands, manufacturers, and marketplace sellers, and each one hands you its own version of messy supplier data.
- Incomplete or incorrect on arrival. A meaningful share of the data merchants receive is missing product attributes fields outright or simply wrong, which shows up downstream as missing product attributes on the live listing.
- Every channel adds its own rulebook. Most merchants sell across multiple channels, and each one enforces a different definition of what counts as acceptable data.
This is the reason most AI shopping tools underdeliver for operators who adopt them too early. The tool was never broken. The precedent was missing.
How Most Merchants Are Handling It, and Why It Stops Scaling
Three responses show up over and over, and each one runs out of road at a different point.
- Manual cleanup. Works at a few hundred SKUs with one or two people dedicated to it. Falls apart the moment you need to scale your product catalog past a few hundred SKUs, more suppliers, or more channels.
- Custom integrations with major suppliers or channels. Solves the problem for your biggest partners and leaves everyone else, often the majority of your catalog, exactly where it started.
- Ignoring it. The most common response by volume. The cost does not disappear, it just moves downstream into lower conversion, more returns, and every AI tool built on top of the catalog underperforming.
None of these are strategies. They are ways of managing a problem that keeps growing faster than the response to it, and the cost compounds fast once a catalog crosses 1,000+ SKUs. Product Data Enrichment Improves the Top and Bottom Line breaks down exactly what that costs in revenue and operating expense.
Where Dyver Fits: the Layer Built to Start From Messy Data
This is also why Dyver was not built as another tool that assumes clean data going in. Most AI systems in e-commerce, search, recommendations, chatbots, and agentic commerce are built to consume AI-ready product data. None of them are built to produce it from the messy supplier data that exists in most catalogs.
Dyver was built specifically for that step, the one every other AI tool depends on but does not do itself.
- Discover finds the missing pieces: attributes, specs, and images scattered across supplier sites, PDFs, and web pages, wherever they actually live.
- Organize normalizes everything into structured product data, mapping every supplier's format and every channel's rulebook into a single internal standard.
- Create fills in what is still missing: descriptions, translations, and channel-specific formatting, in your brand voice, producing AI-native product content from day one.
The output is not just a cleaner spreadsheet. It is the structured layer that every downstream AI tool needs to actually deliver on what it promises, whether that is a shopping assistant, a recommendation engine, or an AI-powered search integration. Dyver does not compete with those tools. It is the precedent that makes them work as advertised.
Takeaways for E-Commerce
- Every AI tool e-commerce operators are adopting, search, recommendations, chatbots, agentic commerce, depends on structured product data existing before it runs. None of them create that structure themselves.
- Software development already proved this pattern. Tasks got done in half the time or less where the underlying work was structured, and that reversed into a measured 19% slowdown, per METR, where it was not.
- Product data is e-commerce's version of a messy codebase: no shared standards, too many source formats, missing product attributes, and a different rulebook per channel.
- Manual cleanup, custom integrations, and ignoring the problem all describe how merchants cope today. None of them scale past a few hundred SKUs or a handful of channels.
- Adopting a customer-facing AI tool before fixing product data quality does not skip the problem. It just moves the failure downstream to where customers can see it.
- Dyver exists specifically to take on AI product data enrichment, turning messy supplier data into the structured layer every other AI investment depends on.
Software development already ran this experiment and published the results. E-commerce does not need to repeat every mistake to get to the same outcome. It needs the precedent fixed first, before the next AI tool goes live on top of it.
See how Dyver builds the data layer for every AI tool in your stack. →