B2B e-commerce faces a $3 trillion data quality crisis. Product information is incomplete, inaccurate, and scattered across legacy systems. The question is whether generative AI can solve this problem or whether the technology introduces more risk than it resolves.
The scale of the data problem
Harvard Business Review estimated the cost of poor data quality at $3 trillion six years ago. The number has likely grown since. Survey data reveals the extent: 96% of companies lack proper data infrastructure, and only 3% meet basic quality standards. Gartner found that 40% of businesses fail to achieve their objectives due to missing, incomplete, or inaccurate data.
The consequences are real. When asked what makes B2B buyers go elsewhere, insufficient or inaccurate data tops the list alongside products that are unavailable or hard to find. Sixty percent of companies say they will switch suppliers if data needs are not met.
The generative AI moment
ChatGPT achieved the fastest technology adoption in human history, reaching 100 million users in two months. Google Search took 12 months. Netflix took 18 years. The technology is now embedded in commerce platforms, CRMs, and enterprise systems. Amazon invested in Anthropic. Salesforce, Commerce Tools, and others are building AI into their core offerings.
Use cases are emerging rapidly: product comparisons, installation instructions, customer service scripts, CRM prompts based on conversation analysis. One practitioner achieved 94% accuracy using optical character recognition to convert a table screenshot into HTML and Excel in 20 seconds versus 20 minutes manually.
The debate format
Recorded live at B2B Online in West Palm Beach, Team AI Not Ready featured Tim Lavender from Beacon Building Products and Sam Schwarz from Jansen Precast. Team AI Ready included Meeta Kratz from Lonza and Michael Schultz from commercetools. All brought hands-on experience implementing AI in B2B environments.
Round 1: Accuracy and completeness
Meeta Kratz argued for iteration over perfection. Nothing is perfect at launch, whether electric vehicles or data. AI is about learning and improving over time. Waiting sacrifices opportunity cost. Predictive analytics can identify customer needs and enable proactive sales outreach.
Do not sacrifice speed for perfection. AI is about learning, iterating, and reestablishing the norm. We have to come out of the gate strong. Why wait?
Meeta Kratz, Lonza
Michael Schultz distinguished AI from machine learning. Machine learning depends on data quality because it trains machines to interpret and analyze. AI is broader, creating intelligent machines with human-like capabilities. Commerce Tools engineers rebuilt the entire adidas.com website in six hours using AI-generated prompts. The capability already exceeds what exists in the market.
Tim Lavender countered with experience. His team tested AI against actual product data in an acquisition model. They did not reach 50% accuracy, and validation required more work than doing the job manually. ChatGPT loves to make up data. The shingle image example showed AI imagining medieval architecture when asked for realistic roofing products.
Sam Schwarz raised product liability. AI finds patterns where patterns do not exist and generates content without knowing when it makes mistakes. Installation instructions for electrical vaults that power cities cannot be wrong. Coating instructions that require primer application in the correct order cannot be wrong. One test produced instructions that would have triggered multi-million dollar lawsuits.
If we had a person in our lives who started finding patterns where patterns do not exist and making up things that seemed crazy, we might try to find that person help. We are relying on something that makes mistakes and does not know it is making mistakes.
Sam Schwarz, Jensen Pre-Cast
The audience voted that the potential for inaccuracy is not too much risk. Team AI Ready won round one.
Round 2: Integration with existing systems
Meeta Kratz argued that legacy systems should not prevent progress. If the legacy system contains bad data, clean the data, put it in a data lake, and work around it. The system is only as good as the data it contains.
Michael Schultz called this the best excuse to replace legacy systems entirely. Whatever technology throws at companies, whether big data or in-memory processing, legacy systems just produce bad data faster. Use AI to identify what data is clean, what needs enrichment, and start fresh.
Tim Lavender pointed to installation instructions that must be accurate. Lawyers circle looking for mistakes, and AI will not pay the legal bills. Data lives in legacy systems because that is where it was built. Getting that data out requires integration, and integration is hard.
Sam Schwarz noted that even AI advocates admitted the solution was to pull data out rather than plug directly into legacy systems. That confirms the integration challenge. Getting data from data teams is already difficult. Adding AI layers does not simplify the problem.
The audience voted that AI can integrate well enough today. Team AI Ready won round two.
Round 3: Cultural adoption
Meeta Kratz shared experience from her former employer. Using transactional data that was already accurate, her team built predictive analytics that identified cross-sell opportunities. Customer A buying four products should also buy the six products that similar customer B purchases. The insight fed sales teams and digital channels. Within one year, the program generated nearly $50 million in revenue.
If 40% of your data is accurate, why not start with the 40% that is accurate? If we wait to clean it up, we are going to be waiting another 20 years. Start there. This is iterative and self-learning.
Meeta Kratz, Lonza
Michael Schultz quoted Taylor Swift: there will always be haters. The only constant is change. AI increases productivity and accessibility. Low-code and no-code environments make impact possible. The opportunity cost of doing nothing is more scary than the risk of trying.
Tim Lavender warned that AI creates more work when validation is required. Sales reps will not embrace technology that forces them to clean up hallucinations. Sam Schwarz added that leaders must provide tools that help all the time, not 50% of the time. Trust erodes when tools fail half their users.
On governance, Sam Schwarz noted that even OpenAI applies strong oversight and has removed features that violated privacy laws. An app designed to help people with low vision started identifying faces without considering biometric regulations. Vigilance is required.
Tim Lavender announced a new hire: someone to test competitor AI implementations for data leakage from proprietary information accidentally exposed through chatbots.
The audience voted that B2B workers will welcome AI rather than see it as a threat. Team AI Ready won round three.
The verdict: AI is ready, with caveats
Team AI Ready swept all three rounds. LinkedIn polling before the event showed 60% agreement that AI is ready for prime time.
Three conclusions emerged. First, start with accurate data rather than waiting for perfect data. If transactional data is accurate because it drives billing and inventory, use it. Predictive analytics on existing good data can generate results immediately.
Second, match use cases to risk tolerance. AI works for drafting product descriptions that humans will review. AI fails for installation instructions where errors cause safety issues. Choose applications where iteration is acceptable.
Third, governance matters but should not prevent experimentation. Turn off learning functions to protect proprietary data. Monitor outputs for hallucinations. But do not let governance become an excuse to wait another 20 years while competitors move forward.