ORLANDO, Fla. — Artificial intelligence systems depend on more than powerful models. They also depend on the quality of the data used to train, test and evaluate them.
As companies use AI in areas such as finance, healthcare, customer service and public infrastructure, data quality has become a larger business concern. A model may appear advanced, but poor labeling, inconsistent datasets or weak evaluation can lead to unreliable results.
That is why more AI infrastructure companies are focusing on the data pipeline behind the models.
Zihan Wang, co-founder and chief research officer at Abaka AI, said reliability has to be built into AI systems early.
“Reliability is not a feature we add later,” Wang said. “It’s the condition that allows every other feature to exist.”
Abaka AI is one company working in that part of the market. The Palo Alto-based data infrastructure company has announced $8 million in funding to expand its work around multimodal AI data, including collection, annotation and evaluation.
Its platform, MooreData, is designed to help enterprise teams manage datasets across formats such as text, audio, images, video and other data types.
The challenge is becoming more important as AI systems are used in more complex environments. A company building a model for a regulated industry may need to know not only whether a dataset is accurate, but also how it was created, reviewed and changed over time.
Wang said that kind of traceability is especially important when AI systems are used in high-stakes settings.
“Evaluation should be as transparent and traceable as a lab experiment,” Wang said. “When teams can reproduce a model’s behavior with confidence, that’s when AI becomes truly enterprise-ready.”
For companies building AI, the problem often starts before a model is trained. Data may come from different sources, use different labeling standards or require review by people with specialized expertise.
If those steps are inconsistent, model performance can suffer later.
Abaka AI says MooreData uses a layered quality assurance process that includes expert labeling, domain-specific sampling audits and automated error detection. The company also offers deployment options for public cloud, private server and hybrid environments.
Those controls can matter for companies that need privacy, security and audit trails while developing AI products.
The company is also involved with the 2077AI Foundation, an open-source initiative focused on benchmarks and interoperability. Wang said shared standards will become more important as organizations compare AI systems and evaluate performance.
Through the foundation, Wang’s team has contributed to open benchmarks tied to multimodal document parsing. One example is OmniDocBench, a benchmark published at CVPR 2025 that has been used as core evaluation data for DeepSeek-OCR.
“Our partnerships aren’t about scale,” Wang said. “They’re about consensus. The future of AI governance will depend on how well we agree on what ‘good data’ actually means.”
The broader issue for enterprises is that AI reliability is becoming harder to separate from data reliability. As models become larger and more capable, companies still need systems that can explain where data came from and how it was evaluated.
That is especially true as businesses use AI in settings where errors can affect customers, compliance or operational decisions.
For Wang, the next stage of AI development will depend less on model size alone and more on whether companies can trust the information behind those systems.
“The race for larger models will continue,” Wang said. “But the real competition will be over who can trust their data the most.”
As AI becomes more common in business operations, data quality may become one of the biggest factors separating experimental tools from systems companies can safely use at scale.
©2026 Cox Media Group
