...
By Sabrina Tullo
0
Share
Newsweek is a Trust Project member
See more of our trusted coverage when you search.Prefer Newsweek on Googleto see more of our trusted coverage when you search.
Artificial intelligence is scaling faster than the safeguards around it. Enterprise use is moving into the core workforce, while the technology continues to produce confident errors that can turn a productivity tool into an unforeseen liability. OpenAI reports an eightfold increase in ChatGPT Enterprise message volume over the past year, while hallucinations remain a stubborn weakness even as models grow more capable.
The consequences are moving from awkward chatbot answers to professional risk. Recently, the Connecticut Supreme Court warned lawyers after fake AI-generated case citations surfaced in legal filings. Days earlier, the Financial Times reported that several PwC Middle East reports contained fabricated footnotes and other errors attributed to AI hallucinations. The incidents point to a problem enterprises cannot solve by simply buying access to the newest model: the system surrounding the model has to make its output dependable.
Security is becoming equally difficult to ignore. Reuters reported last week that Obsidian Security raised $85 million at a $1.1 billion valuation as demand rises for technology securing AI agents accessing sensitive business information. Nearly 70 percent of its clients now allow AI agents to interact with company data, according to the company’s CEO.
Read More on Essays
Some industry estimates suggest AI hallucinations may be costing businesses tens of billions of dollars annually, with organizations reporting losses that can reach millions of dollars when errors go undetected. As AI gains permission to move through corporate systems, what it can access becomes inseparable from whether businesses can trust it.
Ryan Taylor, the CTO of OSIBytes, believes the industry has spent too much time measuring AI by what it can generate and too little examining how it reaches an answer. His argument starts with the architecture underneath the interface: businesses need systems that control where information comes from, keep proprietary data contained and test whether an answer deserves confidence.
Taylor’s perspective comes from decades spent building enterprise infrastructure. That background shaped his approach, particularly after he began investigating hallucinations during his master’s studies in 2024. “It really comes down to the mathematical formula that’s inside their model structures,” he says, arguing that the architecture governing how information is collected and extracted can determine the reliability of the final answer.
He points to a basic failure pattern in conventional AI systems, wherein a model may pull from a limited slice of available information, then assemble an answer from an incomplete context. “The theory goes that the more data you throw at it, the better it becomes. We’ve seen some of that improvement, but at the end of the day, the extraction is where the failure is, not the injection,” Taylor says.
He saw the commercial stakes firsthand while using AI across multiple platforms to help build a website. After spending roughly $5,000 on development, he recalls, the system deleted the entire program because it could not determine what to do with the file. He says, “It was hallucinating, going down a path, trying to keep track of things, and the memory available to it was limiting what it could actually do.”
According to Taylor, that episode captures the gap between impressive demonstrations and enterprise-grade infrastructure. His approach uses what he calls a hyper-agent structure, with 40 agents collecting information from separate data libraries before additional agents reconcile the samples and identify common ground. He compares the architecture with a library in which different agents retrieve information from specific sections, allowing the system to work with narrower, controlled pools of data.
The architecture also addresses a problem executives may consider even more sensitive than hallucinations: intellectual property. Taylor says PyBlox models, OSIBytes’ private AI and agentic code-to-cloud platform, are built for individual customers and kept specific to their use case, with customer information kept separate from other models. “We don’t cross-contaminate models from one client to another,” he says. For businesses weighing whether to place proprietary material into AI systems, that boundary can become a prerequisite for adoption.
Taylor argues that infrastructure will gain importance as AI moves deeper into corporate operations because every new capability creates another point of access that must be controlled. “If you give AI access to your infrastructure, you don’t really want it to go out of that area,” he says.
The question, from his viewpoint, becomes architectural. What permissions does an AI system have, what information can it reach and what happens when its reasoning fails? The premise of OSIBytes, a digital AI technology partner with PyBlox as its AI infrastructure brainchild, reflects Taylor's view that useful AI depends on the machinery surrounding the model. He says the company is developing partnerships with businesses and venture partners interested in testing where its infrastructure can support enterprise needs.
The larger shift may already be underway. AI is becoming easier to deploy, yet the harder business debate is whether organizations can give it access without surrendering control. Taylor’s stance remains cautious. He says, “Don’t trust AI; it’s not taking our jobs, it’s taking away the mundanity from our day-to-day tasks. AI can produce code faster than any person could ever create it, but if that code is broken, what’s the point?” Human oversight, he insists, still has to determine whether the output works.
The AI industry has spent years treating scale as the ultimate competitive advantage: bigger models, more computing power, more data, faster outputs. The next fight may be decided by a less glamorous question: what happens when those systems are wrong? As AI moves deeper into enterprise infrastructure, the companies with the strongest claim on the future may be those solving the unglamorous engineering problems that determine whether businesses can safely hand AI their data, their workflows and eventually their decisions.
View 0 Comments
Submit Correction
View Editorial & AI Guidelines