Introduction
Artificial intelligence continues to evolve at breakneck speed. Business owners across industries are grappling with a critical question: should they integrate Large Language Models (LLMs) into their operations? From customer service chatbots to automated content creation, from inventory management to financial analysis, LLMs promise to revolutionize how small businesses operate. However, the latest generation of AI models presents a complex paradox that every business leader needs to understand before making the leap.
Meet Sarah, who runs a small-sized accounting firm. She’s been evaluating whether to implement an AI assistant to help with client communications and research for preliminary tax advice. Sarah is is excited about the potential efficiency gains but worried about the risks and costs involved.
She is facing the reality that while newer models like GPT-5 offer impressive capabilities, they come with trade-offs that could significantly impact her bottom line and her business reputation.
Hallucinations: Better, But Still a Business Risk
One of the most persistent challenges with LLMs is their tendency to “hallucinate” or generate incorrect information with complete confidence. While the latest models show improvement, the problem remains far from solved.
Recent data from OpenAI reveals that GPT-5 makes incorrect claims 9.6 percent of the time, compared to 12.9 percent for GPT-4o. This represents a 26 percent reduction in hallucination rates and 44 percent fewer responses with major factual errors. While this sounds promising, it means that roughly one in ten responses from the most advanced model could contain false information.
For business applications, this statistic should give you pause. Imagine if Sarah’s AI-powered customer service system provided incorrect information to 10% of her customers, or if her automated researcher gave faulty guidance one time out of every ten interactions. In industries where accuracy is paramount, such as accounting, healthcare, legal services, or financial planning, even a 9.6% error rate could expose your business to significant liability risks.
The legal landscape around AI-generated errors is still evolving, but early court cases and regulatory guidance suggest that businesses remain fully responsible for AI-generated content and decisions. The Federal Trade Commission has made clear that companies cannot simply blame their AI systems when things go wrong. In regulated industries, using AI tools that produce inaccurate information could violate compliance requirements, potentially resulting in fines, lawsuits, or loss of professional licenses. Even in less regulated sectors, customers harmed by AI-generated misinformation may have grounds for breach of contract or negligence claims.
Even small businesses using LLMs for tasks like writing marketing copy, generating product descriptions, or creating training materials need robust fact-checking processes. The cost of implementing these verification systems often negates much of the efficiency gains that initially made AI adoption attractive.
Speed: Faster Processing, But At What Cost?
The good news is that newer LLMs are generally faster than their predecessors. This improvement comes from several technological advances including more efficient attention mechanisms, better memory management, model compression techniques, and specialized hardware optimizations.
Modern LLMs benefit from architectural innovations like Flash Attention and mixture-of-experts designs that streamline computations. Techniques such as quantization reduce model size by using lower-precision numbers, leading to faster processing times. Additionally, dedicated hardware like NVIDIA’s H100 chips and optimized software frameworks have dramatically improved response times.
For businesses like Sarah’s, this means AI assistants can handle customer inquiries more quickly, content generation happens in near real-time, and data analysis tasks that once took hours can be completed in minutes. A restaurant chain using AI for menu optimization (skip the Elmer’s glue, if the cheese keeps sliding off the pizza) or a retail store leveraging AI for inventory forecasting can see immediate operational benefits from these speed improvements, provided the forecasts are accurate.
However, faster processing often comes with increased complexity and resource requirements. While individual responses may be quicker, the underlying infrastructure needed to support these advanced models requires significant investment in both hardware and ongoing maintenance.
Energy Costs: The Hidden Environmental Impact
Perhaps the most overlooked aspect of newer AI models is their dramatically increased energy consumption. While individual businesses using cloud-based AI services may not see these costs directly on their electricity bills, the environmental and societal impact is very real and growing.
Research from the University of Rhode Island found that GPT-5 can consume up to 40 watt-hours of electricity to generate a medium-length response of about 1,000 words. To put this in perspective, 40 watt-hours is about what is needed to nuke your lunch, a significant increase over previous models like GPT-4o.
When you use cloud-based AI services like ChatGPT, Claude, or integrated AI features in business software, the energy costs are absorbed by the service providers and ultimately reflected in the massive power consumption of data centers worldwide. With ChatGPT reportedly handling 2.5 billion requests daily, the collective energy demand is staggering. The total consumption of widespread GPT-5 usage could reach the daily electricity demand of 1.5 million US homes.
This hidden environmental cost has several business implications. First, as energy costs rise and environmental regulations tighten, AI service providers will inevitably pass these costs to customers through higher subscription fees or usage-based pricing. Second, businesses with sustainability commitments may find their AI usage conflicts with carbon reduction goals, even if they’re not directly paying the electricity bills. Third, as society becomes more aware of AI’s environmental impact, customers and stakeholders may increasingly scrutinize companies’ AI usage practices.
For businesses running AI models locally or on private cloud infrastructure, these energy costs become direct operational expenses that can substantially impact budgets. A marketing agency generating 100 AI-assisted blog posts per month could see their electricity costs increase substantially when running models on-premises.
OpenAI and other major providers have been notably secretive about publishing detailed energy consumption data, making it difficult for businesses to assess the true environmental impact of their AI usage or anticipate how energy costs might affect future pricing models.
The Bottom Line for Business Owners
The latest generation of LLMs presents a classic business dilemma: impressive capabilities coupled with significant trade-offs. While these models are faster and somewhat more accurate than their predecessors, they come with higher error rates than many business applications can tolerate and energy costs that could strain operational budgets.
Before integrating LLMs into your business processes, consider these key factors:
Risk Assessment: Can your business tolerate a 10% error rate? Do you have systems in place to catch and correct AI-generated mistakes before they reach customers?
Legal and Regulatory Compliance: Understand that you remain fully responsible for AI-generated content and decisions. Consult with legal counsel about liability issues, especially in regulated industries.
Cost Analysis: Factor in not just subscription or usage fees, but also the likelihood of rising prices as energy costs increase, and the expense of implementing verification processes.
Environmental Impact: Consider how AI usage aligns with your company’s sustainability goals and stakeholder expectations regarding environmental responsibility.
Use Case Selection: Focus on applications where speed matters more than perfection, such as initial draft creation or brainstorming, rather than final customer-facing content or critical business decisions.
Gradual Implementation: Start with low-risk applications and gradually expand as you better understand the models’ limitations and your organization’s ability to manage them.
The promise of AI is real, but so are the risks and costs. The businesses that succeed with LLM integration will be those that approach adoption strategically, with clear-eyed understanding of both the benefits and the hidden expenses of this powerful but imperfect technology.
For business owners like Sarah, the key is not whether to adopt AI, but how to do it intelligently, with appropriate safeguards and realistic expectations about what these impressive but flawed tools can actually deliver.
Sources:
Hallucination Rates:
Cecily Mauran, “OpenAI says GPT-5 hallucinates less — what does the data say?”, Mashable, August 7, 2025, https://mashable.com/article/openai-gpt-5-hallucinates-less-system-card-data?test_uuid=003aGE6xTMbhuvdzpnH5X4Q&test_variant=a
Energy Usage:
Aisha Kehoe Down, “OpenAI will not disclose GPT-5’s energy use. It could be higher than past models” The Guardian, 09Aug2025 https://www.theguardian.com/technology/2025/aug/09/open-ai-chat-gpt5-energy-use
This post is a collaborative effort between several AIs and the AI experts at GunderFish