Maximizing AI ROI Through Smarter Model Usage

AI ROI is not determined by how much you use AI but by how well you use it. Model usage and cost are of course secondary considerations within the broader AI business case. The primary goal must be to capture measurable business value. Hence, the ultimate goal of any AI strategy must be to maximize the output of meaningful, innovative work with minimum AI budget input. The focus should be on maximizing business outcomes first, then optimizing model performance and cost. 

Still, efficient model usage matters because it directly impacts unit economics and determines how much value is retained. Organizations that operationalize AI efficiency through better prompting habits, smarter tooling, and tighter governance will capture more value. Smarter model usage matters because:

• AI adoption is ramping up hidden costs. Every foundation model prompt, every uploaded file, every long-running conversation consumes tokens. And tokens translate directly into spend, energy use, and ultimately ROI. The issue is not increased AI usage, it is inefficient foundation model usage.
• This is not just a financial issue. Inefficient AI usage also increases energy consumption and carbon emissions, creating friction with ESG commitments. Token discipline, therefore, is both a cost and sustainability lever.

Most Organizations Underestimate How Quickly Costs Scale

Bloated inputs, unnecessary images, excessive conversation history, and overreliance on premium models all compound token consumption. What looks like marginal usage at the individual level becomes material at scale. Multiply inefficient workflows across hundreds of employees, and foundation models quickly shift from value drivers to cost centers.

Inefficient Foundation Model Usage Is Controllable

AI ROI is less about limiting usage and more about improving discipline:

• Start with model selection. Not every task requires a frontier model. Reserve premium models for complex reasoning and use lighter models for execution, formatting, and refinement. This ‘model portfolio’ approach alone can deliver significant savings without sacrificing output quality.
• Address input efficiency. Many teams upload raw PDFs, slide decks, or screenshots – formats designed for humans, not machines. These carry hidden overhead like metadata, formatting, images, and layout instructions that inflate token counts. Converting inputs into clean text or markdown and minimizing image use can reduce costs dramatically while improving model performance.
• Master conversation management. Long-running chats reprocess the entire context with every new prompt, creating a compounding cost effect. Separating exploration from execution and regularly starting fresh, focused conversations reduce cost and noise.
• Design AI governance structures and AI training for your workforce. Employees rarely understand the cost implications of their AI usage. Embedding cost awareness into training, setting clear usage guidelines, and auditing tools and integrations can prevent silent inefficiencies from scaling.
• Leverage technical optimizations where relevant. Prompt caching, retrieval-based approaches, and scoped workflows ensure that models process only what is necessary, and only once. The objective is simple: minimize tokens per unit of useful output.

In PAC’s report Maximize AI ROI Through Smarter Model Usage, we outline steps to maximize returns for AI model investments as part of everyday AI usage at work.