As we navigate the midpoint of 2026, the landscape of enterprise technology has undergone a seismic shift. The predictable days of flat-rate software licenses have largely vanished, replaced by a complex, high-velocity economy of tokens and consumption-based metrics. Leading this strategic conversation is Vijay Raina, a renowned expert in enterprise SaaS technology and software architecture. With years of experience helping organizations scale their digital infrastructure, Vijay specializes in the intersection of AI performance and financial sustainability. In this discussion, he breaks down the hidden mechanics of AI spending, the shift toward agentic workflows, and the rigorous governance required to turn experimental pilots into profitable production models.
Organizations are shifting from predictable flat-rate licenses to consumption-based models where every request and token carries a cost. Why has it become so difficult for leadership to pin down exactly what their AI initiatives are costing them?
We are currently operating in an era where AI maturity is still being defined by what I call “napkin math.” When a company moves from a controlled pilot to a full-scale production environment, our data shows that AI spend typically quadruples almost overnight. This happens because the costs are no longer consolidated; they are fragmented across AI providers, indirect cloud compute buckets, and AI features embedded within existing SaaS tools. I recently worked with a bank where the initial projected spend was roughly $3 million, but it surged to over $25 million once employees gained ungoverned access to frontier models. The difficulty lies in this lack of visibility—executives are seeing a “sensory overload” of invoices that don’t always map back to specific business value.
To understand these invoices, we have to look at the fundamental unit of the AI economy. How should a non-technical executive visualize a “token” and its impact on their bottom line?
Think of a token as the smallest unit of information a model processes, similar to a syllable in a word or a specific set of pixels in a digital image. In 2026, tokens have become the universal currency for billing, usually calculated in increments of one million. While a simple greeting like “How are you?” might only consume six tokens, the costs escalate when users provide massive context windows filled with images or dense reference documents. It is also vital to distinguish between input and output tokens; the tokens generated by the model are significantly more expensive than those we feed into it. Furthermore, modern reasoning models generate “intermediate thinking tokens” that remain invisible to the user but still appear on the final bill.
We’ve seen a rapid transition from per-seat licensing to these consumption-based models throughout the first quarter of this year. What is driving this evolution in how vendors charge for their software?
This is a fundamental shift in economic strategy by labs like OpenAI and Anthropic to align their revenue with the actual compute resources consumed. In the first few months of 2026, we saw frontier companies move away from unlimited seat-based pricing because it was no longer sustainable given the massive compute power required for advanced reasoning. For the end user, this means the habits learned during the era of fixed pricing—like inefficient prompting or dumping entire databases into a chat window—are now becoming incredibly expensive liabilities. If you use the same amount of AI today as you did last year, but under a consumption contract, your costs could effectively triple instantaneously. This transition forces a level of “token hygiene” that most organizations simply haven’t trained their staff for yet.
It seems counterintuitive that while the price per token has reportedly fallen by nearly 90 percent, many enterprises are reporting that their total AI bills are actually climbing. Is this a version of the Jevons paradox playing out in real-time?
That is exactly what we are seeing: as the unit price of AI intelligence drops, the demand and total consumption are exploding. We saw a similar trend years ago with mobile data; as the price per gigabyte fell, people stopped just sending emails and started streaming high-definition video. In the AI world, we have moved into an “agentic paradigm” where we aren’t just chatting with a single model, but deploying networks of agents. You might have one lead agent orchestrating four sub-agents to complete a task, which creates an exponential growth in the number of tokens being sent and received. By April and May of 2026, many executives realized that while tokens were cheaper, a single power user could still rack up thousands of dollars in costs in a single day by running these complex, multi-agent workflows.
When we look at the cost of running a single AI agent, what are the specific factors that can make one agent thirty times more expensive than another?
The cost of an agent is largely determined by the efficiency of its design and the “stateless” nature of the underlying models. Because models don’t naturally remember past interactions, you have to feed the entire history back into the system to maintain context, which causes costs to compound with every turn of the conversation. A poorly designed agent might repeatedly make model calls for simple tasks that could be handled by traditional, deterministic code, such as checking a repository for a specific device type. We also have to account for the “failure rate” cost; if an airline agent fails to rebook a flight because the seat was taken during the processing time, the entire token-heavy process must be repeated. Beyond the inference costs, you also have the overhead of vector storage, orchestration, and the human hours spent validating the output.
With these compounding costs, how can a business leader determine if a specific AI tool or agent is actually worth the investment?
The most effective metric is calculating the cost of the agent per successful output and comparing that against the “fully loaded” cost of the human alternative. For example, an airline deploying a rebooking agent handles thousands of requests daily, making the high development cost worth it through sheer volume and speed. Conversely, an agent designed for a task performed only once a year, like auditing certain tax filings, might never reach a positive ROI. I often categorize AI value into three buckets: personal productivity tools, automation for efficiency, and complete business model reinvention. For automation, you must strictly ask: “How much am I saving in labor and maintenance versus the ongoing token spend?”
What practical steps should companies take right now to bring these escalating costs under control without stifling innovation?
Control starts with ownership of the “harness”—the software and governance layer that sits around the language model. Organizations need a unified AI control plane where they can track every request back to an individual user and understand why a specific model was chosen for a specific task. I’ve seen teams achieve 15 to 20 percent savings simply by automating controls that route simple requests to smaller, open-source models rather than expensive frontier models. There is also a massive amount of “value leakage” when employees use high-end coding assistants to find a single fact in a document, a habit that must be broken through upskilling. Ultimately, if you can’t measure the cost of the task, you cannot manage the success of the initiative.
What is your forecast for the future of AI financial management?
I predict that the “wild west” of uncontrolled AI experimentation will end by the close of 2026 as CFOs demand the same level of rigorous financial observability we see in mature cloud environments. We will see a shift where companies no longer just “throw AI” at a problem, but instead focus on reimagining the underlying workflow before the first token is ever spent. While the temptation to adopt every new model for a competitive advantage is high—much like an athlete feels pressured to use every available tool—the winners will be those who balance that drive with a strict “cost-to-value” architecture. AI will certainly be worth the investment, but only for organizations that treat tokens with the same fiscal discipline as they treat their headcount or capital expenditures.
