There’s a growing problem in the AI world that’s become impossible to ignore: the runaway token spend firms are racking up as they put AI agents to work inside their organizations.
When agentic AI first came to the fore, the thesis was simple — many of the repetitive tasks people used to do could be handled faster, more accurately, and more cheaply by AI. What a lot of teams didn’t fully appreciate is that this capability comes at a real, and often unpredictable, cost.
That cost has caught up with them. Enterprise AI bills have roughly tripled over the past year as agentic workflows have spread, with average monthly AI spend climbing from $63,000 in 2024 to $85,500 in 2025 — a 36% jump, according to CloudZero’s research on AI costs. In some cases, the bill for running an AI agent now exceeds the fully loaded cost of the employee it replaced.
I’m not here to argue that AI is bad at replacing repetitive work — quite the opposite. When a task can be done faster, more accurately, and at equal or lower cost, that’s a genuine value add. But when the cost of running these systems gets so high that managers can no longer tell whether the tool is paying for itself, that’s a signal you need to rethink how you’re deploying the technology.
This is why Serac Ventures recently invested in DocuSOR, a company that helps organizations refine their context before it ever reaches an AI system. If you’d like to read our one-page Funding Announcement, click HERE.
To understand why this matters, it helps to understand where token spend actually comes from. Most people think about AI cost in terms of output — the answer the model gives back. In practice, the bigger driver is input: every character you feed into a model is tokenized and billed, and that includes everything the model has to re-read just to understand what you’re asking. LLMs don’t have persistent memory between calls. Every time an agent takes a new action, it has to re-send the entire conversation history — system prompt, prior tool results, every intermediate step — back to the model just to generate the next output. This creates what researchers have called a quadratic cost problem: a workflow that feels like ten small steps can quietly accumulate 80,000 to 200,000 tokens, because every piece of context introduced early on gets re-billed on every subsequent turn. One recent industry analysis put a number on it: agentic workflows require 5 to 30 times more tokens per task than a standard chatbot interaction, and teams that estimate cost as “turns times average cost per turn” tend to underprice their systems by 3 to 5x. Add to that the fact that many providers don’t do a great job explaining this to their customers, and it’s easy to see how bills spiral before anyone notices.
DocuSOR attacks this problem directly. The company builds a repository of refined, structured context that agents can draw on, so that instead of dumping raw, unfiltered information into an LLM, teams feed it exactly what’s relevant. A leaner, better-curated context means the agent needs fewer tokens to get to the right output — which means lower cost and, often, a faster and more accurate result. We think this becomes a real category: infrastructure for context, sitting underneath the agent layer, built for every company trying to get its token spend under control.
We see this as a substantial market opportunity. DocuSOR’s founders have direct operating experience as developers and executives inside large fintech and software businesses, and they bring a strong network of early users who can help them reach product-market fit and prove out efficacy quickly.
What we also find compelling about the DocuSOR team is where they come from. The two founders (Omari Gaskins & Jordan Williams) are based in the Midwest, outside the traditional AI hubs, and they weren’t caught up in the hype cycle of building more LLMs and more agents. Instead, they looked at the underlying infrastructure problem — how context, tokens, and cost actually interact — and built the tools needed to make this new AI paradigm sustainable.
We’re excited to back a company building the infrastructure layer that helps small and large companies alike get a handle on their token spend. We’re still in the early innings of a much bigger shift in how software companies charge for their products. For years, the industry ran on subscriptions. That’s changing fast: Gartner projects that 70% of businesses will prefer usage-based pricing over traditional per-seat models by 2026, and the share of SaaS companies using some form of usage-based pricing has grown from 45% in 2021 to roughly 61% today. Companies that have made the switch aren’t just following a trend — they’re growing meaningfully faster than subscription-only peers.
That shift changes the calculus for every business buying and building on AI. In a usage-based world, companies have to be just as disciplined about what they consume internally as they are about what they charge externally — because if you don’t understand your own consumption, your pricing to customers won’t hold up either. Companies like DocuSOR are building the backbone of that discipline, and we think it becomes the new norm across the technology ecosystem.
If you’d like to learn more about DocuSOR, visit them at www.docusor.ai, or feel free to reach out to us at Serac Ventures to learn more about the company.
