Uber's CTO Praveen Neppalli Naga made a quietly alarming admission earlier this year: his entire 2026 AI tooling budget, set aside for the full fiscal year, was gone by April. Not because of a bad vendor deal or runaway infrastructure spend, but because engineers were using Claude Code, and usage-based token costs scaled faster than anyone had modeled. "I'm back to the drawing board," he said, "because the budget I thought I would need is blown away already."
This isn't a one-off story. It's becoming a pattern.
Around the same time, Nvidia VP of applied deep learning Bryan Catanzaro offered his own uncomfortable observation: "The cost of compute is far beyond the costs of the employees." For his team at one of the most AI-forward companies on the planet, the token and infrastructure bills for running AI at scale already dwarf the salaries of the people it's supposedly replacing.
The Pitch Didn't Include the Fine Print
The standard AI-as-cost-cutter narrative goes something like this: deploy agents, reduce headcount, watch costs fall. It's clean and intuitive. It maps neatly to a spreadsheet. What it ignores is the actual structure of AI pricing and how dramatically that pricing can depart from the "per seat" software models most finance teams understand.
Traditional SaaS is easy to budget. You pay per license, per month. The cost is predictable. AI agents priced on token consumption work completely differently: costs scale with usage intensity rather than user count. An autonomous agent solving a complex problem doesn't just consume one API call. It might chain together dozens of reasoning steps, each one burning tokens. Research on agentic workloads shows they can cost 100x more than simple chatbot interactions, because agents make 3 to 10 times more LLM calls per task and often work with long context windows that amplify input token costs.
The math compounds quickly. Jason Calacanis, investor and co-host of the All-In podcast, described running Claude API agents that hit $300 per day per agent, approaching $100,000 annually per agent, at only a fraction of an employee's output. Chamath Palihapitiya put it bluntly: once you aggregate token spend across an organization, you hit a threshold where you're asking "do we need each of our employees to be at least twice as productive to justify this?"
That question doesn't appear anywhere in the typical boardroom pitch.
What "Fully Loaded Cost" Actually Means for AI Agents
When companies compare AI agent costs against human salaries, they usually make the comparison incorrectly, understating AI costs while understating human costs too. Here's what the numbers actually include:
Token and inference costs are just the starting point. Input tokens, output tokens, and for advanced models, "thinking" tokens all add up. For a heavily used agent running on a frontier model via API, this alone can run into tens of thousands of dollars annually.
Orchestration and infrastructure add another layer. Logging, monitoring, vector storage, retrieval systems, and agent orchestration platforms all carry their own costs. According to practitioners tracking these deployments, token spend is typically only 30 to 60% of the real total cost. Teams that report only inference costs are understating their actual AI spend by 40 to 70%.
Quality assurance and error recovery are where costs accumulate without showing up in anyone's dashboard. AI agents fail in ways that require human intervention. They misinterpret edge cases, get stuck in loops, or produce outputs that look correct but aren't. A well-publicized case involved a fintech engineering team whose two LangChain agents entered an infinite reasoning loop for 11 days, running up $47,000 in inference costs on a pipeline budgeted at under $200 per month.
Human-in-the-loop (HITL) roles are the most overlooked cost. The human oversight layer doesn't disappear when you deploy AI agents. It transforms. Someone still needs to review edge cases, handle escalations, audit outputs for regulated workflows, and intervene when agents go sideways. That human time has a fully loaded cost that rarely makes it into the ROI model.
Gartner's updated 2026 figures put the scale of this problem in perspective: the average enterprise AI deployment ends up costing 2.8x the original estimate, with one in four projects abandoned mid-deployment due to budget overruns rather than technical failure.
The 23% Rule
A 2024 MIT study examined which tasks are actually cost-effective to automate with AI and found that, even accounting for all the productivity optimism, automation makes clear financial sense in only 23% of applicable roles today. Human workers remain the more economical option in about 77% of cases, particularly at smaller scales where fixed AI deployment costs can't be amortized across high volumes.
That figure will change as model costs drop. But right now, betting on AI automation because it will eventually be cheaper than human labor is a different business case than betting on it because it's cheaper today. Many companies are making the latter argument with the former's math.
The Workflow Sorting Problem
Not all work is equally automatable, and the economics differ so sharply by workflow type that where your task lands may matter more than your choice of model or vendor.
Green Zone: High-Volume, Low-Variability Tasks
These are the workflows where AI agents genuinely shine. The task structure is predictable enough to keep error rates low. High volume amortizes the fixed deployment costs. And when an error does happen, it's usually recoverable.
Good candidates: invoice processing, data normalization, standard customer service routing, scheduled reporting, form completion, log analysis, and structured document extraction. IBM's Enterprise Advantage platform claims 50 to 60% cost reductions for agentic workflows in these categories, and those numbers are plausible when the task profile matches.
Yellow Zone: Moderate Complexity, Moderate Stakes
These tasks can work with AI agents, but require more careful design around when to escalate to humans and how to handle exceptions. The ROI is real but requires investment in HITL architecture that most initial deployments underestimate.
Good candidates: contract review support, research summarization, onboarding workflows, HR query handling, and financial close support. Workato's Otto, launched in 2026, is explicitly designed for this tier. It works within existing enterprise governance frameworks, keeps credentials away from the LLM, and routes decisions requiring human judgment to the right people via Slack or Teams rather than trying to resolve everything autonomously.
Red Zone: Complex Judgment Calls with High Error Stakes
These are the cost traps. Be wary of automating tasks where the cost of an AI error (legal, financial, or reputational) materially exceeds the savings from automation. This isn't because AI can't attempt these tasks. It's because the oversight, quality assurance, and error recovery infrastructure needed to make them safe often costs more than the human labor being replaced.
Watch out for: compliance sign-offs, complex customer dispute resolution, strategic financial analysis, high-stakes hiring decisions, and anything where "the agent got confused" creates regulatory exposure.
Building a Business Case That Survives the CFO's Questions
Most AI ROI models collapse when a CFO asks the right questions. The ones that don't share a few structural features.
Start with fully loaded cost, not token cost. Your model should include inference costs at realistic usage volumes (not demo-scale), orchestration infrastructure, HITL labor priced at fully loaded cost rather than base salary, engineering maintenance and prompt tuning, and quality assurance overhead. If your model only captures token spend, multiply it by at least two before presenting it as the real number.
Model token costs at the 90th percentile of expected usage, not the median. Agentic workloads spike. Budget for worst-case usage and treat median cost as the upside scenario, not the base case. Uber's CTO modeled the median. You know how that turned out.
Set hard token budgets and monitor them weekly. The major AI providers allow spending caps per API key or organization. Set them before you deploy, not after your first budget blowout. Token spend without guardrails compounds invisibly; by the time the bill lands, you've already committed the spend.
Measure outcome ROI, not just completion rate. The formula that matters: business value of outcomes produced divided by fully loaded agent cost. An agent that completes 95% of tasks but routinely produces outputs that require significant human rework is not generating the ROI the completion rate implies.
Define a 90-day kill criterion before you start. What does failure look like? At what cost-per-outcome ratio will you pause and reassess? Having this defined in advance changes the deployment posture. You go from hoping it works to actually measuring and adjusting. That's the governance discipline that separates teams that get this right from expensive experiments.
What Getting This Right Actually Looks Like
The enterprise platforms that have emerged over the past year reflect a more sophisticated understanding of where the economics break down. IBM's Enterprise Advantage, announced in January 2026 and expanded at Think 2026, is built around governed agentic workflows where every agent action is logged, auditable, and traceable back to a policy. Rather than deploying general-purpose agents on complex tasks, they've built a pre-packaged catalog of agents scoped to high-volume, structured workflows: document extraction, regulatory reporting support, customer query resolution.
Workato's Otto takes a similar approach. It doesn't replace human decision-making; it handles coordination, follow-ups, and multi-system execution while escalating to humans when actual judgment is required. The design assumption is that the agent and the human work together.
On the SMB side, Sage's expanded partnership with AWS focuses on finance-specific agents covering accounts payable, payroll, cash flow, and compliance. Domains where the task structure is well-defined and the volume-to-complexity ratio genuinely favors automation. For mid-market finance teams, that's a more defensible ROI story than generic agent deployment.
Every platform in this space that's working built governance, scoping, and human escalation into the core architecture from the start. That order of operations is not accidental.
Measure First, Deploy Second
AI agents can absolutely deliver ROI. The companies achieving it aren't the ones that moved fastest or cut headcount most aggressively. They're the ones that measured the right things before they committed.
Token-based pricing is not a reason to avoid AI agents. It's a reason to deploy them with the same financial discipline you'd apply to any significant capital decision. That means modeling the full cost stack, choosing workflows where the task profile matches the technology's economics, and building governance infrastructure before you need it rather than after an 11-day token loop empties your Q2 budget.
The pitch deck version of this story was always a simplification. The question was never whether AI can replace employees. The better question: for which specific tasks, at what cost, with what oversight, does deploying an AI agent deliver better outcomes per dollar than the current approach?
That question has clear answers. They're just less exciting than the boardroom version, and far more likely to survive contact with reality.