The "open tab" fallacy perfectly mirrors one of the most famous hospitality blunders in history: The Savoy Hotel’s Uncapped Omelette.
In the early 20th century, a wealthy patron at London's Savoy Hotel told the chef to "keep making omelettes for the house" until he said stop, assuming a fixed, reasonable hospitality rate. The chef, seizing the opportunity, used rare white truffles and vintage Madeira wine in every single dish. The patron forgot to close the tab, went to bed, and woke up to a bill equivalent to roughly $15,000 today.
Treating Generative AI like a traditional SaaS license is exactly like ordering "the omelette" without checking the ingredients. Leaders assume they are paying for a standard egg dish (a fixed SaaS seat), but employees are unknowingly ordering white truffles (using 200k-context reasoning models to summarize a three-sentence email). Every single prompt is not a click; it is a custom-cooked dish with variable-cost ingredients (tokens).
Enterprise AI is an open tab at a restaurant with no menu prices. Leaders are telling employees to consume as much AI as possible, which is measured in tokens and credits. However, few are checking the bill until it arrives. For one unnamed organisation discussed in the previous post, that bill is $500 million for a single month of Claude usage. This is an astonishing structural failure of leadership and financial governance, reminiscent of early shocks when people started to use cloud computing indiscriminately.
The problem is the "Open Tab" fallacy: the belief that AI is a fixed-cost utility like a SaaS seat. In reality, generative AI is a variable-cost engine where every prompt is a financial transaction. When companies fail to establish spending caps, they expose themselves to uncapped liabilities.
The Problem: Uncapped Consumption and "Checking the Weather"
The recent surge in surprise AI bills is a direct result of a lack of data fluency in the C-suite. Organisations are treating AI tokens as "free" resources during the adoption phase. This approach is dangerous.
Recent data highlights the scale of the issue:
- Microsoft reportedly canceled internal licenses for Claude Code after per-engineer costs reached $500 to $2,000 monthly (Hedgie Markets, 2026).
- Uber exhausted its entire 2026 AI budget by April (Hedgie Markets, 2026).
- Anthropic and other providers offer enterprise consoles with usage limits, yet many companies deploy models to thousands of employees without configuring a single spending cap.
When usage is the only metric, waste is the inevitable result. One CTO noted that employees are using high-end reasoning models to perform trivial tasks, such as checking the weather. This is the equivalent of hiring a partner-level consultant to file a single expense report.
The Diagnosis: Usage is Not Value
The "Open Tab" fallacy is a manifestation of Goodhart’s Law: "When a measure becomes a target, it ceases to be a good measure." Many organizations are measuring "AI adoption" by total token usage. This metric is a perverse incentive.
If employees are ranked on an internal leaderboard based on AI usage: as seen in recent Amazon and Uber examples: they will "game" the system. They will run fake tasks, use overly complex models for simple work, and ignore cost efficiency.
The primary problem is a disconnect between unit economics and business outcomes. Organizations are paying for "Silicon Salaries" without a clear job description for the AI. To fix this, leaders must move from tracking "tokens used" to tracking "cost per successful action."
The Prescription: A CFO Checklist for AI Auditing
To gain control over AI spending, leaders must implement a mechanical approach to cost management. This is not about stifling innovation; it is about establishing the guardrails that make innovation sustainable.
1. Implement Hard Spending Caps at the Provider Level
Most enterprise AI platforms, including Anthropic and Azure, offer usage limits and dashboard alerts. These are not suggestions; they are requirements.
- Set daily and monthly hard limits.
- Ensure the account stops requests once the threshold is met, rather than just sending a notification.
- Separate environments (Production, Staging, Experimentation) with independent caps.
2. Apply Model Routing Based on Logic, Not Prestige
Not every task requires a frontier model like Claude 3.5 Sonnet or GPT-4o.
- Route classification and formatting tasks to smaller, cheaper models.
- Reserve "expensive" reasoning tokens for complex, high-value tasks.
- Automate model selection via an orchestration layer to prevent employees from defaulted to the most expensive option.
3. Define the "Silicon Salary" for Every Agent
A "Silicon Salary" is the cost of AI labor per unit of work.
- Calculate the total tokens (input + output) required for a specific task.
- Compare this cost to the human labor it replaces.
- If an agentic workflow costs $50 per run but only saves 10 minutes of a $40/hour employee's time, the unit economics are negative.
4. Establish Real-Time Observability
The era of waiting for the monthly invoice is over.
- Build a dashboard that tracks cost by feature, team, and individual.
- Identify the top 10 most expensive prompts and audit them for efficiency.
- Use tools like Certinia Veda, which provides a transparent consumption tier for agentic AI within a professional services context.
To close the open tab and establish financial governance, organizations must transition from blind consumption to unit-economic precision. Here are specific, actionable ideas for auditing your AI spend:
5. Establish a "Cost Per Query" (CPQ) Metric Matrix
Classify Tasks by Tier: Segment all corporate AI tasks into tiers (e.g., Tier 1: Trivial/Weather, Tier 2: Standard Drafting, Tier 3: Complex Coding/Reasoning).
Map Model to Task: Audit which models are fulfilling which tiers. If a Tier 1 task uses a high-end frontier model, flags should trigger automatically.
Calculate Margin Impact: Determine the exact financial return of a prompt. If an engineer uses $10 worth of tokens to fix a bug, did it save $10 worth of manual labor time?
6. Implement "Smart Gastronomy" (Cascading Router Architecture)
Default to the Cheapest Model: Program your internal AI gateway to route all user requests to the smallest, cheapest model by default (e.g., Claude 3.5 Haiku or GPT-4o-mini).
Escalation Protocol: Allow the system to escalate to a premium model (e.g., Claude 3.5 Sonnet or Opus) only if the smaller model fails a programmatic quality check or if the user explicitly requests it with a business justification.
7. Enforce Token Budgets and Hard Caps
Set Employee Allowances: Treat AI tokens like corporate expense accounts. Allocate a monthly token budget per employee based on their role (e.g., $50/month for marketing, $300/month for software engineers).
Automated Circuit Breakers: Configure the enterprise console to automatically throttle or freeze access once an individual or department hits 80% of their monthly allocation, requiring management approval to open the tab further.
8. Audit Context Window Waste
Identify Chat Longevity: Audit how long employees keep a single chat thread open. In modern LLMs, every new prompt sends the entire historical conversation back to the model.
Enforce Thread Expiry: Implement an automated policy that clears chat history or forces a new thread after 5–10 turns to prevent exponential token consumption from bloated context windows.
9. Transition to "FinOps for AI" Governance
Daily Spend Dashboards: Move away from monthly billing reviews. Create real-time dashboards that track daily token spend by department, team, and individual user.
Anomaly Detection: Set up automated alerts for sudden spikes in usage, which often indicate an infinite loop in a developer's script or an unoptimized automated agent running wild overnight.
The Strategic Move From Budgeting to Governance
Strategic failure occurs when a company is forced to fire human workers to pay for unmanaged AI bills. This is a real risk when workforce cuts are used as "the only lever" to offset surprise OpEx spikes.
Managing AI costs is a matter of strategic integrity. It requires a shift from passive budgeting to active governance. By treating AI as a variable resource with measurable unit economics, organizations can move beyond the "Open Tab" fallacy and toward a sustainable data strategy.
For more on the psychology behind these failures, read our Saturday Strategy on the Planning Fallacy in Enterprise AI. For a deeper dive into financial trust, see our guide on AI KPIs and CFO Trust.
COST”>
Monday Ledger Exclusive: The CPQ Tracker Giveaway
To help you transition from unthrottled consumption to financial clarity, I have created and built the AI Unit Economics & Cost Per Query (CPQ) Tracker. This Google Sheets tool is the operational blueprint you need to audit your AI spend and establish unit-economic precision.
I’m giving this away for free this Monday! Keep an eye on my LinkedIn for the download link, or get in touch if you want to start auditing your "open tab" today.
Conclusion
AI is an essential tool for the modern enterprise, but it is not a blank check. The $500 million bill is a warning. Organisations that fail to master their unit economics will find that their AI "innovation" is actually a liability.
I help organisations build the Data Fluency and Agentic AI Foundations required to scale safely. The first step is simple: turn on the spending caps.
