Why 'wasted' AI spending might be your best investment

A recent discussion on the AI Daily Brief tackles "tokenmaxxing"—the practice of encouraging high AI model usage. While criticised for promoting waste, the counterargument is that this period of aggressive experimentation is essential for business survival and growth. For organisations moving from simple AI assistance to building autonomous agentic workflows, viewing this initial high cost as a learning expense, rather than an operational inefficiency, is key to staying competitive.

The AI Daily Brief recently provided a counterpoint to the growing backlash against what's been termed "tokenmaxxing". This term describes the internal corporate practice, sometimes encouraged via leaderboards, of racking up high usage of large language models (LLMs).

The debate around AI usage

Critics argue that tokenmaxxing leads to significant waste, poor return on investment, and perverse incentives where employees are rewarded for usage rather than for creating value. The concern is that businesses are burning through cash on AI tokens without a clear, measurable benefit.

However, the podcast argues this perspective misses a crucial shift underway in the enterprise: the move from assisted AI to agentic AI.

• Assisted AI: A tool that helps a human perform a task more efficiently (e.g., ChatGPT helping to draft an email).
• Agentic AI: An autonomous system that performs multi-step tasks and workflows with minimal human oversight (e.g., an agent that manages customer service inquiries from initial receipt to resolution).

According to this view, the high token consumption seen today isn't necessarily waste. Instead, it's the unavoidable cost of research and development as organisations learn how to design, build, and deploy these more complex agentic systems. Companies willing to absorb these initial costs to foster experimentation will learn faster and ultimately build more valuable, productive workflows than their more cautious competitors.

Two mindsets on AI spending

This creates a clear division in strategic thinking for business leaders:

For business owners and operators, this debate reframes AI spending from a simple line-item expense to a strategic investment in future capability. The core insight is that you cannot achieve the transformative potential of agentic AI by applying the cost-control mindset of mature technologies.

Treating AI spend as R&D

The shift to agentic systems is not about optimising existing tasks; it's about redesigning entire workflows. This requires experimentation, failure, and iteration. Framing your initial high token usage as an R&D budget or a staff training cost is a more accurate and productive approach.

Organisations willing to burn tokens on valuable mistakes will outpace those waiting for perfect ROI.

This perspective has different implications depending on your organisation's scale:

• Small to Medium Businesses (SMBs): You can't afford to burn cash indefinitely. However, you can be agile. Your "R&D" might involve a small, fixed budget for a few team members to experiment with automating one or two core, repetitive processes. The risk of not learning how these systems work is that a competitor will, fundamentally lowering their cost to serve and out-compete you.
• Large Enterprises: The risk of inaction is systemic. Falling behind on agentic workflow adoption could mean losing significant market share over the next few years. The challenge is creating a culture that allows for this experimentation. Instead of company-wide token leaderboards, which can be counterproductive, it means creating 'sandboxes'—dedicated teams with protected budgets tasked with exploring high-potential agentic workflows without the immediate pressure of demonstrating positive ROI.

The risk of premature optimisation

Ultimately, the biggest risk is not 'wasting' tokens on experimentation. The biggest risk is prematurely optimising for cost and, in doing so, failing to build the internal knowledge and experience required to deploy AI agents effectively. The companies that learn how to get AI agents to reliably perform complex work will have a durable competitive advantage. That learning process has a cost, and that cost is paid in tokens.

Business leaders should move from passive observation to active, structured experimentation. Tolerating a period of seemingly inefficient AI usage is a prerequisite for long-term gains.

1. Reframe the Conversation and the Budget: Stop treating all AI usage as a standard operational cost. Carve out a specific, time-bound budget for "AI Experimentation and Training." Communicate to your team that the goal of this budget is learning, not immediate efficiency.

2. Establish Pilot Programs, Not just Tools: Instead of just giving everyone a ChatGPT licence, identify 2-3 high-potential workflows for automation. Form a small, cross-functional team to focus on building an agentic proof-of-concept for one of them. For example:

   • An agent to handle initial triage and data gathering for customer support tickets.
   • An agent to research competitors and generate a weekly summary report.
   • An agent to process inbound sales leads and enrich them with public data.

3. Measure What Matters: Ditch simple metrics like 'cost per query' for your experimental phase. Instead, track learning-based KPIs:

   • How many iterations did it take to get a reliable output for a specific task?
   • What have we learned about structuring prompts for multi-step processes?
   • What is the 'brittleness' of our agent? When does it fail and require human intervention?

4. Embrace the 'Cost of Learning': Review token consumption in the context of your pilot programs. High usage within an active experiment is a positive signal of engagement and learning. High usage with no corresponding progress or learning is the 'bad waste' that needs to be addressed. This contextual analysis is far more valuable than a simple leaderboard.

5. Plan a Phased Transition: Recognise that your approach to cost will evolve. Your AI strategy should have distinct phases:

   • Phase 1 (Experimentation): High tolerance for cost, focus on learning.
   • Phase 2 (Pilot): Moderate cost focus, begin measuring process-level ROI.
   • Phase 3 (Scale): Low tolerance for waste, focus on efficiency and optimising proven workflows.