Uber began 2026 with a problem many organizations might envy: employees were enthusiastically adopting AI, with engineers using tools such as Claude Code and Cursor.
Then the bill arrived.
Uber reportedly exhausted its planned annual budget for AI coding tools by April 2026, prompting spending caps, usage dashboards and an exception process for additional capacity. More concerning was the uncertainty about what the spending had produced. The company could not yet draw a clear line between rising AI consumption and more useful features reaching customers. Uber’s experience exposes a broader misunderstanding: increased AI usage and spending does automatically create business value.
The cost of obtaining a given level of AI capability has generally declined. Yet organizations are adopting more advanced models, processing more information and embedding AI into enterprise workflows. A lower price per token can therefore coexist with a rapidly rising total bill.
However, many organizations still budget for AI like conventional software: buy it, deploy it and pay the license. But AI economics are shaped by three moving variables: unit price, consumption and value. The six myths that follow show how common assumptions can cause organizations to underestimate costs, mismanage consumption and overstate benefits.
Myth 1: AI Consumption is Predictable
The token rate may be visible, but actual consumption is much harder to predict.
Many AI services charge by tokens: small units that may represent a word, part of a word or punctuation. Charges can include the information sent to the model, the response it generates, cached inputs and services such as search or tool use.
Agentic workflows make consumption less predictable. The same request may retrieve different information, call different tools or retry failed steps. A study of coding agents found up to a 30-fold variation in token use across repeated runs of the same task, while higher consumption did not consistently improve accuracy.
Key takeaway: Token consumption is a cost signal but isn’t evidence of value.
Myth 2: The Vendor Price Is the Total Cost
The vendor quote may cover the license, subscription or token rate. It rarely captures the full cost of deploying, operating and governing the AI system.
Even the quoted rate may change. Claude Sonnet 5 launched at introductory rates of US$2 per million input tokens and US$10 per million output tokens. On 1 September 2026, those rates increase to US$3 and US$15. A business case based on launch pricing may therefore understate future costs.
Other expenses emerge during implementation. For instance, an AI contract-review tool may require integration, security and privacy assessments, testing, training, legal review and ongoing oversight. These costs often sit across different budgets and are easily missed.
Assess the investment across its lifecycle:
- Acquire: Contracts, models and licenses
- Deploy: Data, integration, testing and training
- Consume: Tokens, infrastructure, storage and support
- Govern: Security, privacy, assurance and human review
- Change or exit: Migration, redevelopment and control reassessment
Changing solutions may also require integrations to be rebuilt, prompts rewritten and controls retested.
Key takeaway: Budget for the AI lifecycle, not just the initial purchase.
Myth 3: More Advanced AI Always the Right Solution
More advanced models may improve results, but they can also increase cost, response time and operational complexity. The model, deployment method and level of autonomy should therefore match the task. Advanced reasoning may be justified for regulatory analysis or complex security investigations, while routine content summarization may be handled effectively by a smaller model. The goal is reliable performance without paying for unnecessary capability.
Deployment creates another trade-off. With a provider-managed service such as Amazon Bedrock, AWS operates and scales the underlying AI infrastructure, while the organization manages its application, data and controls. A self-hosted model such as Meta Llama provides greater control but shifts responsibility for infrastructure, security and maintenance to the organization.
Autonomy can increase costs further. A chatbot generates a response, while an agent may plan, search, call tools, retry and verify its work. In Anthropic’s research system, agents used approximately four times as many tokens as chat interactions, while multi-agent systems used approximately 15 times as many.
Key takeaway: Match capability, deployment and autonomy to the task.
Myth 4: A Successful Pilot Proves the Business Case
A pilot can show that an AI solution works under controlled conditions. However, it doesn’t show what the solution will cost when hundreds of people use it every day.
For example, an AI-enabled contract-review pilot may perform well with selected documents, a small user group and close legal oversight. At production scale, however, the organization must support larger data volumes, more integrations, ongoing monitoring, and user questions. Here, the business case should be reassessed before expansion, using forecasts that reflect workload growth, support requirements and the cost of maintaining reliable performance.
Key takeaway: A successful pilot proves feasibility, not production value.
Myth 5: Individually Justified AI Investments Create an Efficient Portfolio
A strong business case for each AI initiative does not guarantee an efficient portfolio.
An organization may enable AI meeting summaries in its collaboration platform, purchase a separate note-taking assistant for sales and allow employees to expense other generative AI tools. Each decision may solve a legitimate need, yet the organization can end up paying several times for transcription, summarization, drafting and action-item tracking.
Reviewing AI purchases separately can conceal overlapping capabilities, underused features and tools operating outside the formal inventory. A portfolio may also appear diversified while multiple applications depend on the same model provider, cloud platform or data environment, creating duplicated costs and concentrated risk.
Key takeaway: Review AI investments collectively; identify overlapping capabilities, underused tools and shared dependencies.
Myth 6: AI Efficiency Automatically Creates Business Value
Faster work does not automatically reduce costs or improve results.
Klarna initially reported that its AI assistant handled 2.3 million conversations in one month, cut average resolution time from 11 minutes to under two and performed work equivalent to 700 full-time agents. By 2025, the company was recruiting human support again after its CEO acknowledged that an excessive focus on cost had reduced service quality.
AI’s rising operating costs are also creating an unexpected question: could a person sometimes be cheaper? Nvidia Vice President Bryan Catanzaro said the AI compute used by his team cost more than its employees. Although specialized, the example shows why organizations should compare AI and human work at the task level, including model use, infrastructure, oversight, errors and rework. Quality, volume and speed also matter, since higher AI costs may still be justified by greater scale or capability.
Key takeaway: Efficiency creates value only when it lowers costs, releases usable capacity or improves outcomes without unacceptable effects on quality or risk.
From AI Cost Control to AI Value Governance
The goal of AI cost management is not to minimize every token, license or technology expense. Some organizations may need to invest significantly more to remain competitive, improve services or address emerging risks. What matters is that spending remains visible, intentional and connected to value.
Good AI cost governance is not about spending as little as possible. It is about knowing when to invest more, when to reduce spending and when to change direction. The question is no longer simply, “What does the AI product cost?” It is: “What will it cost at scale and what improved because we spent it?”
