AI agents change the economics of automation
When software can take many steps in the background, businesses need to measure not only what it costs to run—but how much useful work actually gets done.
A big forecast is not a business case
Huawei's 2035 outlook describes a 100,000-fold increase in global computing capacity or demand for computing. That is a Huawei forecast about infrastructure—not measured fact, independent consensus or a prediction that each business will use 100,000 times more AI.
From paying for seats to paying for work
AI services may add costs for model use, tokens, API calls, searches, compute, storage, integrations and external services. Agent retries and repeated tool calls can use resources without producing a finished result.
Measure the useful outcome
- What outcome was accepted and completed?
- How often was work corrected, escalated or discarded?
- How much human review and exception handling was required?
- What did usage, integration, oversight and maintenance cost in total?
A useful comparison is full cost per accepted outcome, compared with the current process—not cost per prompt or number of actions.
The “because we can” problem
More reports, messages or research steps do not automatically create value. Start with the business problem, set quality and cost limits, define when a person must intervene, and track whether the result improves the work.
See The Real Cost of Business AI and Which Parts of Your Team's Work Should Move to AI?.