Cheap technology is not a cheap task
Model fees are one component. Preparation, review, permissions and downstream correction may remain human work. Without changes there, tool efficiency may not become an operating result.
AI task-value discovery and use-case selection.
Start AI investment with what has changed in task economics.
Cheaper inference makes more tasks feasible to try. Feasible and worthwhile remain different. We believe the enterprise must compare the value of the whole job: whether saved time can be used, how errors affect operations and whether the new method moves a workflow bottleneck.
A SIGNAL WORTH EXAMINING>280×
reduction in inference cost at a specified capability level1
2025 · Nov 2022–Oct 2024 · GPT-3.5-level performanceTHE ZHICE INTERPRETATION
As models become cheaper, worthwhile problems become the scarce resource.
Falling inference prices repeatedly redraw the economic boundary of tasks. They do not remove the costs of process integration, human review or error handling at the same pace. We believe AI investment should move from a fixed tool list to a revisable task portfolio, reassessing automation and human control as capability, prices and consequences change.
THE ARGUMENT BEHIND THE VIEW
A judgement needs a complete argument.
Start AI investment with what has changed in task economics.
EXAMINING AI THROUGH THE TASK
THE REASONINGThe business conditions behind the view
Value often sits in the adjacent step
Faster drafting does not necessarily shorten approval; better prediction does not necessarily improve trading. The connection between a task and subsequent action determines whether an improvement is usable.
Sequence changes the value of learning
Related tasks can reuse knowledge, evaluations and interfaces. Selecting isolated demonstrations can miss clusters that accumulate enduring capability.
IMPLICATIONS FOR LEADERS
What should leaders reconsider?
Compare the complete task chain
Examine preparation, processing, review and execution together. Identify where time is consumed and outcomes constrained before assigning work to AI.
Prioritise verifiable business improvement
Consider frequency, consequences of error, assessability and value together. Valuable but hard-to-verify tasks require a different investment rhythm.
Keep the portfolio revisable
Update priorities as models, prices and constraints change. Previously uneconomic tasks may become attractive; deployed tasks may warrant narrower scope.
WHEN TO RECONSIDER
Where the binding constraint is regulation, physical capacity or an immovable approval policy, AI may not be the next best investment. Establish what technology can actually change.
What action will the organisation take differently after this task becomes faster, and which outcome should improve?
FROM PERSPECTIVE TO PRACTICEPut this perspective to work.
ExpandCollapse
Put this perspective to work.
Demos and departmental requests can dominate use-case lists without establishing volumes, rework, expert review or failure consequences. Faster drafting may create more downstream review. Value needs to be estimated across the complete task chain.
Reconstruct the work
Select tasks such as maintenance knowledge support, contract checking or operating analysis. Record triggers, inputs, steps, review and exceptions, and verify effort using samples.
Compare human-machine designs
Compare rules-based automation, AI assistance and controlled execution, considering acceptable errors, data permissions and learning costs rather than assuming model involvement everywhere.
Build a testable portfolio
Select pilots by value, feasibility and risk. Specify baselines, full costs, evaluation periods and continuation conditions; identify missing capabilities for deferred use cases.
Work products
Task-chain and workload baseline
Use-case value, readiness and risk portfolio
Pilot charter and continue/stop criteria
What to measure
- End-to-end handling time
- Include preparation, generation, human review and rework, comparing at equivalent quality.
- Cost per accepted task
- Include model, integration, knowledge maintenance, human review and operating costs.
Initial value estimates are investment hypotheses and require representative-task validation before becoming business targets.
Data notes and research foundationsExpandCollapse
- 1Stanford HAI · AI Index 2025
Comparison of inference cost at GPT-3.5-level MMLU performance; not a reduction in total enterprise AI or project costs.
Stanford 2025 · Nov 2022–Oct 2024 · GPT-3.5-level performance - 2Stanford HAI · AI Index 2026
OSWorld is a structured computer-use benchmark; the result is not an energy-production success rate or a deployment threshold.
Stanford AI Index 2026 · OSWorld benchmark - 3IBM Institute for Business Value · 2025 CDO Study
Self-reported status in a cross-industry, cross-geography survey, not a measured maturity score for energy enterprises.
IBM 2025 · 1,700 senior data and analytics leaders globally
THE NEXT CONVERSATIONThe question worth discussing:
Contact ZHICE