Tempest AITempest AI: Helping you understand the financial impact of when AI works for your business, and when it doesn't.
If your agents really run your business, show me the inference economics.
ScrollBetween the business and the AI.
I work between the business, its users and what AI can actually do, and I make sure the numbers still work once it's running.
AI viability & inference economics
Whether an AI use case pays, what a reliable result really costs, and what happens when the conditions change.
AI workflows & automation
Workflows built across multiple technologies, not just business APIs, from data pipelines and MCP servers to CAD.
AI adoption & training
Helping teams put AI to work, from prompt-based workflows in tools like ChatGPT and Claude to new ways of working.
From requirement to solution
Working between the business, its users and AI capability: understand the problem first, then build what solves it.
Cheap execution is not necessarily cheap operation.
Token price is an input, not the economic outcome. The unit that matters is the cost of a reliable business outcome.
Failed runs still cost money, and so does human correction. A cheap model that needs repeated attempts can cost more per accepted result than a pricier one that gets it right first time.
Inference Economics is not a forecast that prices will rise. It is a stress-testing discipline: the aim is not to prove AI businesses fail, but to find which AI dependencies are economically robust and which are being mispriced.
Don't ask only whether AI can do the work, or how cheap the tokens are. Ask what it costs to produce a reliable outcome, how dependent the business becomes on the inference that produces it, and what happens when those conditions change.
Where AI economics break.
Five ways to look at an AI-dependent business, each with the question it has to be able to answer.
Falling inference costs do not guarantee falling AI operating costs.
Cheaper intelligence makes it rational to use far more of it: longer tasks, more reasoning, more agents, more context, more checking. When competitors do the same, the inference needed just to keep up can grow faster than its price falls.
During one of my own AI-assisted software builds, I hit usage limit after usage limit, and the finished system still failed the specification I'd written. Plenty of activity, very little accepted work. Most regular AI users will recognise the feeling. High activity is not high yield.
What happens if prices fall, but the workflow needs far more inference to stay competitive?
If the inference disappears, how much of the business disappears with it?
When AI replaces operating capacity, inference becomes a factor of production, priced and controlled by a supplier. A small tail of failures nobody noticed can outweigh a great many cheap successes, and removing people can remove a fallback the business didn't know it relied on.
How far can price, volume, quality and availability move before this process stops producing an acceptable outcome?
Exit is not free simply because another model exists.
AI can make execution dramatically faster while making adaptation slower and more dependent on engineering. Moving provider means reproducing the outcome, not just swapping the API, and running models locally moves the cost rather than removing it.
How quickly, and at what total cost, could another inference source reproduce the outcome you depend on?
Facts should be shared. Interpretations should be attributed. Decisions should be traceable.
Each new AI workflow can quietly rebuild its own view of the same business data, paying again to reconstruct the same reality. Worse, an AI interpretation saved as if it were fact gets picked up by the next system as evidence, until the business is reasoning from its own previous reasoning.
Can you still tell which of your records are facts, which are AI interpretations, and what now depends on them?
Capability, governance and economics: an AI system has to clear all three.
A system can be capable and well governed and still fail on economics. Measure the outcome you can rely on, the human work that remains, and the capability actually delivered, not the model name on the invoice.
If you turned the inference off tomorrow, would you know where the business is exposed?
Start with one workflow.
A practical diagnostic for deciding whether an AI-dependent process is economically resilient enough to scale. One workflow at a time, run as a compact workbook rather than a consulting theatre exercise.
- 01
Map
Where does inference sit in the business outcome?
The workflow, its dependencies, owners and what counts as an accepted outcome.
- 02
Measure
What does the outcome actually cost, and how reliably is it delivered?
Baseline cost per reliable outcome, reliability, human intervention and exposure.
- 03
Stress
What happens when the inference assumptions move?
Price, volume, quality, availability, migration and adaptation stress results.
- 04
Decide
Is the dependency economically acceptable?
A decision record, with residual risk explicitly accepted by its owner.
- 05
Monitor
Which assumptions must stay true in production?
Thresholds, owners, alerts and the triggers for reassessment.
Pass all critical gates first. Then compare economics.
There is no single maturity score. A workflow that fails a critical continuity, governance, data-integrity or recovery gate cannot make up for it with cheap tokens. Each assessment ends in one of six decisions:
The depth matches the stakes: a light review for low-value assistive work, deeper measurement where AI controls revenue, service continuity or irreversible actions.