How much is your organisation spending on AI usage?
As teams adopt more agentic workflows, token and inference costs can grow quickly. That is not necessarily a bad thing. Higher AI spend can be a sign that people are actually using these tools and finding value.
The operating challenge is more useful:
How do you optimise AI costs without limiting people's ability to experiment, build and find new productivity gains?
Use the right model for the work
One approach worth exploring is a hybrid model:
- Run open-weight models locally for tasks that do not need frontier-level intelligence.
- Keep API calls to frontier models for the harder problems.
Local inference will not make AI free. Hardware, setup, maintenance and operational trade-offs remain. But the marginal cost of many repeatable tasks can fall dramatically.
Local inference is becoming accessible
The surprising part is how accessible this is becoming.
With high-spec Apple Silicon devices, or purpose-built local AI hardware, teams can set up useful local inference servers without building a full infrastructure platform from scratch.
That creates a practical routing decision for every workflow:
- Does this task need the reasoning quality of a frontier model?
- Is the work repeatable enough to benchmark against a smaller model?
- Do privacy, latency or volume make local execution more attractive?
- What quality threshold should trigger escalation?
For many agentic workflows, "good enough and local" may be the right default. Frontier models then become the escalation path, not the baseline for everything.
The goal is not to minimise AI spend at all costs. It is to match model capability and cost to the value and risk of the work.