Cloud computing created FinOps because variable infrastructure spending needed active management.
Agents are about to create the same requirement for AI.
The reason is straightforward.
A conventional SaaS licence is relatively predictable. A seat costs a known amount each month.
Agent costs are behavioural.
They depend on:
- how often the agent runs;
- which model it uses;
- how much context it receives;
- how many reasoning steps it takes;
- how many tools it calls;
- whether tasks are retried;
- how much output is generated;
- whether sub-agents are spawned.
Two agents with the same number of users can therefore have radically different economics.
More usage is not automatically better
This is already becoming visible.
Reports in 2026 have highlighted very high token consumption among advanced coding-agent users. That usage can be entirely rational if it creates proportionally greater output, but token volume alone tells a business almost nothing about value.
A million tokens spent preparing a complex regulatory submission might be cheap.
The same million tokens repeatedly summarising an unchanged document might be waste.
Enterprise AI therefore needs a more useful unit of economics.
Examples include:
- cost per successfully completed workflow;
- cost per case resolved;
- cost per report produced;
- cost per qualified sales opportunity;
- cost per software issue closed;
- cost per hour of human effort avoided.
Once AI is measured that way, architectural decisions become clearer.
Model routing becomes a financial control
Not every step needs the most powerful model available.
A sophisticated agent might use:
- a small model to classify an incoming request;
- deterministic code to validate a field;
- enterprise search to retrieve evidence;
- a medium model to extract structured data;
- a frontier model only for the part requiring difficult reasoning.
That is not simply an engineering optimisation.
It is margin management.
At large scale, intelligent routing may determine whether an agentic workflow is commercially attractive.
Memory can reduce repeat spending
Persistent memory also has a direct cost dimension.
Without memory, agents repeatedly reconstruct context.
They re-read documents.
They ask users for information already supplied.
They process the same project history.
They rediscover the same decisions.
A good memory layer can reduce that duplication by retrieving concise, relevant state.
This makes memory an economic component of the architecture, not only a user-experience feature.
The dashboard enterprises will eventually need
Every production agent should arguably expose four classes of metric:
Value
- workflows completed;
- human hours avoided;
- cycle-time reduction;
- revenue influenced.
Quality
- accuracy;
- exception rates;
- corrections;
- successful approvals.
Risk
- tool actions;
- policy violations;
- escalations;
- human interventions.
Cost
- tokens;
- model spend;
- tool/API spend;
- infrastructure spend;
- cost per completed outcome.
The organisation can then answer the question that matters:
Is this agent economically better than the process it replaced?
The future of enterprise AI will not be won by the company using the most tokens.
It will be won by the company getting the most business value from each unit of intelligence it consumes.
Sources
- Microsoft Research, Efficient AI applications: context engineering and agents
- Contentstack, The 2026 Agentic Enterprise Report
- Business Insider, OpenAI reveals how much its researchers are spending on AI coding (2026)