Your AI bill keeps rising and nobody knows what it has delivered
By Xavier Franco, CEO de Elternativa · Grupo Zemsania
· 6 min read
Written with the help of AI and reviewed by Xavier Franco

In this article
Not long ago, AI was an experiment funded from an innovation budget. Today it is one more budget line: licences, model consumption, vendor agents and in-house agents. If you ask a chief technology officer how much AI has cost this month, the answer usually comes quickly. If you ask what it has delivered in return, not so much.
That gap between what it costs and what it delivers is exactly what AI observability has to close.
AI cost is already managed, but blindly
This is not an impression. The FinOps Foundation's State of FinOps 2025 report, with 861 responses representing around 69 billion dollars of public cloud spend, finds that 63% of teams already manage AI spend, up from 31% the year before. In one year, it has doubled.
But the same report says something revealing: in AI, the priority is not yet to optimise but to understand the spend: allocate it, collect the data, detect anomalies and forecast. In other words, most teams are still trying to find out what is going on.
And Gartner, in its report on best practices for optimising the cost of generative and agentic AI, forecasts that through 2028 at least half of generative AI projects will exceed their budget, because of poor architecture decisions and a lack of operational experience.
In practice: ask for last quarter's AI bill and try to split it by agent or by use case. If you cannot, that is your first problem.
Why the bill is opaque
The AI bill has three features that make it hard to read:
- It is usage-based. You do not pay a fixed price: you pay for every call to a model. A badly configured agent, one that repeats queries or drags along extra context, can multiply the cost without anyone noticing.
- It is scattered. Agents from different vendors, each in its own system: the CRM assistant, the ERP assistant, the flows of a low-code tool and your own agents. Each with its own bill and its own console.
- It does not say what was delivered. You know how many tokens were consumed, but not whether the task went well, whether the answer was correct or whether anyone used it.
The result is familiar: model consumption grows out of control by agent and by use case, and what each agent does ends up in logs that only a specialist understands.
What AI observability is
Observing AI means seeing, in real time, what each agent does: which tools it uses, with which data and which model, how much it costs and with what quality. It is not a monthly report: it is a trace of every operation.
To be genuinely useful, it needs four things:
- Inventory. Knowing which agents you have, whoever built them. Without a living register there is no policy to apply.
- Telemetry. Traces of every operation, with cost, latency and errors. The sensible way is to use open standards such as OpenTelemetry, so you are not tied to a vendor.
- Quality. Knowing how much it costs is not enough: you need to know whether the agent completed the task, whether it used its tools properly or whether it let personal data out.
- Rules. Thresholds for cost, latency and sensitive data that apply themselves.
In practice: start with the inventory. It is impossible to control the cost of something you do not know exists.
Warn first, block later
There is a temptation when you discover the bill: to cut. I understand, but cutting blindly also has a cost: you stop use cases that do deliver.
My recommendation is to go in phases. First, observe: connect the agents and watch what happens for a few weeks. Then, warn: thresholds that alert you when an agent spikes. And only then, block whatever exceeds the limits you have agreed.
And one important nuance: those rules have to be deterministic. If the control of AI spend is done by another AI, you do not have control: you have another opinion. And another bill.
From cost to value
The underlying question is not how much you spend, but what each euro returns. With observability you can finally compare agents: which costs most per task solved, which fails most, which nobody uses. With that you can decide what to scale, what to fix and what to switch off.
And along the way you get something else: evidence. The same traces that tell you what each agent costs can be used to show what it does, with which data and under what supervision, which is what the European Artificial Intelligence Regulation asks of high-risk systems.
In practice: define an indicator per agent that is not just cost: tasks solved, answers validated or incidents avoided. That is what turns the bill into a conversation about value.
Five questions for every AI vendor
Many of your agents were not built by you: they come inside the CRM, the ERP or the service platform. That does not take them off your bill or out of your responsibility. Before renewing or expanding, I would ask every vendor five questions:
- How much does each execution cost? Not the licence price, but the real consumption per task.
- Can I see the traces? What the agent did, with which data and with which model, in a standard format I can bring into my own system.
- Can I set limits? Spend and usage thresholds per agent or per use case.
- How do I measure its quality? Whether it completes the task, whether it uses its tools properly and whether it lets personal data out.
- What happens if I disconnect it? If switching off an agent breaks a process, you need to know beforehand.
If a vendor cannot answer these questions, the problem is not only cost: it is control. And with the European Artificial Intelligence Regulation under way, control is no longer optional for high-risk systems.
In practice: send these five questions to the three AI vendors that bill you most. The answers will tell you where to start.
Where to start
- Take an inventory of AI agents and use cases, your own and your vendors'.
- Connect telemetry for the biggest consumers, without modifying them.
- Split the cost by agent and by use case for one month.
- Set thresholds that warn before they block.
- Review every month cost, quality and value with management.
How we help
At AI Governance we help you see and govern all your AI: a living inventory of agents, knowledge bases and integrations; deterministic rules on cost, latency and sensitive data that first warn and then block; and a dashboard with costs and an agent ranking.
We do it with Regentic.AI, the AI agent observability and governance platform of which we are exclusive distributors in Europe. It connects through OpenTelemetry, MCP, REST or A2A, on any framework and model, without touching your agents.
This article is informative and does not replace your organisation's legal advice.
Do you know what AI cost you this month, and what for?
In a demo we show you how Regentic.AI traces every agent, its cost and its quality, and applies thresholds that first warn and then block.

