AI agents – that is, software that not only answers tasks but independently works through several steps – are currently on many project lists.Two current reports counter this euphoria with sobering figures and experiences: the analyst firm Gartner warns of cost risks with token-based billing, and a founder publicly reports why he switched off his first agent again.
The token bill comes later
According to Gartner, AI agents can exceed cost expectations by more than 50 percent when billing is per token.A token is a text unit – roughly speaking a word component – that a language model processes on input and output.So payment is not per user or per month, but per processed volume.
That is the crucial difference from classic licensing models: with an agent, the processed volume is hard to estimate in advance.An agent that solves a task in several steps calls the model repeatedly, passes intermediate results, reads up context and corrects itself.Each of these steps generates tokens.What looks harmless in a prototype with a few test cases scales in production operation with the number of transactions – and with the length of the contexts that the agent carries along.
At the same time, Gartner predicts more autonomous IT processes and hybrid AI infrastructures, i.e. operating models that combine cloud services and own environments.The two go together: the more routine is handed over to agents, the more relevant the question becomes of where and under what conditions the computing load arises.
The first agent is often switched off again
The second perspective comes from practice.Snipki founder Jens Polomski has switched off his first AI agent, named Neo, again.He has not abandoned the idea behind it – but he has the concrete implementation.As the central problem he names tool hopping: the constant switching between new tools, which costs companies time and focus without producing any robust benefit.
The pattern is likely to be familiar to many.A new tool appears, a prototype is created in a few days, enthusiasm is great – and three months later the next platform is up for discussion, while the first agent never found its way into a stable process.What gets switched off then is not because the technology failed, but because there was no clearly defined use case behind it.
Two viewpoints, one common finding
The two sources look at different levels: Gartner argues from the perspective of IT budgets and infrastructure, the practice report from the perspective of daily work.However, they meet at one point: the bottleneck is rarely in model quality, but in the missing preparatory work.
Anyone who does not define before the start which task an agent takes on, how often it runs and how its benefit can be measured gets both – incalculable costs and a tool landscape that turns faster than the processes behind it.Gartner's figure of more than 50 percent deviation is a forecast and not an empirical value from a concrete project; it describes a risk, not a law of nature.
What can be derived from this for project planning
From the two findings, some sobering consequences emerge for companies that want to integrate agents into processes or into software development:
- Use case first, tool afterwards. An agent needs a clearly defined task with a defined start and end point.Only from that does it emerge which platform is even an option.
- Calculate the cost model before the prototype. What matters is not the price per million tokens, but the expected number of runs multiplied by the average token consumption per run – including the internal intermediate steps.
- Anchor upper limits technically. Limits per transaction, per day and per cost centre belong in the architecture and not in a works agreement that nobody reads.
- Make consumption visible. Without monitoring that assigns token consumption to individual use cases, a cost deviation can only be recognised on the monthly invoice.
- Limit tool changes deliberately. A change should be a decision with justification, not a reflex to a product announcement.
- Plan switching off as a normal project step. If an agent does not fulfil its purpose, ending it is the cheaper option compared with letting it continue silently.
Hybrid infrastructure as a cost lever
Gartner's reference to hybrid AI infrastructures is more than a side note in this context.If it is foreseeable that certain tasks arise at high frequency and with a consistent pattern, the calculation changes: processing in one's own or in a dedicated environment can become economical compared with pure billing per token.However, this decision can only be made if robust consumption data is available – that is, after a pilot phase with measurement, not before.
At the same time, the following applies: not every task needs an agent.Many recurring processes can be solved more cheaply, faster and above all more traceably with classic automation.
An agent is worthwhile where decisions have to be made under changing conditions – not where a deterministic process is sufficient.
Classification for web and software projects
For web and software projects, this means: AI agents should be treated like any other component – with a defined use case, a calculated operating cost model and measurement points that demonstrate the benefit.A pilot with a narrow scope and hard consumption limits yields more insight than a broad-scale agent landscape whose costs only become visible in retrospect.And switching off an agent is not a failed project, but a result – provided that the requirements for the next attempt emerge from it.
Sources
- Gartner warns of cost risks from AI agents — DataCenter-Insider | News | RSS feed
- Experiences with AI agents: Why Jens Polomski switched off the first one again — t3n.de - News