The Agentic Revenue CompanyRevenue as a Service
Blog

Reducing and controlling AI costs: rules first, then models

Reducing AI costs does not start with the model. How rules before AI, the right model choice, caching and budget limits keep AI agents in sales economical.

An agent sorts your sales team’s inbox. Every incoming e-mail goes to a large language model: the customer enquiry as well as the automatic out-of-office reply and the delivery receipt. The results are good. The bill at the end of the month is not. If you want to reduce AI costs, you often look for a cheaper model first. The more effective question is: which of these tasks needs a model at all?

Why AI costs decide whether agents succeed

Agentic sales – AI agents in sales – is meant to secure and grow revenue. Whether that succeeds does not depend on impact alone. An agent that achieves a lot but costs more than it brings in is not a success. It is an expensive experiment.

Gartner predicts that over 40 per cent of agentic AI projects will be cancelled by the end of 2027 – due to costs, unclear business value or inadequate risk controls (Gartner, 25.06.2025). This is a forecast, not a measurement. But it shows that agents can fail not only on technology, but also on economics.

Costs are therefore not a side issue you optimise after go-live. They belong in the design from the start.

Reducing AI costs before the model starts computing

Many steps in sales follow clear rules. Whether an e-mail is a bounce, whether a contact has a block, whether a customer is in an approved territory: all of this can be checked without asking a language model. In the REVAGO system, such rules therefore run before the model is called.

This does more than save costs. Rules are faster, traceable and always deliver the same result for the same data. The model is used where it plays to its strengths: when a free-form customer reply has to be understood, classified or summarised. How agents go about this in the inbox is described in AI email triage.

This approach reflects our understanding of an AI-first operating model. AI-first does not mean every step runs through a language model. It means designing the process starting from the result – with rules, classic automation, models and people each in the right place.

The right model for each task

Not every task needs the most powerful model. A short classification step has different requirements from a call briefing that condenses information from many sources.

That is why we do not choose the model for a task by gut feeling, but on the basis of a sample: the candidate models are tried out on it, and a person checks whether the results are good enough for the purpose. Only then is the decision made. That way, “cheaper” does not quietly turn into “worse”.

Two further levers have a direct effect on costs: the length of the output and the depth of reasoning are limited. An agent that only has to return a category does not need a long text or a detailed justification for it.

Caching, budget limits and cost tracking

Many model calls start with the same instructions and the same background knowledge. With prompt caching, this recurring part does not have to be processed in full again for every call.

Fixed budget limits apply to external data sources. One example is the web search that an agent only uses as a fallback when checking for defunct companies, if commercial register data and the customer’s website do not give a clear picture. More on this in Clean up your customer database.

The basis for all decisions is cost tracking per call. It shows what a work step actually costs and where the next improvement is worthwhile. Without this transparency, every optimisation would remain a guess.

Costs are part of value

In the REVAGO model, V stands for value: the effectiveness of each commissioned use case, meaning what it achieves for revenue and what it costs. Schematically: impact minus effort equals value. Model costs are part of that effort. How we make the value of agents assessable is described in Value: How we make the business value of AI agents assessable.

You pay for results – not for activity. More model calls are therefore not a result, but effort that shows up in the economics of a use case. That is why we design costs in from the outset: during implementation, we define with your functional leads where rules are enough, where a model is needed and where a person decides. In operation, we track costs and adjust rules and model choice when tasks or models change.

What this looks like in practice is shown by customer reactivation: there, rules check channel and data situation before a model analyses a customer reply or a person decides on the next step.

Where an additional model check is worth its cost and where a person checks more cheaply is covered in the wiki article on LLM-as-a-judge.

Your next step

Do you know what a single work step of your AI costs today – and which of them could do without a model? If you want to find out how AI agents in your sales organisation can work effectively and economically at the same time, check your REVAGO potential.


← All articles