What does private AI cost per useful task?

Consider an illustrative manufacturer serving a European dealer network. A dealer asks whether a replacement part will fit an older machine. The employee needs to identify the model and year, find the correct documentation and decide whether there is enough information to prepare a proposal.

A useful result is more than a fluent answer. It refers to the right source, identifies missing information and does not invent compatibility. A missing serial number should prompt a follow-up question, not a plausible guess.

Agree with the team what successful completion means. Which facts must be correct? Which sources may the model use? What can it prepare, and which commitment remains with an authorised colleague? This makes the trial an assessment of the work rather than the quality of a demonstration.

Collect ordinary enquiries and difficult cases: incomplete details, superseded documents, exceptions and questions the available information cannot answer. Give a compact model and a more capable alternative the same tasks and the same permitted sources.

Review the results with the people who understand the work:

• Are the answer and source reference correct?
• Does the model recognise missing information?
• How much checking or correction is still needed?
• Is it responsive enough during busy periods?
• Does the application respect the agreed permissions?

Keep the unsuccessful examples. They help distinguish a capability problem from an information problem. A more expensive model will not automatically resolve inconsistent product records. Google Cloud likewise recommends relating quality and performance to business goals and evaluating design choices through experiments. Google Cloud on AI performance.

Our preference is private AI: a suitable open model in a controlled environment, with responsibility for access and business information clearly assigned. A virtual private server can suit a narrowly defined task. A larger model may need more memory and computing capacity; long documents and concurrent users also affect the requirements. Choose the model and its hosting together. Google Cloud explains how the task, quality requirements and capacity shape this choice.

Where suitable, we consider European models, including those from Mistral, which can run on your own infrastructure. The decision still depends on whether the model performs the work reliably. Mistral on self-hosting.

Bring five cost areas into the comparison:

Cost area	The question to answer
Implementation	Which integrations, source selection and access controls are needed?
Capacity	What does the environment cost at normal demand and at peak usage?
Review	How much time do employees need to check responses?
Correction	How will errors be discovered and resolved?
Maintenance	Who maintains the sources, models, security and tests?

Divide recurring costs by the number of tasks completed to a useful standard. Keep the initial investment visible alongside that figure. This prevents a low price per model request from dominating a decision while employees spend time correcting the output.

Your own server also costs money when it is lightly used. For an occasional task, an external service may be a better commercial fit, provided its data terms meet your requirements. Higher utilisation, particular data requirements or a tightly defined task can favour self-hosting. The answer depends on the application.

For European organisations, model origin and processing location are separate decisions. A European model does not, by itself, mean that data stays in Europe. Specify where the service runs and how access, logging and support are arranged.

Imagine completing 1,000 tasks each month to the same agreed quality standard. Value employee time at €40 an hour. The figures below are illustrative assumptions, not current supplier prices or measured customer results.

On a small screen, scroll horizontally to read the full table.

Here the larger model needs less checking, but the difference does not justify the extra operating cost. The compact model is a better fit. If testing shows that it needs thirty hours of checking and corrections rather than ten, its monthly total rises to €1,500. The decision can change.

Keep the initial development investment visible alongside the monthly comparison and account for differences between the options. Do not count unresolved tasks as successes. An application must meet the required quality standard before a comparison on price becomes meaningful.

If a compact model handles routine enquiries well but struggles with a few complex cases, the largest model need not process every request. You can keep routine tasks within clear limits and refer exceptions to a knowledgeable colleague or a more capable model.

Test that handover too. Does the application reliably recognise when to stop? Who takes responsibility next? Adding another model does not help much if the first step cannot identify an exception.

This does not automatically require an agent. A fixed rule may be sufficient. An agent becomes relevant when completing a task requires different next steps depending on what has been found. Read our approach to private AI and selective use of agents.

A useful trial produces a clear choice: the task to support, the model that meets the requirements, expected operating costs and the point at which an employee takes over. Sometimes the answer is a compact model. Sometimes it is a larger model for part of the work. Sometimes the priority is better information before introducing AI.

The aim is to give employees a stronger basis for decisions and help them serve customers well. It is not to leave them supervising an inexpensive system that creates more correction work.

Start with a handful of actual customer enquiries, the sources your team consults today and the judgements people make. Define a useful answer and the commitments that remain with a person. That provides a more practical starting point than a list of models and subscriptions. We can work through the decision for your organisation as a standalone consulting engagement.