When a company budgets an AI project, it almost always thinks about development: building the assistant, connecting the model, delivering the tool. That makes sense, it is the visible part. But in the projects we support, development is only part of the total cost, and rarely the part that brings bad surprises.

This article reviews the cost items to plan for, then offers a simple method to estimate return on investment before getting started. The goal is not to scare anyone: a well-budgeted project is one with every chance of reaching the finish line.

1. Data preparation

This is the most often underestimated item. A document assistant is only as good as its documents: outdated versions to discard, duplicates, contradictory information, poorly structured files, access rights to clarify. An automation project depends on the quality of master data: suppliers, customers, products.

This work is often done partly by business teams, which makes it a hidden cost: it does not appear in the quote, but it takes internal time. It must be planned explicitly, with named people who are available.

2. Integration with existing tools

A model that works in a demo does not yet create value. It does when it is connected to the CRM, the ERP, email, the intranet, and fits into how teams work. This integration requires connectors, access rights, sometimes changes to existing tools, and tests with real flows.

It is often the difference between a convincing prototype and a tool people actually use.

3. Model usage

Language models are billed by usage, generally by the volume of text processed in input and output. For moderate use, this cost is low. It can become significant as volume grows, when long documents are routinely sent as context, or when an agent chains many calls for a single task.

A few design choices change the bill a lot: sending only relevant passages, using a lighter model for simple tasks, caching what can be cached. The right reflex is to estimate this cost from the prototype stage, based on the real expected volume, rather than discovering it in production.

4. Infrastructure and security

Hosting, a vector database for a RAG system, logs, monitoring, backups, access management. If the data is sensitive and requires European hosting or a model deployed in your own infrastructure, this item grows. It is a justified cost, but it must be known from the start, because it shapes architecture choices.

The cost of an AI project does not stop at go-live. That is where the longest part of its life begins.

5. Maintenance and evolution

An AI system is not software you deliver once and for all. Documents change, processes evolve, model providers release new versions and retire old ones. You need to monitor answer quality, update the index, adjust instructions, handle user feedback.

This recurring item is rarely costed at decision time. Yet a system that is not maintained degrades, and users abandon it.

6. Team adoption

A tool nobody uses has zero return on investment, whatever its technical quality. Training, support in the first weeks, adjusting the tool to field feedback and internal communication are part of the project. It is often the most profitable item: a little effort on adoption completely changes the outcome.

Estimating return on investment

To decide, these costs must be weighed against an honest estimate of the gains. Here is the method we use with our clients.

  • Measure the current situation: how often the task is done each month, how long it takes, who does it, how many errors or rework it causes. Without this baseline, no gain can ever be demonstrated.
  • Estimate a realistic gain, not a maximum one: AI almost never removes an entire task. It shortens it, handles part of it, improves its quality. A cautious assumption is better than a spectacular promise.
  • Count all the gains: time saved, but also errors avoided, shorter lead times, customer satisfaction, the ability to absorb more volume without hiring.
  • Compare against total cost over several years: development, but also model usage, infrastructure, maintenance and adoption. Comparing an annual gain with development cost alone gives a distorted picture.
  • Set the success threshold before starting: for example, “the project is a success if processing time halves on this scope”. This threshold avoids endless debates afterwards.

A concrete example

A support team is considering an assistant to help its agents answer faster. The first budget only covers building the tool. Going through the list, several items appear: cleaning up the knowledge base, which contains many outdated articles; integration into the ticketing tool, without which agents would juggle two windows; model usage cost, estimated from the real ticket volume; and a few hours a month to update the index and track quality.

On the gains side, the team measures average ticket handling time and the number of escalated tickets. The success threshold is set before launch. The final budget is higher than first planned, but it is complete, and the decision is made with full knowledge of the facts. Above all, three months after launch, it is possible to say precisely whether the project delivered.

Our recommendation

Before asking “how much does this project cost?”, ask “how much does it cost over its whole lifetime, and what must it deliver to be worth it?”. Start with a narrow, measurable scope that lets you validate both real costs and real gains before investing more widely. It is the best protection against drifting projects, and the best argument for getting the means to go further.