Why the AI Does Not Use Its Tools
Many teams connect MCP servers and skills and assume the AI will use them. Our own measurement shows otherwise. What is behind it, and what teams can learn from it.
Anyone developing with AI agents today builds tools for them. An MCP server for code search, one for the ticket system, a few skills for recurring tasks. The assumption behind it: whatever the AI needs, it will fetch.
We measured that.
In a measurement with 45 test runs, the AI called available tools on its own in only 16 cases. That is why ai-boost supplies the context with every prompt, without the AI having to ask.
In all other cases the AI keeps working anyway, just without the knowledge the tools would have provided. That is rarely noticed right away. The code can run and still not fit the conventions and decisions of the project.
How an agent decides
At the start of a session, an AI agent sees the available tools with short descriptions. Whether it calls one of them is decided by the model at every step, based on those descriptions and the task. That leads to three weaknesses.
The model does not know what it does not know. It cannot miss a team convention or a past decision as long as it has no idea it exists. So it does not ask for it.
Its own knowledge seems sufficient. For many tasks the model has a general solution at hand. A tool call costs a step and, from its point of view, brings no visible benefit.
More tools make the choice harder. Every description takes up space in the context window and competes with all the others. The more tools are connected, the less clear it becomes which one is meant when.
Offering versus delivering
Offering a tool means: the AI may ask. Delivering context means: the AI gets the answer before it has to ask.
Agents like Claude Code provide hooks for this. A hook is a program that runs at fixed points in a session. The UserPromptSubmit hook, for example, runs on every prompt before the model sees it and can attach additional context. Whether that context arrives is then no longer up to the model.
Tools still make sense when the AI should look something up on purpose, such as opening a specific ticket. But the basic knowledge about the project does not belong in a drawer the AI first has to open.
What teams can learn from this
Some of it can be put into practice right away, even without an additional platform.
Fewer, clearer tools. A few tools with unambiguous descriptions beat many similar ones.
Do not leave the essentials to chance. Whatever has to apply in every session belongs in the prompt, not in a tool the AI might call. But only what applies to the task, not everything.
Measure tool calls. Logging which tools are actually called quickly shows which ones only exist on paper.
The hook is the easy part
A hook is set up quickly. The hard part is what it should deliver. Sending everything fills the context window with rules that do not apply right now, and the answers get worse. Only the context that fits the task is useful: the right place in the code, the convention that applies here, the ticket the change belongs to.
That takes an index across code, history, tickets and documentation that stays up to date. A check whether the results actually answer the question. And a rollout that works the same on every machine in the team, in every project.
How ai-boost solves it
ai-boost takes care of exactly this part. Code, conventions and commit history are semantically searchable, tickets and documentation are connected and kept up to date, and every ticket is linked to the commits that implemented it. With every prompt, ai-boost selects the context the task needs and delivers it through hooks. After every search, a model checks whether the results actually answer the question. False hits are discarded.
Setup takes one command per machine and one per project, defined centrally for the whole team. The tools remain available on top, for everything the AI should look up on purpose.
That way, quality no longer depends on whether the AI happens to think of asking at the right moment.
Available does not mean used. Context has to be delivered.
The question is not which tools your AI has. It is what it knows when it writes the first line.