Language models have a peculiar weakness: they're brilliant at language and reasoning but unreliable at exact computation — arithmetic, precise data lookups, current facts. The fix isn't to make them better at guessing; it's to let them stop guessing and use tools.
Why models shouldn't guess
A model predicts likely text, which is the wrong tool for exact answers. Ask it to multiply two large numbers and it produces a plausible-looking result that's often wrong — because it's pattern-matching, not calculating. The same applies to precise data, live information, and anything requiring exactness. Guessing at these is where a lot of AI unreliability comes from.
A calculator never "usually" gets arithmetic right — it just gets it right. The smart move is for the model to know when to call one.
The tool-use answer
The solution is to give the model tools and teach it to reach for them: a code interpreter for math and data, a search tool for current facts, a database query for precise lookups. The model's job shifts from being the calculator to knowing when to use one and how to interpret the result. It reasons about the problem, delegates the exact part to a reliable tool, and incorporates the answer.
Why this is the right division of labor
This plays to each side's strengths. The model is good at understanding the problem, deciding what's needed, and reasoning over results — genuinely hard things. Tools are good at exactness — arithmetic, execution, precise retrieval. Combining them yields systems far more reliable than the model alone: the reasoning is the model's, the exactness is the tool's.
The takeaway
The most reliable AI systems don't ask the model to do everything — they let it orchestrate tools. For any task involving computation, current data, or precision, the design principle is clear: don't make the model guess what a tool can know for certain. A model that reasons about when to compute, and then computes, beats one that confidently guesses every time. It's a core pattern behind trustworthy agents.
