Why Good Tools Beat a Smarter AI

Ask a chatbot for last quarter’s margin and it may hand you a believable number it made up. Ask an AI that can call a tool, and the tool runs a real query on your data and hands back the real figure, the same one every time. The AI decides what to look up, and a tool someone built carefully does the exact work. When you want answers you can trust, put your effort into better tools, and keep the AI as the one that asks and explains.

A chatbot guesses when it does not know

A chatbot on its own has never seen your books. It has read a huge part of the public internet, but not your invoices, your job hours, or your price list. Ask it about your margin and it has nothing real to work from.

It often answers anyway. In 2025 a team of researchers at OpenAI wrote that large language models “sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty.” They also found these made-up answers “persist even in state-of-the-art systems.” So even the newest models still guess sometimes.

A guess that looks right is the dangerous kind. Nobody checks a number that seems about right. That is a big part of why an AI project needs real work around the model, as I explain in why an AI project costs real money.

Tool calling, in plain English

Think of the AI as a dispatcher and the tools as the trades. The dispatcher takes the call, works out what the customer needs, and sends the right person. The dispatcher never fits the boiler, and the fitter never takes the calls.

In AI, a tool is a small program someone writes to do one job: look up an order, total a month of sales, or price a part. Anthropic’s guide to tool use puts it this way: the AI “determines when to call a tool based on the user’s request and the tool’s description.” It then sends back a request, and your own software runs the tool.

Here is the full trip for one question:

1. You ask. “What was our margin on steel jobs last quarter?”
2. The AI picks a tool. It chooses the margin tool and fills in “steel” and the dates.
3. The tool runs. It reads your real records and works out the figure, using your rules.
4. The AI explains. It gives you the answer in plain words, with the source beside it.

The AI does steps 2 and 4. It is good at those: understanding a messy question and explaining an answer clearly. The tool does step 3, the part that has to be exact.

Good tools give the same answer every time

A good tool works like a calculator. Put in the same thing and you get the same answer every time. When your prices or rules change, someone updates the tool once, and every answer after that uses the new rule. It also holds your specifics: your units, your rules, and your odd exceptions, like the customer who always gets a different rate.

Think of a costing engine, like the one I built for metal castings. It turns estimators’ rules into a price, and the same part gets the same price every time. If I put an AI in front of an engine like that today, the AI could explain the price, compare it to last month, or answer a customer’s question about it. It should never invent the price, because the rules live in the engine.

A tool is where you write down what you want, once, in a form a computer follows exactly.

Good tools also outlast models. New AI models come out often. A new model might change how the answers are worded. It does not change what your margin tool returns, because the tool reads the same records with the same rules. Your effort keeps its value.

The AI can still pick the wrong tool

Tools fix the maths, but the AI still decides which tool to use and what to pass in, and it can get that wrong. Anthropic’s own guide warns that if a question leaves something out, the AI “might guess values you didn’t supply.” Ask about “last quarter” and it might pick the wrong dates.

Careful builders plan for this. They:

• Keep each tool narrow. One tool, one job. A tool that only reads closed jobs cannot mix them up with open ones.
• Check what goes in. The tool rejects a date range that makes no sense, or a customer that does not exist, instead of guessing.
• Show the source. Every answer says which records it used, so a person can check it in one click.

This work is most of what separates a demo from a system you can rely on. AI also comes in more than one kind, and each kind needs different checks, which I cover in is all AI the same thing?

Even with good tools, a person stays responsible for what the AI says. Who carries the blame when it goes wrong is its own question, covered in who is responsible when AI gets it wrong.

Ask every vendor where the numbers come from

When someone sells you an AI system for your business data, ask one plain question: where does this number come from?

A good answer names a record, a report, or a rule you can check. If the answer is “the model works it out,” walk away. That system will give you a believable figure with nothing behind it.

On our AI Builds page we make one promise about this: every answer points to the record it came from, so your team can check it in a click. Tools are what make that possible.

Tools are not always needed. If your team wants help drafting emails or tidying up notes, a general chat tool is fine, and there is no number to get wrong. The tools matter once the answer is a figure someone will act on. For the difference between a chat tool and a system built on your data, see why your team can’t just use ChatGPT for this.

Start with one number

Pick one number your team looks up every week, the kind that takes someone an hour to pull together. That is often a good first tool.

Bring that number. Book a call from the AI Builds page, and I’ll tell you if a tool can answer it and what it would take to build.