In commerce and finance, a lesson is only as good as its numbers. Language models are fluent with words and unreliable with arithmetic. We think that settles how an AI tutor for these subjects has to be built.
A commerce student working through a café’s GST return, a shop’s day book or a bank statement is learning two things at once: the concept, and the habit of getting the figures right. An AI assistant that explains the concept well but slips on a multiplication teaches the wrong habit. Worse, it does so confidently, in the same calm tone it uses when it is correct.
Language models produce text that sounds right. Arithmetic is a place where sounding right and being right come apart. A model can set up a depreciation schedule perfectly and then get one line of it wrong, or apply a rate that is close to the current one but not quite it. For a student, the errors are hard to spot precisely because the explanation around them is so convincing.
In most subjects, a tutor that is occasionally wrong is a nuisance. In accounting, tax and finance, the numbers are the subject. A student who learns to trust a figure because the explanation sounded sure has learned something harmful.
In Teach-Back, the assistant does not do mental maths. Every number comes from a tool call: calculators and rate tables that are built for the job, with the call visible to the student. The model’s job is to decide what to calculate and to explain why; the tool’s job is to produce the figure.
| Step | Who does it | What the student sees |
|---|---|---|
| Decide which calculation applies | The assistant, with the student | The reasoning, in words |
| Look up the rate | A rate table | The rate and where it came from |
| Do the arithmetic | A calculator | The inputs and the result |
| Explain what the result means | The student, in a teach-back | Questions until the reasoning holds |
Getting the number right is half of it. Showing how the number was got is the other half. When the tool call is visible, the student can see the inputs, check them against the case and notice when they set the problem up wrongly. That is the skill an exam, an audit or a job will test: not whether you can multiply, but whether you chose the right thing to multiply.
It also helps faculty. A visible trail of tool calls shows where a student’s reasoning went off course, rather than just a wrong final answer.
A reasonable worry is that a tutor with calculators becomes a homework machine. It should not, and in Teach-Back it is not. The assistant is Socratic: it asks, hints and checks, and it never answers graded work. The tools are there so that when the assistant does show a worked example, the example is correct, and so that when the student works a problem, they practise with the same calculators and rate tables. Our comparison of a Socratic AI tutor and an answer chatbot covers that difference in more detail.
If the answers are vague, the tool may still be good for explaining ideas, but students should be told not to trust its figures. Tax rates and rules also change, so for anything beyond coursework, students should confirm the current position with a qualified adviser.
For how the teach-back loop itself works, see the teach-back method for students. More on the Teach-Back guides page.
Not if the model does the arithmetic itself. A sound tutor for commerce and finance should take every number from a calculator or rate table and show the calculation to the student.
Every number comes from a visible tool call to calculators and rate tables, never from the model’s own mental maths.
In Teach-Back, no. The assistant is Socratic: it asks, hints and checks, and never answers graded work.
See it on your own data. Teach-Back — Learn it, then teach it back. Book a 30-minute working session with an engineer.
