Economics, Business & Finance: How AI can best tutor this subject

What you'll be able to do: Use AI for the part of economics that's model-building and judgement, and recognise the one failure mode that makes this subject more dangerous to learn from AI than almost any other on this list.

What economics assessment is actually testing

Not whether you can shift a curve. Anyone can memorise that a supply increase moves price down and quantity up.

Economics tests whether you can build a model explicitly enough to say what it assumes, read a graph as a compressed argument rather than a picture, tell a testable claim from a value judgement, and say who actually bears an effect rather than just what happens "on net". Accounting tests something adjacent but distinct: whether you can use a formal language, the accounting equation, the T-account, the journal entry, correctly, not whether you can add up numbers. Finance sits on top of both.

The through-line is identification: what, specifically, rules out the obvious alternative explanation for what you're seeing. A rising sales figure after a marketing campaign, a falling unemployment rate after a policy change, a wage gap correlated with a credential, in every case there is a boring alternative story, and the discipline is the set of tools for telling them apart. That should determine how you use the tool: for models, representations and the structure of disagreement, not for the verdict.

What it is genuinely good at here

Translating between representations of one model. Equation, graph, table and plain statement of a demand curve, a production function, the accounting equation, an NPV calculation, these are one object seen four ways, and AI will produce all four on request, tirelessly, for as many models as you want swept.

Running comparative statics on demand, and correctly. "What happens to equilibrium if input costs rise and consumer income falls at the same time" is exactly the kind of compound, patience-testing question AI will work through directly, as many times and with as many variations as you ask for.

Playing a counterparty. The sceptical investor, the demanding client, the competitor undercutting your price, the regulator asking why your merger doesn't lessen competition. Business and finance are argued disciplines as much as calculated ones, and rehearsing the argument against a persistent opponent is where AI earns its keep.

Explaining the plumbing of accounting once the numbers are trusted. Why a transaction hits two accounts, why depreciation isn't a cash flow, why retained earnings connects the income statement to the balance sheet. This is genuinely good tutoring territory, because it's mechanical and rule-governed.

Building the strongest version of a position you reject. Ask it to steelman rent control, protectionism, a competitor's pricing strategy, and it will, which is more than most textbooks attempt.

Generating problems with a single right numeric answer. Elasticity calculations, present value, ratio analysis, tax-incidence arithmetic. Rich, patient, correctable practice.

What it is bad at here, specifically

This is the part that makes economics unusual in this corpus, and it has two faces.

Face one: it will give a settled-sounding answer to a question specialists genuinely dispute. The size of the fiscal multiplier. The employment effect of a specific minimum-wage increase. The right discount rate for valuing climate damage decades out. Whether a given recession was mostly a monetary shock or a real one. Ask any of these and you will typically get one fluent, well-organised, confidently delivered answer, not "this is actively contested, and here is the shape of the disagreement". The prose register doesn't change between a settled fact and a live dispute, so nothing in the sound of the answer tells you which one you got.

Face two: it will confidently reproduce manufactured public controversy on questions where the profession is not, in fact, split. Rent control's long-run effect on the supply and quality of housing is about as close to consensus as empirical economics gets, and it is simultaneously one of the most politically contested policies there is. Who actually bears the cost of a tariff is a question with a fairly settled empirical answer, substantially the importing country's consumers and firms, not the exporter, and it is routinely argued the opposite way in public. Ask AI a politically loaded version of either question and it may hedge toward "there are two sides", not because the evidence is balanced but because the discourse is.

You cannot tell these apart from the outside. Minimum wage sits right next to rent control in the textbook, both are simple price controls in a supply- and-demand diagram, and one of them is a live dispute among labour economists (monopsony power, search frictions, how far above equilibrium the floor sits) while the other is closer to settled. AI will answer both with identical fluency. The only way to know which situation you're in is to ask directly, see method 07, field disagreement mapping.

Numbers. Specific statistics, study results, historical figures, plausible and frequently wrong or stale. Verify before you use one.

Jurisdiction and currency. Which accounting standard applies, this year's tax thresholds, the current policy rate. Generic and often out of date.

Citations. Ask for the paper that found a result and you may get an author, journal and year that sound exactly right and don't exist.

The shape of a good session

1. Classify the question first: before any content, ask whether this is something economists broadly agree on or genuinely dispute, and how you'd check that claim independently of the AI's own say-so.
2. State the model in at least two representations together: equation and graph, or graph and table, before accepting any conclusion drawn from it.
3. Name the ceteris paribus conditions the conclusion depends on, and ask what's being held constant that you didn't say out loud.
4. Separate positive from normative the moment a "should" appears.
5. Ask who bears it, not just what happens on net.
6. If it's contested, get the steelman of the other side before you settle on a position.

None of that is arithmetic, and all of it is what the discipline is actually assessing.

The instruction to set

For this session you are tutoring me in economics. Rules, before answering any question with a policy or empirical component, tell me whether this is a question professional economists broadly agree on or one they genuinely dispute, and say how confident you are in that assessment; keep positive claims (what will happen) and normative claims (what should happen) visibly separate; whenever you state a model's conclusion, state the ceteris paribus conditions it depends on; when I ask who is affected, answer who bears the cost specifically, not just what happens on net; and never give me a specific statistic, study result or citation as settled fact, tell me what to verify and where.

The failure that looks like success

A complete, well-organised, correctly reasoned answer to "does raising the minimum wage cost jobs": delivered in exactly the same voice whether the honest answer is "yes, robustly, above a certain threshold" or "genuinely disputed among labour economists, and hinges on employer market power in the specific labour market". Every sentence is defensible in isolation. The paragraph gives no sign of which kind of question it just answered.

This is the subject's specific version of a general problem: fluency is not calibration. An answer can be well-formed and still misrepresent the state of knowledge behind it, and economics has an unusually large number of questions where the state of knowledge is the actual content of a good answer.

The tell: you've never had the AI volunteer "this is disputed" without being asked, on any topic, regardless of how contested it actually is. Either you've only ever asked settled questions, or it isn't telling you.

What it cannot replace

The professional literature's map of who believes what, on what evidence, and how that's moved over time. A tool with no stake in being right defaults to sounding authoritative regardless of what's actually settled, and only a discipline-native survey of the disagreement, or a person who reads the literature for a living, can calibrate that for you.

It also can't replace real data, real prices, real financial statements with real consequences attached, or the professional norm of citing your source when a claim is genuinely contested. Use it to build and compare models. Don't let it tell you, unprompted, how settled the ground you're standing on is.

Where to go next

01 Ceteris paribus is the foundation everything else sits on: the phrase that carries the whole discipline's worth of unstated conditions. 07 Field disagreement mapping is the most direct answer to this article's central warning, the method for finding out, rather than guessing, whether you're in a settled or a contested question.