When AI Gets the Numbers Right but the Judgment Wrong

After 30 years in tax and accounting, I have learned to be careful with answers that arrive before the right questions have been asked.

A spreadsheet can balance and still describe a business poorly. A forecast can calculate perfectly and still rest on assumptions that will never survive contact with the market. A tax strategy can produce an attractive number while leaving a business owner struggling to meet payroll.

That is why my interest in artificial intelligence comes with a question I cannot ignore: What happens when a system gets the numbers right but gets the judgment wrong?

As CEO of Theogony Financial, LLC, and through my work in AI tax and accounting consulting, I think about that question from both sides. I see the usefulness of tools that help professionals examine information and develop ideas. I also see how easily a convincing answer can be mistaken for a responsible recommendation.

I welcome the technology. But I want its usefulness measured by the quality of the decisions it supports, including what happens after somebody follows its advice.

A correct answer to an incomplete question

Consider a hypothetical business owner who asks whether buying equipment could reduce taxable income. For this example, assume the purchase qualifies for the tax treatment being considered and the calculation is accurate.

The proposed purchase costs $50,000. An illustrative calculation using a 25 percent tax rate shows a potential $12,500 reduction in tax. The owner sees the savings and becomes interested.

I see another question: Does the business need the equipment?

If the purchase is unnecessary, the owner may be committing $50,000 to obtain a tax benefit worth substantially less. The timing of payments, financing costs, and actual tax treatment would require further review, but the basic business question already matters. A deduction does not turn an unnecessary expense into free money.

I would want to know about cash reserves, upcoming obligations, expected demand, and the equipment’s contribution to operations. If customers are paying slowly and payroll is approaching, preserving liquidity may deserve more attention than reducing taxable income.

The hypothetical AI calculation has answered a narrow question correctly. It has not yet established whether the purchase is sensible.

That gap is where professional judgment begins. I cannot evaluate a tax strategy without considering the person or business that must live with it.

The elegance of a forecast can conceal its weakness

I have a similar concern about financial modeling. Give a system assumptions for revenue growth, gross margin, operating expenses, and capital spending, and it can help construct a detailed forecast. The formulas may be consistent. The presentation may look ready for a boardroom.

But who examined the assumptions?

If I assume revenue will grow 20 percent annually, I can create a model that makes expansion look attractive. That does not establish that customers will buy more, that the sales team can acquire them, or that the business can deliver what it sells.

I want to understand the source of the growth. Is it a signed contract, a tested sales process, a price increase, or simply management’s ambition? What does the business have to spend before the additional revenue arrives? How much cash gets tied up while invoices remain unpaid?

Those questions change the meaning of the forecast.

I also want to know what happens if the plan falls short. A business might survive a modest revenue miss but struggle badly if slower collections occur at the same time. A forecast that overlooks the timing of cash can make a profitable business appear more financially secure than it is.

My responsibility is to challenge the model until I understand what it assumes and where it becomes fragile. A polished projection should invite examination. Its appearance should never substitute for evidence.

Missing information deserves a visible place

Another problem begins when the available information is incomplete.

Imagine someone asking for advice about a rental property without providing its full history. They supply the purchase price and current rent but omit prior depreciation records, periods of personal use, or details about ownership.

An answer can sound thorough while depending on facts nobody has established.

In my work, I want those gaps made visible. I need to know which documents support the analysis, which statements came from the client, and which facts remain unresolved. If an answer changes materially depending on a missing document, that uncertainty belongs in the recommendation.

This is also how I think about evaluating AI work. I do not want to reward a response merely because it reaches a tidy conclusion. I want to examine whether it recognized an incomplete record and asked for the information needed to proceed responsibly.

Sometimes the strongest response is conditional. Sometimes it is a request for a document. Sometimes the calculation should wait.

I regard that restraint as part of competence. A professional who recognizes the limits of the evidence gives the client something valuable: a clearer understanding of what can actually be concluded.

Even reading an analog clock can expose the unevenness of AI’s abilities. Researchers have found that some general-purpose models struggle to interpret clock hands accurately, while specialized systems perform better. I find that revealing: success on a complicated task does not guarantee competence on a seemingly simple one. In my professional work, that is another reason to verify the particular answer in front of me rather than rely on a system’s reputation for intelligence.

Confidence is easy to read and difficult to audit

I understand the appeal of a clear answer. People come to professionals because uncertainty is tiring. A business owner already managing employees, customers, and bills may want someone to explain what to do next.

AI can present information in language that feels decisive. That makes it especially important for me to distinguish clarity from certainty.

When I review a recommendation, I ask what supports it. Can I trace the figures to the records? Have the assumptions been stated? Does the conclusion depend on a fact that was guessed? Would another reasonable interpretation change the result?

These questions matter even when every arithmetic step checks out.

For example, an expense can be entered accurately but classified incorrectly. A financial report can faithfully reproduce a source file while carrying forward the source file’s errors. Checking the totals will not resolve those problems.

I want review to reach beneath the surface of the answer. Otherwise, speed simply helps an overlooked mistake travel farther.

Responsibility continues after the recommendation

The philosophical question that stays with me is this: What does intelligence mean when it cannot bear responsibility for its advice?

I am not suggesting that a tool must become a person before it can be useful. I am asking how we should assign responsibility when a person uses its output to make a consequential decision.

If I present a recommendation to a client, I remain responsible for my professional work. I cannot excuse a weak analysis by saying the software produced it. The client relied on my review and judgment.

That obligation requires me to understand enough of the reasoning to explain it. I need to be able to defend the assumptions, describe the limits, and identify what additional information could change my view.

I also need to consider the client’s priorities. Two business owners can face similar numbers and make different reasonable choices. One may be planning an expansion. Another may be preparing to retire or supporting a family member whose care makes financial stability especially important.

A recommendation becomes useful when it takes those circumstances seriously. The client’s life is part of the decision.

How I want AI to support my work

I see a productive role for AI in preparing questions, organizing information, exploring scenarios, and helping draft explanations for review. Those uses can create more room for the work that requires careful attention.

But I want the process to make verification easier.

Before relying on an output, I want the underlying records identified. I want material assumptions stated in plain language. I want missing information separated from established facts. Where practical, I want important calculations checked independently.

I also want the recommendation tested against alternatives. What if the client delays the purchase? What if revenue grows more slowly? What if collections deteriorate? A useful analysis should help me understand how the decision changes when circumstances change.

The amount of review should match the consequence of the decision. Drafting a routine explanation and recommending a major financial commitment require different levels of scrutiny.

My goal is to use the time technology saves to ask better questions and examine the answers more carefully.

The standard I intend to keep

I have spent enough years working with financial information to know that the most important issue is sometimes absent from the first question a client asks.

A request to lower taxes may reveal a cash problem. A request for a forecast may reveal an expansion plan that has not been tested. A request for a quick answer may reveal records that still need to be reconciled.

I want AI to help me notice those issues. I want it to strengthen the work between the initial question and the final recommendation.

The numbers matter. I have built my career around understanding them. But numbers acquire meaning through the circumstances they describe and the decisions people make with them.

Before I recommend a course of action, I need to ask whether the evidence supports it, whether the client understands it, and whether I can stand behind it.

Getting the calculation right is essential. My work continues until I understand what that calculation means for the person trusting me with the decision.