Getting the Correct Answer From AI

Why the best results come from collaboration, challenge and professional judgement – not simply by asking a question and accepting the first answer

The lesson from this example is simple: AI can be extremely useful, but it is not a substitute for professional scepticism. The answer improves when the user supplies context, asks for sources, challenges weak assumptions and keeps testing the reasoning

The question that started the discussion

The issue: Whether a UK trust, holding investment assets, whose income is mandated to its beneficiaries, needs to register under the Automatic Exchange of Information rules, commonly referred to as AEOI or CRS.

That sounds like a specialist tax question, but the wider point is not really about AEOI. It is about how to use AI properly when the answer matters. The initial response from Google Gemini and Microsoft Copilot in this discussion, was essentially the predictable answer: a trust with investment assets may be an Investment Entity and may therefore need to register. That answer was not absurd. It reflected a common starting assumption. But it was not the end of the analysis.

Why the first answer was not enough

The initial answer treated the trust rather generally: investment assets, professionally managed portfolio, potential Financial Institution status. But the key fact was more precise. The income was mandated to the beneficiaries. In UK trust tax language, that can mean the trustees are neither entitled to nor in receipt of the income.

That distinction matters. The CRS Investment Entity test turns on whether more than 50% of the entity’s gross income is attributable to investing, reinvesting or trading in financial assets. The question therefore becomes: what is the trust’s gross income?

How the answer changed

Stage	What happened	Why it mattered
Initial AI answer	Gemini and Copilot initially treated the trust as likely to fall within the Investment Entity analysis because it held investment assets.	This was a plausible starting point, but it did not focus enough on the mandated income point.
Request for sources	The answer was tested against the CRS wording, HMRC guidance and the absence of a definition of gross income.	Asking for sources turned a general answer into a more structured legal and technical discussion.
Challenge to assumptions	The key challenge was whether income mandated to beneficiaries was actually gross income of the trust at all.	This exposed the weak link in the initial response: it assumed the trust had gross income without proving it.
Revised conclusion	The final view became more nuanced: there is a strong argument that if the trust has no gross income, it cannot satisfy the 50% investment income test.	The conclusion changed because the reasoning was challenged, not because the original question was simply repeated.

The important lesson: AI responds to the quality of the challenge

The most useful part of the exercise was not the first answer. It was the dialogue. We asked for the source, challenged the use of undefined terms, pointed out the relevance of HMRC’s own manuals, and forced the AI to separate assumption from authority.

That is how AI should be used in professional work. It is not a magic answer machine. It is closer to a very fast researcher or debating partner. It can spot issues, summarise arguments and change its mind when presented with better reasoning. But it needs direction.

The danger is not that AI gives a wrong answer. The greater danger is accepting a confident answer before testing the reasoning behind it.

What made the final answer better

The final conclusion was better because the discussion moved from labels to logic. The question was no longer simply, “Does this trust have investment assets?” It became, “Does this trust have gross income for CRS purposes at all?

That is a much sharper question. If the trust has no gross income, then its gross income cannot be more than 50% attributable to investing. The arithmetic is straightforward once the premise is accepted. The real issue is whether the UK mandated income rules genuinely leave the trust with no gross income for this purpose.

Practical takeaways for using AI

Do not stop at the first answer, especially where the issue is technical, legal, tax-related or judgment-based
Ask the AI to identify the source of each key proposition rather than relying on a confident summary
Challenge undefined terms. In this example, “gross income” was the critical phrase
Separate the factual assumption from the legal conclusion. Here, holding investment assets was not the same thing as proving the trust had gross income
Use AI collaboratively. The best result came from questioning, refining and testing the answer step by step

Conclusion

The point of this example is not that AI was useless. Quite the opposite. AI was useful because it provided a starting point, then responded to challenge. But the reliable answer emerged only after the reasoning was interrogated. In addition AI’s research capabilities showed that that there was no alternative definition for Gross Income.

For professional users, that is the key lesson. AI should be treated as part of a collaborative reasoning process. Ask questions, ask for sources, challenge assumptions, and be prepared to keep pressing until the analysis is properly aligned with the issue you are actually dealing with.

Important note: This article is intended to illustrate how to work collaboratively with AI. It should not be relied on as advice on whether any particular trust needs to register under AEOI/CRS. That question depends on the precise facts, the governing documentation, the applicable law and HMRC guidance at the relevant time.