Tech-Zone Article

Are You Outsourcing Your Thinking to AI? What CAs Should Never Delegate

It’s 7.45 pm on a Tuesday.

You’re staring at a messy GST classification issue. The client’s facts are incomplete, the circular is ambiguous, and your junior has just sent a cheerful message:

“Sir, AI has given a clean answer. Looks perfect.”

You open the screen.

The language is polished. The structure is neat. The conclusion is delivered with the quiet confidence of someone who has never lost a single night’s sleep over a notice.

For half a second you feel relieved.

Then a small, uncomfortable thought arrives:

“Have I just outsourced the thinking?”

Most of us have experienced some version of this moment. AI tools have become the new silent partner in many CA offices — always available, never tired, and remarkably good at sounding sure of themselves.

The question is no longer whether we should use them. That debate is becoming pointless. The more important question is whether, while outsourcing some of the work, we are slowly outsourcing the one thing the profession cannot afford to lose: independent professional judgment.

Microsoft CEO Satya Nadella has described this risk in stark terms. When organisations hand over too much of their data, prompts and reasoning processes to external AI systems, he said, they risk effectively outsourcing their thinking.

For Chartered Accountants the parallel is uncomfortable. Because the issue isn’t really about AI. It is about who is doing the thinking when the answer matters.

Assistance is useful. Outsourcing thinking is different.

There is a clean line between the two — at least initially.

Using AI to draft a first version of a letter, summarise a long circular, compare two documents, extract information from a PDF or list possible interpretations of a provision is assistance. The machine does some of the heavy lifting. The CA still decides what is relevant, what is missing, what needs to be verified and what the final position should be.

Outsourcing thinking begins when the machine’s output becomes the conclusion rather than an input into the reasoning. It can happen without anyone deliberately deciding to do it. The answer looks complete. The language sounds authoritative. The explanation appears logical. The junior says, “AI checked it.” And because the answer confirms what we were hoping to find, we stop asking the next question.

That is where the risk begins.

The difference is not technical. It is professional. The final work product, opinion, certification or advice still belongs to the Chartered Accountant. The responsibility does not move to the AI simply because the AI produced the paragraph.

What CAs should never fully delegate

AI can accelerate many parts of professional work. But some responsibilities should remain firmly on the human side of the desk.

Final professional judgments and opinions

Audit opinions, tax positions that may need to be defended, GST classifications carrying rate implications, materiality assessments, going-concern conclusions — these are not merely drafting exercises. They are professional conclusions. AI can help prepare the working paper. It can identify issues. It can suggest questions worth investigating. But the conclusion must still be formed, evaluated and owned by the professional. AI cannot sign the opinion. More importantly, it cannot take responsibility for it.

Professional skepticism and risk assessment

AI can flag an unusual transaction. But an unusual transaction is not automatically a risk. A CA has to ask: Why is it unusual? Is there a legitimate explanation? Does the supporting evidence make sense? Is this an isolated transaction or part of a pattern? Is management’s explanation consistent with everything else we know?

That last question is important. Professional skepticism often begins with a small internal reaction: “Something doesn’t quite fit.” That instinct is developed through experience — through years of messy files, contradictory information and difficult conversations. AI can help you investigate that feeling. It should not replace it.

Ethical and integrity decisions

Independence threats, client acceptance or continuance, conflicts of interest, pressure to soften a finding, or questions about whether a particular course of action is professionally appropriate are not simply optimisation problems. They involve professional ethics and responsibility. A tool may help identify the relevant rules. The decision about what you should do remains yours.

Context-heavy client advice

The best professional advice often depends on information that never appears neatly in a document — the promoter’s history, the client’s business plans, past transactions, commercial relationships, the client’s tolerance for risk, an informal understanding between parties, or a fact the client mentioned casually over a phone call but never put in writing.

AI works with the information provided to it. Professional advice often depends on understanding what has not been provided, what has been left unsaid, and what questions should have been asked in the first place. That is why advice is not merely an answer. It is an answer applied to a particular situation.

Ownership of formal work

If an AI-generated analysis, note or draft eventually enters a workpaper, goes to a client or appears under the firm’s letterhead, the CA still owns the reasoning behind it. The fact that AI wrote the paragraph does not make the paragraph someone else’s responsibility. The tool can produce text. The professional has to decide whether that text deserves to survive.

Where AI is extremely useful — and where it isn’t

None of this is an argument against AI. Quite the opposite.

A CA who spends an hour manually comparing two versions of a 60-page agreement when AI can identify the differences in seconds is not demonstrating professional judgment. The CA is simply doing mechanical work the machine is better suited to perform.

AI can help you:

► Summarise lengthy documents and extract important clauses

► Identify changes between two versions of a document

► Pull dates, amounts, names, clauses and other information from unstructured documents

► Convert a mass of information into a table, checklist or structured format

► Prepare a first version of a letter, note or client communication

► Look for patterns, inconsistencies or unusual items in a dataset

► Suggest questions, risks or areas that deserve further investigation

The common thread is simple: let AI do the heavy lifting. But don’t confuse heavy lifting with decision-making.

The “AI said so” trap

Consider these statements appearing in a professional conversation:

► “ChatGPT says this is correct.”

► “Copilot checked it.”

► “The AI has analysed the entire file.”

None of these statements establishes correctness. AI can produce an answer that is factually wrong, based on outdated information, based on incomplete facts, technically correct but irrelevant to the particular case, based on an assumption that was never stated, or confidently expressed but unsupported.

There is an especially dangerous feature here: the quality of the language can be completely unrelated to the quality of the reasoning. A wrong answer written badly is usually easy to challenge. A wrong answer written elegantly is much more dangerous.

That is why the professional’s question should not be “Does this answer sound right?” A better question is: “What would I need to verify before I was prepared to rely on this?”

That one change in mindset makes AI considerably more useful.

The quiet cost of handing over the thinking

The obvious risks are familiar: an AI error finds its way into a client deliverable, an incomplete analysis is treated as complete, or a professional conclusion is reached without adequate independent application of mind.

There is, however, a quieter cost.

Suppose a junior encounters a complicated issue. Earlier, the junior might have read the facts, formed a preliminary view, looked at the relevant provisions, considered the possible alternatives, and discussed the issue with a senior.

Now imagine a different workflow. The junior gives the facts to AI. AI produces an analysis. The junior edits the language and puts it into the file.

The work may be faster. But what happened to the junior’s reasoning?

This matters because professional judgment is not downloaded into a person at the time they qualify. It is developed through repeated exposure to problems, uncertainty, exceptions and disagreement. If AI consistently removes the difficult thinking from the process, the firm may become more efficient while simultaneously weakening the very capability that makes the firm valuable.

That is a much bigger issue than whether AI occasionally gets an answer wrong.

The bigger risk isn’t hallucination — it’s dependency

Hallucination gets attention because it is visible. Dependency develops quietly.

Imagine a CA firm where:

► staff gradually stop learning the underlying process because AI performs it;

► nobody remembers how to perform an important task without the tool;

► workflows become dependent on one particular AI platform;

► employees cannot explain an important conclusion without opening the AI conversation;

► the firm's accumulated reasoning increasingly exists inside an external system rather than within the firm.

The firm may become extremely efficient.

► Until the AI is unavailable.

► Or produces an unexpected result.

► Or changes how it works.

► Or the underlying information changes.

► Or the client asks the deceptively simple question: “Why?”

That is the test. A healthy use of AI should make a CA faster and more capable. If removing AI makes the CA incapable of understanding or completing the work, something has gone wrong. AI should create leverage, not helplessness. If AI disappears tomorrow, a CA should become slower — not incapable.

A simple rule for AI-assisted CA work

A useful framework is:

Or, even more simply:

Six habits that keep the thinking with you

1. Form a preliminary view before asking AI. You don’t need to have the answer. You simply need to know what you think the problem might be. That gives you something against which to evaluate the AI’s response.
2. Ask AI to challenge itself. Ask what assumptions it is making, what facts could change the conclusion, and what is the strongest argument against the position.
3. Verify what matters. The more consequential the conclusion, the less acceptable “AI says so” becomes as verification.
4. Make the human reasoning visible. A good workpaper should show what the CA considered, what was verified, what conclusion was reached, and why.
5. Don’t let AI become the only way to perform the task. If a process is important to the firm, people should still understand the underlying process even if AI performs much of it.
6. Keep ownership of the intellectual capital. Prompts, client-specific context, corrections, exceptions and professional reasoning are valuable knowledge. If all of that exists only inside an external AI conversation, the firm may be borrowing its own intelligence rather than building it.  

The CA’s value is not the mechanical work

There is an interesting paradox here. The better AI becomes at summarising documents, comparing information, extracting data, preparing drafts and performing repetitive analysis, the less valuable those activities become as differentiators of professional expertise.

That is not necessarily bad news. It may actually be one of the biggest opportunities AI creates for the profession. If a machine can reduce five hours of mechanical work to thirty minutes, the CA has five hours less of mechanical work. What should replace it?

More thinking.
More questioning.
More analysis.
More understanding of the client’s circumstances.
More attention to exceptions and risks.
More time spent explaining consequences rather than merely producing answers.

The value of the CA increasingly lies in the things that remain difficult to automate: understanding context, questioning assumptions, evaluating evidence, recognising exceptions, exercising judgment and taking responsibility for the conclusion.

So the objective should not be to compete with AI at doing the work. Nor should it be to reject AI because it sometimes gets things wrong. The objective should be to use AI intelligently enough that it frees the CA to spend more time doing the thinking that actually matters.

Because the next time a polished AI answer lands on your screen at 7.45 pm, there is one question worth asking before you hit “send”:

“Am I still the one thinking?”

Explore more articles in "Tech Zone"