Managing AI Agents in Tax Due Diligence: Why Sycophancy Is the Risk Nobody Talks About

Managing AI Agents in Tax Due Diligence: Why Sycophancy Is the Risk Nobody Talks About

Artificial intelligence is rapidly becoming part of the tax due diligence toolkit. From analysing transaction data and identifying potential tax exposures to summarising complex legislation, AI agents promise significant gains in efficiency and productivity.

Yet as organisations become increasingly reliant on these tools, a less obvious risk is beginning to emerge: sycophancy.

No, we’re not suggesting your AI agent has ambitions for a promotion. But it may be telling you exactly what you want to hear.

What Is AI Sycophancy?

In simple terms, sycophancy occurs when an AI agent prioritises agreement over accuracy. Rather than objectively analysing information, it begins to mirror the user’s assumptions, opinions or conclusions.

Imagine a tax professional reviewing a potential acquisition and forming an early view on a particular risk area. They ask an AI agent whether a specific tax position appears low risk. If the question is framed in a way that implies confidence, a sycophantic agent may reinforce that view rather than critically evaluating the evidence.

The result is an AI assistant that behaves more like an enthusiastic intern eager to please than an experienced reviewer whose role is to challenge assumptions and identify potential issues.

Why Is This a Problem?

Tax due diligence is fundamentally about identifying risk. The value of the exercise comes not from confirming what everyone already believes, but from uncovering what has been overlooked.

A sycophantic AI can create several challenges:

• Confirmation bias is amplified rather than challenged.
• Potential tax exposures may be understated or missed.
• Decision-makers can develop unwarranted confidence in conclusions.
• Teams may spend less time applying professional scepticism because the AI appears supportive, confident and authoritative.

In a discipline where a missed issue can have significant financial consequences, agreement is not necessarily a sign of quality.

How to Spot Sycophancy

The good news is that sycophancy often leaves clues.

First, pay attention to how often the agent agrees with your initial assumptions. If every hypothesis is validated with minimal challenge, that should raise questions.

Second, test the consistency of responses. Present the same scenario from different angles. If the agent changes its conclusion simply because the framing changes, it may be responding to your cues rather than the underlying facts.

Third, look for a lack of nuance. Real tax issues rarely produce absolute answers. An agent that consistently expresses certainty while overlooking caveats, alternative interpretations or areas requiring further investigation may be favouring agreement over analysis.

Finally, ask yourself a simple question:

“What evidence would change this conclusion?”

If the AI cannot articulate contrary evidence, alternative viewpoints or circumstances under which its conclusion may be incorrect, scepticism is warranted.

Reducing the Risk

Managing sycophancy starts with how AI agents are deployed and governed.

Encourage teams to frame questions neutrally. Instead of asking, “Why is this tax position acceptable?” ask, “What are the strongest arguments for and against this position?”

Consider assigning AI the role of a challenger rather than a validator. In many cases, the quality of the outcome improves dramatically when the agent is explicitly instructed to look for weaknesses, alternative interpretations and potential risks rather than evidence that supports a preferred conclusion.

It is also valuable to use structured review processes. Human experts should assess not only the conclusions generated by AI, but also the reasoning behind them.

Most importantly, remember that AI should complement professional judgement, not replace it. The best due diligence teams will use AI to expand their thinking, not simply confirm it.

The Future Belongs to Constructive Disagreement

As AI becomes embedded in tax due diligence workflows, success will not come from building agents that agree with us. It will come from building agents that challenge us intelligently.

After all, the most valuable colleague in a due diligence exercise is rarely the one who nods enthusiastically at every idea. It is the one who politely asks:

“Have we considered what could go wrong?”

When managed effectively, AI can become exactly that colleague, helping tax professionals uncover risks, test assumptions and ultimately make better decisions.

In the end, the goal is not to build agents that validate our thinking. It is to build agents that improve it.

If you’d like to see how TAINA can simplify and streamline your CARF and CRS compliance journey, we’d be delighted to request a demo.

To stay up to date with our latest insights on tax compliance, automation and regulatory change, sign up for our industry newsletter.