Managing AI Agents: Why the Best Results Come from Asking, Not Telling

Artificial intelligence has quickly become a fixture in professional services, and tax due diligence is no exception. Firms are experimenting with AI agents to automate research, analyse large datasets, draft reports and streamline workflows.

Yet as many organisations are discovering, successfully managing AI agents requires a different mindset than managing traditional software.

The instinctive approach is often to tell an AI agent exactly how to perform a task. After all, that is how we have historically interacted with technology. We create detailed process maps, define every step and expect the system to execute instructions precisely as written.

However, AI agents are not simply software tools. They are reasoning systems capable of evaluating objectives, identifying dependencies and recommending approaches. As a result, one of the most effective ways to manage them may be surprisingly counterintuitive: instead of telling them how to do the work, ask them how they think the work should be done.

I recently experienced this firsthand while developing an agentic workflow for software development.

My initial approach was straightforward. I created a detailed process describing exactly how the AI agent should execute the development lifecycle. I specified the sequence of activities, defined checkpoints, outlined documentation requirements and dictated how information should flow between stages.

The results were acceptable, but they fell short of what I had expected.

The agent followed instructions diligently, yet it frequently encountered friction points that I had unintentionally created. In some cases, the process was overly rigid. In others, I had assumed dependencies that were not actually necessary. The workflow reflected my assumptions about how the work should be performed rather than the agent’s ability to optimise the process.

Eventually, I changed tactics.

Instead of presenting a predefined methodology, I gave the agent a simple objective: design an effective lifecycle for agentic software development.

I then asked two questions:

• What information do you need to perform this task effectively?

• How would you recommend structuring the development process?

The difference was remarkable.

The agent immediately identified information requirements that I had overlooked. It requested clarity on stakeholder objectives, testing expectations, governance requirements, feedback loops and deployment constraints.

More importantly, it proposed a workflow that was significantly more efficient than the one I had designed.

Rather than forcing activities into a traditional linear sequence, the agent recommended iterative validation cycles, automated quality gates, continuous documentation generation and structured feedback mechanisms. The process was not only more aligned to the strengths of AI agents, but also more adaptable to changing requirements.

The experience highlighted an important lesson for organisations adopting AI.

Many professionals approach AI agents as junior employees who need detailed instructions. In reality, they are often closer to highly capable specialists who benefit from clear objectives, sufficient context and well-defined outcomes.

The objective remains a human responsibility. What changes is that the agent can often contribute valuable insight into how that objective is achieved.

The role of the human shifts from process designer to outcome owner.

This does not mean relinquishing control. Governance, oversight and professional judgement remain essential—particularly in tax due diligence, where accuracy, compliance and defensibility are paramount.

However, it does mean recognising that AI agents can contribute valuable perspectives on how work itself should be structured. In many cases, they can identify unnecessary complexity, challenge assumptions and suggest more effective approaches than the ones we would naturally create ourselves.

For tax due diligence teams, this has practical implications.

Whether an agent is reviewing transaction data, identifying tax risks, generating workpapers or supporting report preparation, teams should consider starting engagements with questions rather than instructions.

Ask the agent:

• What information do you need to complete this task?

• What assumptions are you making?

• What risks or gaps do you see?

• How would you recommend approaching this analysis?

• What process would produce the highest-quality outcome?

These questions often reveal blind spots, improve efficiency and reduce unnecessary complexity.

There is a certain irony in all of this.

For decades, professionals have been taught that expertise means having the answers. Yet one of the most valuable skills in the age of AI may be knowing which questions to ask.

As AI agents become increasingly sophisticated, the organisations that gain the greatest advantage will not necessarily be those with the most detailed instructions. They will be the ones that learn how to collaborate effectively with intelligent systems.

Sometimes the fastest path to a better process is not telling the agent what to do. It is asking the agent how it would do it.

And occasionally, the answer is better than the one we would have come up with ourselves.

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.