For several years, the standard advice for using artificial intelligence has been to keep a “human in the loop.”
That remains sound advice, but it is becoming incomplete. There’s a little more to it than that now.
AI tools’ capabilities have moved beyond short prompts and isolated tasks. Agentic AI can work through multistep assignments, use several tools, analyze documents, prepare drafts, and continue working for extended periods before asking for human input.
This creates meaningful opportunities for CPAs. It also changes the way firms and companies need to think about oversight.
Having staff watch an AI system complete every step would eliminate much of the productivity benefit. At the same time, allowing AI to operate without clear direction, safeguards, or review creates unacceptable risk.
AI governance for CPA firms and organizations should therefore focus less on keeping a person involved in every action and more on keeping people in control of the work.
The better principle is human in the lead.
Human in the Lead Does Not Mean Constant Supervision
Keeping humans in the lead means the organization remains responsible for the objective, the boundaries, and the outcome.
People decide what the AI should accomplish, what information it may use, what actions it may take, and which decisions require professional judgment. They also determine how the work will be reviewed and who is accountable for the final result.
Consider an AI workflow that reviews client documents, extracts relevant information, identifies missing items, prepares a preliminary workpaper, and drafts a client follow-up request.
A staff member should not need to watch every step. Instead, the firm should define the approved source documents, specify the information to be extracted, identify the exceptions that require escalation, and require a qualified professional to review the completed work.
The AI performs the routine steps. The human designs the assignment, evaluates exceptions, and owns the result.
That is what it means to be human-led, AI-enabled, and risk-controlled.
Good AI Work Starts With a Good Assignment
Prompts matter because a prompt is more than a question. It is the assignment given to the AI.
The familiar principle of “garbage in, garbage out” still applies. Vague instructions, incomplete facts, and unclear expectations usually produce weaker results.
A well-designed prompt should give the AI enough information to do useful work. It should explain the objective, provide relevant context, identify constraints, describe the expected output, and state how important claims or calculations should be checked.
Creating prompts like that requires an understanding of the tool itself. Staff need to know what the AI does well, where it tends to struggle, what information it can access, and how it may respond when facts are missing or uncertain.
Even an excellent prompt cannot compensate for the wrong tool, incomplete source material, poor workflow design, or the absence of appropriate review.
Prompts should also be tested rather than treated as permanent templates. A prompt that performs well with one model may behave differently with another. Model updates, new features, and changes in available tools can also affect previously reliable results.
Good prompts remain essential. They simply need to be maintained as the technology changes.
Match Oversight to the Risk
Confidence in AI should not become blind trust.
AI can produce work that is polished, persuasive, and wrong. It may overlook an exception, misunderstand an instruction, use an inappropriate source, or make an assumption without disclosing it.
CPA professionals already have a framework for responding to that problem: professional skepticism.
The level of skepticism and review should reflect the potential consequences of an error. Using AI to organize internal meeting notes is not the same as using it to analyze a tax issue, draft client advice, evaluate financial information, or process confidential data.
Low-risk, reversible tasks may require only a brief review. Higher-risk work may require approved sources, defined checkpoints, exception reporting, independent verification, or final approval by an experienced professional.
As AI systems become more autonomous, firms may have fewer opportunities to interact with them during a task. That makes the quality of the remaining touchpoints more important.
The goal is not maximum oversight. It is the right oversight, at the right time, based on the risk.
Reliable Results Require Training, Process, and Controls
When AI results are inconsistent, the solution is rarely just another prompt.
Reliable AI use depends on several elements working together.
Training helps staff select the right tool, provide better instructions, assess results, protect sensitive information, and recognize when a matter should be escalated.
Processes define where AI may be used, what information it may access, how work should move through the firm, and who is responsible for the outcome.
Controls help confirm that AI-supported work is accurate, secure, documented, and reviewed according to its risk.
When these elements are missing, the symptoms are usually easy to recognize: inconsistent output, repeated corrections, duplicated effort, and staff spending more time repairing AI-generated work than the original task would have taken.
That is not a productivity gain. It is a process problem.
AI Learning Cannot Be One and Done
AI is changing too quickly for a single class to keep professionals current.
New tools appear. Existing products add capabilities. Techniques improve. Model behavior changes. A task that was unreliable a few months ago may now be practical, while a familiar workflow may need to be adjusted after an update.
Firms do not need to chase every new product. They do need a practical way to understand what has changed, which developments matter, and how those changes may affect their work.
If AI use in your firm produces inconsistent results, requires frequent rework, or consumes more time than it saves, your team may need better assignments, stronger skills, clearer workflows, and more appropriate controls.
The SC.CPA AI Mastermind learning cohort helps professionals build those capabilities while staying current with the tools, techniques, and strategies changing from month to month. Participants gain timely guidance on what is worth testing, what is working in practice, and where caution is still warranted.
The future of AI in accounting is not about choosing between unrestricted automation and constant human supervision.
It is about preparing people to lead the work.
