Microsoft's research team recently drew a stark comparison: uncontrolled AI agents behave like insider threats. They have access, autonomy, and—without guardrails—the potential to cause damage before anyone notices.
If you run an accounting firm or manage operations at one, this isn't abstract security theater. You handle client tax returns, bank reconciliations, payroll records, and sensitive financial documents. An unguarded AI agent processing that data without human approval is a liability you can't afford.
This post walks through what Microsoft's warning means for your firm, why human-in-the-loop automation matters, and how to think about deploying agents safely.
What Microsoft Actually Said—And Why It Matters
Microsoft's framing is direct: frontier AI models—the powerful ones everyone talks about—can't be fully trusted to make autonomous decisions. Not because they're malicious, but because they're unpredictable. They can hallucinate, misinterpret context, or take actions that seemed logical to the model but are wrong in your business.
In a tech context, that might mean an AI agent sends an email it shouldn't or deletes data by mistake. In accounting, it means an agent could misclassify invoices, send unapproved reminders to clients, or—worse—flag a legitimate transaction as suspicious when it isn't.
The core insight: giving an AI agent full autonomy over your accounting workflows is like hiring someone smart but unpredictable, handing them access to client files, and telling them "figure it out." You wouldn't do that with a person. You shouldn't do it with an agent either.
How Unguarded Agents Create Real Risk in Accounting
Accounting firms already operate under scrutiny. Clients trust you with their financial data. Your practice relies on accuracy, confidentiality, and audit trails. An unguarded AI agent can break all three:
- Accuracy slips silently. An agent might process 500 invoices, get 95% right, and you only notice the errors after a client questions their balance. By then, you've lost credibility.
- Confidentiality gaps appear. An agent might send sensitive client information to the wrong email or summarize data in a log that ends up in the wrong place. GDPR and data privacy regs don't care that it was automated.
- Audit trails vanish. If an agent acts autonomously, you can't easily explain why a transaction was flagged, approved, or changed. That's a problem when a client asks or a regulator does.
- Client relationships suffer. If an agent sends a reminder to a client before you've had a chance to review it, or if it flags something incorrectly, you look disorganized—not efficient.
The pattern: autonomous agents move fast. Mistakes happen at scale. By the time you see the problem, it's already affected multiple clients or processes.
Human-in-the-Loop: The Guard Your Firm Needs
The antidote isn't to avoid AI agents. It's to use them the right way: let them handle the repetitive work—chasing documents, entering invoices, drafting routine replies—but keep approval in human hands.
Here's what human-in-the-loop looks like in practice:
- Agent prepares. The AI agent gathers documents, flags missing info, or drafts a client reminder.
- You review. An office manager or team member sees the prepared output in one place, checks it for accuracy, and approves or edits.
- Agent executes. Only after approval does the agent send the email, file the invoice, or move to the next step.
This approach keeps the speed of automation—no more manual data entry or document chasing—but adds the judgment layer that catches mistakes before they become client problems.
Microsoft's research on advancing computer use with oversight models suggests the industry is moving toward this pattern: agents that can reason and act, but within defined boundaries and with human sign-off at critical points.
Three Practical Checkpoints Before You Deploy
If you're thinking about automating workflows with agents, ask yourself:
- What could go wrong? If this agent makes a mistake, could it affect client trust, compliance, or data security? If yes, you need human approval in the workflow.
- Who reviews before it ships? Is there a clear person or role responsible for checking the agent's work before anything leaves your firm? If not, build it in.
- Can you explain every decision? If a client questions why an invoice was coded a certain way or why they got a reminder, can you trace the logic? If the agent acted alone, the answer is no.
These questions should shape how you think about how the automation process works at your firm. Don't just think about time saved; think about where a human needs to stay in the loop.
Starting Small with Guards in Place
You don't need to overhaul your entire operation at once. Start with low-risk, high-repetition tasks—document collection, invoice data entry, routine client reminders—and build approval workflows from day one.
The pattern works best when:
- The agent handles busywork (chasing, entering, drafting).
- A human makes or approves the actual business decision (sending, filing, classifying).
- Everything is logged so you can explain what happened.
This is why human-in-the-loop controls are built in from the start—not bolted on after something goes wrong. It's also why fixed-price packages for automation should include approval workflows as a core feature, not an afterthought.
Trust Requires Transparency
Clients hire accounting firms for judgment, accuracy, and discretion. When you deploy automation without guardrails, you're betting on the agent's judgment instead of yours. That's a bet your clients wouldn't accept if they knew about it.
The firms that build trust in the AI era are the ones that automate busywork but keep humans in charge of decisions. Microsoft's warning is really just a formal way of stating what good practices have always required: if something matters, someone accountable needs to review it before it goes out the door.
For your accounting firm, that means starting with agents as assistants—powerful ones that save hours—not as replacements for your judgment. Build the approval layer first, then turn on the speed.
If you're ready to think through which workflows should be automated and where human approval fits, get in touch to discuss how this works in practice for your firm.