We are in the age of autonomous agents, writes Aaron Harris, chief technology officer of Sage.
Recently, I, like many others in tech, have been experimenting with always-on AI agent frameworks like OpenClaw. In one week, I burned through over 220 million tokens. The invoices piled up so quickly that I did the only logical thing: I built an accounting agent that I named Arthur, to keep track of all the money my other agents were spending.
At first, I was impressed. Arthur made quick work of the invoices, entering them in a spreadsheet, recording amounts, and categorizing expenses. But then I checked his work. The ledger was a mess. He’d inexplicably changed expense categories from “AI Expenses” to “Cloud Services,” deleted what he assumed was a duplicate invoice, and formatted dates as text strings so the spreadsheet couldn’t even read them. When I asked him to self-assess, Arthur gave himself a 5 out of 10. His own brutal summary: “A spreadsheet this broken would fail any basic formula or pivot table.”
Arthur’s story is an example of a serious issue playing out across businesses. In the rush to adopt generative AI, a dangerous assumption has taken hold: that the primary barrier to deployment is a deficit of machine intelligence. This overlooks the critical importance of systemic reliability. An AI that is impressive in a demo but unreliable in a real workflow is worse than no solution at all. In finance, “almost right” has always been wrong.
Defining the Role of Trust in AI Deployment
Most AI experiments are impressively capable in demos, yet dangerously unpredictable in practice. This isn’t a sustainable model for business-critical functions, and it’s creating a drag on adoption. Recent research Sage conducted with IDC found that finance leaders are spending more than 15 hours a week just validating AI outputs. That’s nearly two full workdays that AI was supposed to give back.
The problem runs deeper than lost productivity. The same research revealed that 70% of CFOs would veto an AI-generated result, even if it were accurate 99% of the time, if they couldn’t understand the reasoning behind it. This isn’t resistance to technology; it’s a rational response to risk. For a business owner, an accountant filing on behalf of a client, or a small business that can’t absorb a compliance failure, the stakes are simply too high. The cost of getting it wrong could be a failed audit or a financial penalty – or worse, a loss of credibility.
The core tension is that accounting and software are systems built on invariants; things that must always hold true. Debits must equal credits. Code with a misplaced semicolon won’t compile. Generative AI, by contrast, is probabilistic. It makes predictions, not promises. The challenge isn’t the agent’s ambition. It’s the absence of systems that govern how agents work and ensure they operate within the deterministic world of finance.
From Demo to Trusted System: The Pillars of AI Governance
The leap from a clever demo to a trusted system depends on establishing a new architecture; one where trust is the foundation, not an afterthought. This requires turning abstract principles into an operating reality through clear governance.
First, this means traceable actions and explainability. Every action an agent takes must be logged and transparent. When an auditor asks why a transaction was posted, the system needs to provide a complete answer, including the prompt that triggered it, the data it saw, and the reasoning it applied. It must be a glass box, not a black box.
Second is human control and oversight. Humans have to remain in charge. Consequential actions should always require approval, and agents must operate within clearly defined guardrails. When an agent’s confidence is low, its programming must compel it to escalate to a human expert rather than guess. Taking Arthur as an example, he never lacked confidence – that was the problem. Asked how to fix his own mistakes, he reasoned his way to a perfectly sensible answer and then recommended I sign up for a competitor’s software, with no idea he was talking to the CTO of Sage. He wasn’t wrong to reason it through. He was wrong to guess when he didn’t have to.
This all must be supported by finance-grade governance. You can’t layer on trust after the fact. This means applying the same rigor to AI agents that we apply to humans: robust authentication, role-based permissions, approval workflows, and segregation of duties. An agent should only be able to act with the authority of the user who invoked it.
Building an Operating System for Trust
At Sage, we are building what we think of as an ‘Agent OS’, an operating environment for agents that makes our vision for trusted AI possible. This isn’t just about creating a single smart agent; it’s about building the systems that govern how all agents operate whether they are built by Sage, our partners, or our customers. It’s the platform that ensures every agent has the right skills and tools for the job. Even the latest LLMs with PhD-level math skills try to do calculations in their head; we’d prefer they use a calculator. Our Agent OS provides those deterministic tools.
This is the control layer that makes Authentic Intelligence possible. And these aren’t just compliance box-ticking exercises. They are the essential bridge between AI’s potential and its practical, reliable application in the real world.