Why ‘Good Enough’ AI Falls Short In High-Stakes Professional Work
The rise of large language models (LLMs) has put powerful tools on every desktop. But power is not the same as trustworthiness, and fluency is not the same as accuracy. For work where a single fabricated citation or a misread regulation can trigger real harm, a new and stricter benchmark is emerging. Thomson Reuters calls this higher benchmark Fiduciary-Grade AI — AI engineered specifically for professionals who carry duties of care and regulatory oversight.
The gap general-purpose AI cannot close
Many general-purpose models are trained on broad datasets that can include public web content, licensed material and other sources. They are remarkable at generating plausible-sounding text, summarizing documents, and answering broad questions. Content drawn from the open web is not always an authoritative source — it can mix accuracy, opinion, outdated material and error. When a professional relies on a model grounded in a broad and uneven information environment, they can inherit that uncertainty and remain fully responsible for the outcome.
This is the core problem. A professional with duties of care or regulatory oversight cannot delegate accountability to a system they cannot trace, verify, or defend. General-purpose AI can produce outputs that are difficult to attribute and hard to independently verify, which may make it useful for a brainstorm but inadequate as the foundation for work that must withstand scrutiny from a regulator, a court, or a client. That risk is not theoretical. In Mata v. Avianca, a federal judge sanctioned two New York attorneys in 2023 after they filed a brief citing six court decisions that a general-purpose chatbot had fabricated.
Four principles that separate trusted AI from the rest
Thomson Reuters frames Fiduciary-Grade AI not by what it produces, but by how it is built, what it is allowed to access, and how its outputs can be governed. The company’s framework outlines four design principles that it says separate Fiduciary-Grade AI from general-purpose systems:
1. Grounded in authoritative content. Substantive outputs must derive from curated, domain-specific content — not information scraped from the open internet. Every material output should be traceable to a source that a qualified professional can independently locate, cite, verify, and trust. If the AI cannot show its work in a way a professional can check, it does not meet the standard.
2. Meaningful human expertise. Credentialed subject-matter experts must be involved in the development and ongoing oversight of the system — not merely consulted after the fact. This is the difference between a system designed and tested around a profession’s terminology, workflows and risk considerations and one that merely mimics its language.
3. Privacy and security by design. Data protection must be a structural feature of the architecture, not a policy layered on top. Client confidentiality is non-negotiable in this kind of work; an AI that sends sensitive matter data into a shared training pipeline may be difficult to deploy responsibly, no matter how capable it appears.
4. Designed for transparent, verifiable reasoning. Outputs must be traceable, verifiable, and defensible by the professional responsible for them. The human remains in charge — the AI’s role is to extend professional judgment, not replace it or obscure it.
Each of these is a design constraint, not a marketing claim, and Thomson Reuters says each is testable. Together, they point to where general-purpose AI can fall short of what professional work demands.
How to evaluate fiduciary-grade AI: a buyer’s checklist
The hardest part of buying professional AI is that a polished demo can hide almost everything that matters. A model that drafts a flawless memo in a controlled setting may still pull from unverified sources, leak confidential inputs, or produce outputs no professional can defend. Professionals with duties of care or regulatory oversight need to probe the architecture underneath before signing. This checklist is a place to start:
- Test with a real matter, not a vendor script. Hand the system a redacted but representative work sample. Does it cite sources you can pull up and check? Does it flag uncertainty, or paper over it?
- Watch for stage-managed demos. A system that performs well only on a vendor’s curated examples may not be fiduciary-grade.
- Put data handling in writing before procurement. Ask whether your inputs train shared models, where matter data resides, and what happens to it when the contract ends.
- Reject vague assurances. General claims about “enterprise-grade security” are not answers on their own; ask for concrete, contractual commitments on data handling.
- Ask for evidence of expert involvement. Look for how professionals shaped the system and how it stays current as law or regulation changes — not just a name on a slide.
- Verify citations yourself. Probe the sourcing on a few outputs and confirm a qualified person can locate and verify each one.
The goal is to confirm the system can survive the same scrutiny your own work is held to.
Learn how Thomson Reuters applies these principles in CoCounsel — the AI assistant built on authoritative content, shaped by experts, and engineered so every output can be traced, verified, and defended. Because in high-stakes professional work, “almost right” is simply wrong.
