Fiduciary-Grade AI™: What it is, why it matters, and how to buy it

Why a Fiduciary-Grade AI standard

For generations, professional trust has been defined by standards, certification, and fiduciary duty. When someone carries a designation like CPA in accounting or JD in law, we understand both their qualifications and the obligations that govern how they must act.

If we expect AI to take on a meaningful share of professional time, we have to apply the same logic. Just as we assess a person's fitness for a job, we must validate an AI system's fitness for the work it is being asked to support.

That is what Fiduciary-Grade AI is — the Thomson Reuters standard for how AI should work in high-stakes professions. It is AI designed for professionals with duties of care and regulatory oversight — drawing on authoritative, domain-specific content; protected by rigorous privacy and security safeguards; shaped by subject-matter experts; and designed to produce transparent outputs that can be verified.

Part 2. The four principles of Fiduciary-Grade AI

Fiduciary-Grade AI is defined not just by what it produces, but by what it is allowed to access, retain, and rely upon. These four principles are the standard that Thomson Reuters established for the profession — and the standard by which any AI in high-stakes work should be evaluated, whether ours or anyone else’s. Each is a design constraint, not a marketing claim. Each is testable. Each reveals where general-purpose AI falls short of what professional work demands.

1) Grounded in authority, with access to the right context

A Fiduciary-Grade AI system must derive its substantive outputs from authoritative, curated, and domain-specific content — not information scraped from the open internet. Every material output must be traceable to a source that a qualified professional can independently locate, cite, verify, and trust.

Authority alone is not sufficient. AI agents must also operate with the full context required to complete professional work. Only when a system can access, know, and act on the specific data, knowledge, systems, and tools required can it complete the complex, multistep tasks that professional work demands.

What to look for

• A content foundation curated and validated by practicing professionals, not assembled from the open web
• Citations that link directly to the authoritative source, not paraphrased references
• A platform that can act across the systems and tools where professional work actually happens, not just within a single application
• End-to-end workflows — research, drafting, analysis, and review — completed in context, not as isolated artifacts that a professional has to reassemble

The same standard applies in tax and audit. A corporate tax team calculating a provision across dozens of jurisdictions needs AI that reasons over current, authoritative tax code — not a general-purpose model's approximation. ONESOURCE is built to that requirement, grounded in validated tax authority, connected to the systems where that work actually happens, and designed to produce a defensible result a tax professional can verify and sign off on.

2) Protected by imperative data privacy and security

Where privacy is paramount, Fiduciary-Grade AI systems are built to protect it. Privacy and security must be structural features of the architecture — not policy overlays, not configurable options, not commitments that loosen when a frontier model is integrated underneath.

What to look for

• A written commitment that your data is not used to train the vendor's models or any third-party models
• A written commitment that your data is not shared beyond your own environment
• Privacy and security commitments that survive partnerships — confirmed in writing for any frontier-model partner the vendor integrates with
• Enterprise-grade controls, such as encryption, access management, audit logging, data residency, and certifications
• Clear answers on what happens to your data if you switch vendors

3) Built with human expertise, not just human oversight

Professional workflows must be designed, tested, and continuously refined with meaningful involvement from credentialed subject-matter experts in the relevant professional domain. Oversight is not the same as expertise — and a system that has been reviewed by experts is not the same as a system that has been shaped by them.

When ambiguity or risk arises, the system must recognize its own limits and bring in a professional, rather than producing an output that overstates its reliability. Also, customers should have access to real-time human support — because in professional work, transparency and trust depend on it.

What to look for

• A named, sizeable team of credentialed subject-matter experts who shape evaluation, content curation, and workflow design — not just review outputs after the fact
• A system that surfaces uncertainty and brings a human in, rather than generating a confident answer that it cannot back up
• Real-time access to qualified human support, not just self-service documentation
• Evaluation against real-world professional workflows, not only synthetic benchmarks

4) Designed for transparent, verifiable reasoning

A Fiduciary-Grade AI system must provide a reviewable trail of what it did and what it relied on — sufficient for a qualified professional, and where applicable, a regulator, court, or auditor, to evaluate the basis for the output and determine its reliability and defensibility.

Making each step in the planning, reasoning, and execution process visible also serves another purpose: it helps younger professionals learn. A system that shows its work is a system that teaches.

What to look for

• A reviewable trail of every step the system took, the sources it consulted, and the tools it called
• Reasoning grounded in authoritative content, not in unattributed web data
• A citation system that lets a reviewer trace every claim to its source in one step
• Outputs ready for human review, designed to be examined, explained, and defended
• A vendor that is willing to be specific about what the system can and cannot do, and where a human must remain in the loop

Why Thomson Reuters

Here at Thomson Reuters, we did not arrive at this standard from a position of aspiration. We arrived at it from more than 150 years of building the content, tools, and expertise that professionals depend on for their most consequential work. CoCounsel. ONESOURCE. Westlaw. Checkpoint. Reuters News. These are not AI experiments layered on top of general-purpose models. They are the infrastructure of professional trust — and Fiduciary-Grade AI is how we ensure that AI earns a place in that infrastructure.

The scale behind the standard

The Fiduciary-Grade AI standard is grounded in what we bring to the profession at scale. More than 2,600 credentialed subject-matter experts — practicing attorneys, CPAs, auditors, tax specialists, and regulatory analysts — shape the content curation, workflow design, and evaluation frameworks that underpin our products. These aren’t labelers or output reviewers; they’re professionals who define what right looks like in the work itself. Their expertise is backed by authoritative content spanning more than 150 jurisdictions, decades of case law, tax codes, and regulatory guidance — content that no general-purpose model has access to by design.

This is not a marketing claim. It is the reason we can build AI that is grounded in authority, protected by structural privacy safeguards, shaped by genuine domain expertise, and designed to produce verifiable work product. Each of the four principles in this guide reflects something we have been building toward for years.

Part 3. Building the business case 

AI investments need to be defended on the same basis as the work itself — rigorously, specifically, and with evidence. The following is our perspective on how to frame that case, drawn from the principles of our standard for high-stakes AI. The metrics and weightings you settle on will depend on your profession, your workflows, and your risk posture.

Three lenses on value

1. Completed work, not just accelerated tasks 

Fiduciary-Grade AI is grounded in authority and has the context to complete multistep professional work. The value to look for is the portion of a professional workflow the system can complete from start to finish — not how much faster it performs a single task.

2. Outputs that can be verified and defended

AI must provide transparent, verifiable reasoning sufficient for a regulator, court, or auditor to evaluate the basis for each output. Work product that the professional can verify, explain, and defend is itself a form of value — and one a buyer can measure through review burden, citation accuracy, and first-pass acceptance.

3. Capacity returned to higher-judgment work

The system should bring in a qualified professional when ambiguity or risk arises, rather than producing an output that overstates its reliability. When that boundary is well drawn, professional time is returned to the work that most needs human judgment. Buyers can frame this as a shift in time allocation rather than a reduction in headcount.

Three lenses on risk 

1. Accountability that stays human

As AI takes on more of the work, it does not assume any of the accountability — that stays with the professional. A buying decision is therefore also a choice about which AI a professional is willing to put their name behind — and which they are not.

2. The cost of "almost right"

Almost right is not a hypothetical risk. Consider a tax filing that cites a regulation amended six months ago, a contract clause that references case law from the wrong jurisdiction, or an audit finding grounded in a standard that has since been superseded. In all those situations, the professional accountable for the output — not the AI — bears the consequences. The asymmetry between the upside of a right answer and the downside of a wrong one is the most crucial factor in any business case for AI.

3. Data exposure that compounds

Privacy and security should be structural features of the architecture, not policy overlays. The risk to weigh here is the difference between commitments written into the contract and those that survive partnerships, and commitments that can be reconfigured later. The first kind is buyable. The second is not.

Part 4. A buying path 

There is no single right buying process. The path below is the one we recommend based on our experience working with professional firms and departments that already use our AI technology. Treat it as a perspective, not a prescription:

• Start with the work, not the tool. Identify the workflows where the consequences of being wrong are highest and where professional time is most concentrated. Anchoring the evaluation in workflows tends to yield clearer decisions than anchoring it in product features.
• Match the category to the work. For each workflow, decide which of the four AI categories in Part 1 fits the stakes. Some work is suited to productivity tools, while high-stakes professional tasks require a Fiduciary-Grade AI system.
• Apply the four principles. Use the principles in Part 2 as your framework for assessing any vendor under serious consideration — in evaluation calls, RFPs, and contract negotiations. Specific answers, written commitments, and a vendor that volunteers the system’s limits are all good signs. The right vendor will welcome the rigor; vague answers are themselves an answer.
• Pilot on real work. Evaluate against actual matters or filings, reviewed by senior professionals. Curated demonstrations are not a substitute. The standard you apply in the pilot should match the standard you will hold in production.
• Put the commitments in the contract. The standard's requirements on authoritative content, structural privacy and security, expert involvement, and verifiable reasoning are the same requirements your contract should reflect. A vendor that meets the standard will be willing to write it down.

This is the standard we build upon at Thomson Reuters, and the standard delivered through CoCounsel for legal, tax, audit, and compliance professionals. It is the bar we hold ourselves to, and the bar we believe these professions should hold any AI to.

Every professional knows the work they would not delegate

Fiduciary-Grade AI is the standard we set for work that professionals would not delegate to anything — or anyone — they could not defend. If you are evaluating AI for professional work, your standard should be at least as high as the standard you set for the work itself. We built ours to meet you there.

Thomson Reuters: chosen by change-makers.

Stay up to date on industry-specific AI insights and the latest CoCounsel developments.