# From Copilot to Colleague: Trustworthy Agents for High-Stakes

Joel Hron, Thomson Reuters | AI Engineer World's Fair 2025 | 19:45

Source: https://www.youtube.com/watch?v=kDEvo2__Ijg
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/from-copilot-to-colleague-trustworthy-agents-for-high-stakes
Published: 2025-07-23
Tags: agents, enterprise, evals, guardrails

## TL;DR
- High-stakes professional software needs AI that produces work and makes decisions, while keeping the risk of wrong answers within what users can accept.
- Agentic behavior is a set of adjustable dials for autonomy, context, memory, and coordination, rather than a binary property of a system.
- Existing domain logic, expert knowledge, and proprietary content can give agents useful tools and make them more trustworthy in legal and tax workflows.

## Summary
Joel Hron describes how Thomson Reuters is moving from helpful assistants toward systems that produce professional work, make judgments, and act inside legal, tax, compliance, audit, trade, and risk workflows. Those settings have little tolerance for incorrect results. Hron presents agency as a spectrum controlled by dials for autonomy, context, memory, and coordination. He explains why evaluation is difficult, since even domain experts can judge the same answer differently at different times, and why agentic systems make source tracking and guardrails harder. Thomson Reuters is using its legacy applications as tools for agents rather than discarding them. A tax system can extract documents, map values into a tax engine, apply rules, and validate its work. A legal research system can search sources, compare and validate citations, and produce a cited report. Hron's advice is to build the full workflow early, then learn which parts need improvement, while making use of the company's distinctive content and domain expertise.

## Key ideas
### Professional AI is being judged by the work it produces
[00:38](https://www.youtube.com/watch?v=kDEvo2__Ijg&t=38s)
Hron says Thomson Reuters began by building assistants whose north star was helpfulness. That standard has shifted toward productivity. Users now expect systems to produce output and make judgments or decisions on their behalf. This matters especially in law, tax, global trade, risk, and fraud investigations, where users do not accept a high risk of being wrong. He frames the move toward agents as a change in what the system is responsible for, not just a change in the interface. The company's customers work in structured professional settings, so an agent has to fit those workflows and their expectations of accuracy.

### Agency is a set of dials that should match the risk of the task
[03:51](https://www.youtube.com/watch?v=kDEvo2__Ijg&t=231s)
Hron defines agentic AI as a spectrum rather than a yes-or-no category. Teams can adjust how much agency a system has based on the use case and on what users will tolerate. For an exploratory task, autonomy may be turned up. For a workflow requiring precision and certainty, it may stay low. The autonomy dial ranges from a discrete task such as summarizing a document to a self-evolving workflow that plans, executes, observes, and replans. Other dials control context, memory, and coordination, from one retrieval source and a stateless interaction to shared persistent memory and multiple agents collaborating.

### Evaluation is difficult because trust expects consistency from variable systems
[07:25](https://www.youtube.com/watch?v=kDEvo2__Ijg&t=445s)
Hron calls evaluation one of the hardest parts of building these products. Users associate trust with certainty and an expected result for a given input, while these systems do not behave deterministically. Human evaluation is variable too. In Thomson Reuters' testing, trained legal experts reviewing the same question and response can show swings of more than 10 percent in accuracy when they repeat the evaluation a week later. Expert review is also expensive, especially when a team iterates every week. Higher agency makes the problem harder because source references can be lost, agent behavior can drift along a trajectory, and guardrails require deep domain knowledge. The team uses detailed rubrics and preference signals to judge whether the system is improving.

### Legacy applications can become the tools that agents need
[10:11](https://www.youtube.com/watch?v=kDEvo2__Ijg&t=611s)
Thomson Reuters has more than 100 years of software development behind its products, including highly tuned domain logic that users expect. Hron says early assistant work often left that logic behind and started again. Agents changed the approach. Teams can decompose legacy applications into components that an agent calls as tools. This turns existing systems from a constraint into useful infrastructure. The approach also gives an agent access to calculations, validation, and other controlled operations that would be difficult to reproduce reliably inside a model. Hron presents this reuse as one of the company's practical advantages in high-stakes domains.

### Building the complete workflow first reveals what actually needs work
[11:09](https://www.youtube.com/watch?v=kDEvo2__Ijg&t=669s)
Hron questions the way teams often interpret MVP. They can spend too much time finding the smallest valuable piece of code before they understand how the whole agentic system behaves. Thomson Reuters found that building the complete system first made it possible to see which components needed focused engineering and which problems were handled by the agentic behavior of the system. This required a mindset change for product teams. His recommendation is to start with the whole problem and workflow, then use what the working system reveals to decide where to invest effort. The point is to learn from system behavior rather than optimize an isolated fragment.

### A tax agent can combine document extraction with controlled tax software
[12:07](https://www.youtube.com/watch?v=kDEvo2__Ijg&t=727s)
Hron demonstrates a tax workflow using fabricated data. The system takes source documents such as W-2s and 1099s, extracts data, maps it to fields in a tax engine, applies tax-law rules and conditions, and generates a tax return. The tax engine gives the AI a controlled tool for calculations. It also contains a validation engine that the system can use to inspect errors, retrieve more information from the documents, and resolve issues before finishing the workflow. The example shows how an agent can coordinate model-based extraction and reasoning with existing software that performs domain-specific calculations and validation.

### Legal research agents need traceable sources and visible work
[14:00](https://www.youtube.com/watch?v=kDEvo2__Ijg&t=840s)
The legal example uses an AI assistant with the tools from a litigation research product. It can search for documents, fetch them, compare citations across cases, and validate citations within cases. The system searches case law, statutes, regulations, legal know-how, licensed content, and other sources while reasoning toward an answer. During the workflow, the model writes notes about what it finds, then combines them into a final report. Hron stresses that the report links to actual cases and statutes through citations. The interface also flags risk associated with citations. The visible trajectory shows how the system arrived at the result instead of presenting only a final answer.

### Unique content and expert knowledge should shape the product strategy
[16:04](https://www.youtube.com/watch?v=kDEvo2__Ijg&t=964s)
Hron says Thomson Reuters asks which assets it has that other companies do not. He points to 4,500 domain experts and more than 1.5 terabytes of proprietary content across its industries. The company also has an applied research lab with more than 200 scientists and engineers and spends more than $200 million a year on AI product development. His advice is to ask the same question of any organization building agentic software: identify the assets that are genuinely distinctive, then use them to create differentiation in the application. For Thomson Reuters, that means combining professional knowledge and owned content with legacy systems that agents can call.

## Notable quotes
- "We're not asking assistants to just be helpful anymore. We're asking them to actually produce output, to make judgments and decisions on behalf of users." (00:54)
- "We view these things as levers that we are able to kind of pull up and down depending on the type of use case and what sort of agency we want to give the system." (07:05)
- "I think we need sort of north stars that guide, and in many ways, we really look at preference at the end of the day to really drive an understanding of are we getting better or are we getting worse." (09:33)
- "It wasn't until we actually built the whole system that we could see the whole system operate and we could understand what components of that system do we need to go spend time on." (11:31)
- "What are the unique assets that you have and how can you perhaps best leverage those to build uniqueness into whatever applications it is that you may be doing?" (17:17)

## Tools & references mentioned
- Thomson Reuters
- Y Combinator
- ThoughtTrace
- FedRAMP
- ISO
- W-2
- 1099
- CISA

## Who should watch
- You are building AI features for legal, tax, finance, government, or another workflow where an incorrect answer has professional or regulatory consequences.
- Your team is deciding how much autonomy, memory, retrieval, and tool access to give an agent instead of treating agency as a single product setting.
- You have mature software, domain experts, or proprietary data and want to understand how they can become useful parts of an agentic system.

## Related talks

- [Navigating AI's Frontier in 2025](https://aietalks.com/talks/navigating-ais-frontier-in-2025) (Grace Isford, Lux Capital, 17:55)
- [Why (Senior) Engineers Struggle to Build AI Agents](https://aietalks.com/talks/why-senior-engineers-struggle-to-build-ai-agents) (Philipp Schmid, Google DeepMind, 10:40)
- [Scaling Agents for Gen AI Products](https://aietalks.com/talks/scaling-agents-for-gen-ai-products) (Anju Kambadur, Bloomberg, 19:38)
- [Agentic Enterprise: What Your CEO Must Know About AI](https://aietalks.com/talks/agentic-enterprise-what-your-ceo-must-know-about-ai) (Hubert Misztela, Novartis, 28:04)
- [Agentic Engineering: Working With AI, Not Just Using It](https://aietalks.com/talks/agentic-engineering-working-with-ai-not-just-using-it) (Brendan O'Leary, Kilo Code, 27:03)