Why governance lets Thomson Reuters move faster on AI

There’s pressure to get AI working fast, and governance often gets left for later. Build the pilot, prove it works, sort out who can see what data afterward. 

Caitlin Halferty has seen where that leads. Teams get the pilot running, and then a problem forces the governance work they skipped, like finding people can see data they shouldn’t.

“They’ll build a really great data capability, perhaps a pilot, and then they’ll go to bolt on the governance later,” says Halferty, head of data and analytics at Thomson Reuters, in an interview with Digital Journal.

Thomson Reuters is a Toronto-based company whose customers are lawyers, accountants, and tax and audit professionals, the kind of work where being wrong carries a cost. Halferty’s argument is to govern the data first and then build AI on top of it.

“Governance for us has enabled us to de-risk and accelerate our AI transformation,” she said on stage at a Snowflake conference in June.

Getting every system to agree on what words mean

Good data isn’t enough on its own. 

A company can connect every source and lock down who sees what, and its systems can still disagree on something as basic as how many customers it has.

“Unless you also invest in a semantic capability that really makes sense of that data, you still find yourself where you’ll arrive at conflicting answers on number of customers or best performing products in market,” says Halferty.

A semantic layer acts as a shared business dictionary, giving every system the same definition of terms like “customer” or “best-selling product,” so reports and AI tools don’t arrive at different answers.

In Thomson Reuters’ case, Halferty puts the system underneath at roughly 37,000 governed tables and 350 data sources feeding one secured source of truth.

The work began with finance and the team expanded from there. The reason? If finance will file its external reports off a data system, it vouches for the system everywhere else.

She said more than 1,500 people across finance and the business use that semantic layer for everyday decisions. 

Work that took months now runs near real time, and a finance task that once meant waiting on Excel files that took 90 minutes to load is gone.

The agreed-upon data also enables AI agents to move faster. When every source already means the same thing by “customer,” an agent doesn’t stop to sort out the difference.

“Because we focused on it early, we’re able to move faster,” says Halferty. 

Curated, governed data, she adds, is “sort of a gold mine for AI agents.”

Her test for whether the work is real is whether it has reached production. 

“This is finance-validated metrics embedded in our key workflows,” she said. 

A standard set by what the customer can’t get wrong

The standard for the data and the AI built on it is set by the people who buy from Thomson Reuters. Its customers are lawyers, accountants, and tax and audit professionals.

“Their name and reputation is on the line, it can’t be wrong,” says Halferty.

She calls the bar “fiduciary standard”. It means more than 175 years of curated content, thousands of subject-matter experts checking outputs, citations built into the product so users see where an answer came from, and a promise not to train models on customer data.

It also means testing before anything ships. 

Halferty says every AI capability runs through the company’s responsible-AI practice, checked for hallucination and bias, with the security team testing for prompt injection, before a customer ever uses it.

Where to start without the scale

Thomson Reuters has scale, budget, and a long head start on clean data. Canada is pushing businesses to adopt AI, and boards are increasingly asking where to begin. For smaller companies, where’s the right place to start?

Halferty suggests starting by learning what your peers are doing first. Bring in outside help if the skill isn’t on staff. Then find where your own data lives.Thomson Reuters had customer data alone spread across 23 sources before pulling it into a single customer master.

“Figuring out how to pull that together across those different data sources, and leveraging a data platform and consulting capability, to identify what are you going after as an organization, what are therefore the data domains I need to prioritize,” says Halferty. 

Get the data that matters most governed first, and the speed comes from there.

Final shots

• Check whether your systems agree on basic definitions before adding more data or AI on top. If two reports count customers differently, that gets worse with AI, not better.
• Build access rules and governance into a project from the start. Adding them after a working pilot is where teams find people can see data they shouldn’t.
• Find where your data actually lives before anything else. Halferty’s was spread across 23 systems, and pulling it into one place came before the AI work.