AI governance

Using AI safely with confidential financial data

Finance teams don't need a ban on AI. They need three clear rules and tools that are set up properly.

Ferdi Schulte · · 3 min read

Walk through any finance department today and you'll find people using AI assistants. Some paste a variance analysis into a chatbot to get a first draft of the commentary. Others ask it to explain a formula or tidy up a forecast narrative. The question is no longer whether finance uses AI. It's whether it does so safely.

The instinctive reaction is to block it. In my experience that backfires: people simply switch to personal accounts on their phones, and you lose all visibility. A better approach starts with understanding where the real risk sits.

Where the risk actually is

The risk is rarely the model itself. It sits in three places:

  1. The account people use. A consumer account may use prompts to train future models and keeps no company record. An enterprise licence typically excludes training, keeps data in agreed regions and provides audit logs.
  2. What goes into the prompt. Unpublished results, payroll data, customer names and M&A plans each carry a different level of sensitivity.
  3. What comes out and where it goes next. An AI-drafted number that ends up unchecked in a board pack is a control failure, not an AI failure.

Three rules that work in practice

1. Use the company tools, never personal accounts. Make the approved enterprise AI tool the easiest option. If it is harder to reach than the free one, people will choose the free one.

2. Classify before you paste. A simple traffic-light model is enough. Green: public or already published information, usable anywhere. Amber: internal data such as budgets and forecasts, allowed only in approved enterprise tools. Red: personal data, unpublished results before release and deal information; explicit approval or it stays out.

3. A human owns every number. AI can draft, summarise and suggest. The person who signs off remains accountable, and figures are always traced back to the source system.

Make it easy to do the right thing

Rules alone don't change behaviour. What does help:

  • A short list of approved use cases with example prompts, such as drafting variance commentary from an exported, anonymised table.
  • A small group of champions in finance who try things first, share what works and answer colleagues' questions.
  • A clear route to ask "can I use this data here?", with an answer within a day.

The bottom line

Safe AI use in finance is mostly ordinary good governance: the right tools, clear data classes and human accountability. Get those three in place and AI becomes a productivity gain instead of a risk you hope nobody notices.

Read next

Let's talk.

About a role, an advisory question or a workshop for your team on data, analytics, AI and FP&A.