Read time: 3 minutes

Last week I ran a live AI workshop for the finance team of a digital healthcare company in San Francisco.

Accountants, FP&A analysts, GTM team members - a full cross-section of the function.
The brief was simple: help them use the tools they already have access to and cut their month-end close cycle in half.

Here's what happened.

The gap is almost never the tools.
Claude, ChatGPT, and Perplexity are all accessible.
The gap is that nobody has shown finance teams how to actually prompt them for finance work. Most people type a question, get a mediocre answer, and conclude that AI doesn't work for their job.

So we started with prompting fundamentals.
How the three tools work differently from each other and when to reach for which one.
The RACEF framework: Role, Action, Context, Examples, Format - as a repeatable structure that gets useful output rather than generic output.

Then we moved into their actual workflows.
Variance commentary. Month-end reporting narratives. Reconciliation prep.
Board-pack summaries drafted from raw data.
The tasks everyone on that team does every month, in the same structure, that currently take the most time.

The demos were where it shifted.
Watching Claude produce structured variance commentary - with the right tone, the right format, the right materiality framing - in under a minute.
Not a perfect first draft. A strong first draft that needs editing, not building from scratch.

That's the thing that changes how people think about it. The time is not saved by producing perfect output. It's saved by never starting from a blank page again.

But here's the thing I noticed that I wasn't expecting.

The people who got the most out of the session weren't the most technical people in the room.
They were the ones who could articulate their own workflow clearly.
The analyst who could describe exactly what she does every month-end - step by step, input by input - got dramatically better output than the person who just said "help me with reporting."

AI doesn't reward technical skill.
It rewards self-awareness about your own process.

If you can explain what you do to a smart intern in clear steps, you can get AI to handle most of it. The problem is that most finance professionals have never been asked to articulate their workflow that precisely.
The AI exercise is almost incidental - the more valuable thing is finally mapping out what you actually do every month.

That's what RACEF is really for.
Not just a prompting framework.
A way of forcing you to be specific about your own process before you ask AI to help with it.

By the end of the session, everyone had a workflow they could run the next morning. Not a pilot. Not a proof of concept.
Something that replaces a specific task they do every month.

That's the only bar that matters: does it change what they do on a Tuesday?

The team's energy when we showed them what Claude could actually do in real time!!

Every Monday at Hewlett Packard Enterprise (HPE), 40 to 50 business leaders join a 90-minute performance review.
The finance team used to prepare around 100 pages of PowerPoint for that call.
It took hundreds of hours across the business.
People scrambled over weekends to pull reports, reconcile data, and format slides.

HPE's CFO replaced it with an AI analyst named Alfred.
Built on Deloitte's platform. Custom NVIDIA infrastructure.
Deployed across 3,000 people.

The results: 90% reduction in manual effort.
Financial reporting cycle down 40%.
Processing costs down 25%.

She spent millions.
You just need 15 minutes and a Claude subscription.

Here's how to build your own version.

Step 1: Create a new Project in Claude

Go to claude.ai, open Projects, and create a new one.
Name it after your company or finance function.

In the Description field, paste a system prompt that tells Claude what your business is, how it should analyse, and what output style you want.
Think of this as the briefing document for your AI analyst.
The more specific it is, the more useful everything that follows becomes.

A good system prompt covers: what the company does, your FY dates and reporting currency, what materiality thresholds you use (flag variances over £15K or 5%, for example), and what to lead with.

That last point is the most important instruction you'll write: tell it to lead with the insight, not the number.
"Gross margin fell 520bps in March driven by emergency freight costs" is useful.
"Gross margin was 36.5%" is not.
One line changes the quality of everything that comes back.

Step 2: Upload your source files

The power of a Project is that Claude holds multiple documents in context simultaneously and cross-references between them.
Upload your GL export, AR aging report, bank summary, ops dashboard, and any relevant board minutes or management commentary.

When you ask a question, Claude is not just looking at one file. It's pulling from all of them and connecting the threads. Three different source files, one coherent finding. That's the moment it stops feeling like a chat tool and starts feeling like an analyst.

Step 3: Run three prompts

The first asks for a weekly performance brief.
What are the three to five things that require a decision or escalation?
Where are we versus budget and why does the variance matter?
What does cash look like over the next 30 to 60 days?

The second builds a scenario model for anything that needs stress-testing.
A supplier delay, a revenue shortfall, a cost overrun.
Describe the scenarios and ask Claude to model base, stress, and worst case with the assumptions clearly stated so your board can challenge them.

The third builds the board pack.
Describe who's in the room and what they need to see.
Claude builds a structured output: executive summary, P&L with variance highlights, scenario comparison, recommended actions with owners and timelines.

The work that takes your team a week takes 15 minutes.

The 100-slide Monday deck was never the real problem.
The problem was that it consumed the time needed to actually lead.
This is what getting that time back looks like.

If you want the system prompt template I use to set this up with clients, hit reply and I'll send it over.

I want to share how people in this community are putting AI to work in their real lives. Not the demos. Not the hype. The actual stuff.

This week, two that stood out.

Sarah in London is on maternity leave and used Claude to build a custom dashboard tracking her baby's feeds, naps, and development.
Every morning she gets an automated email summary of the previous day, with coaching tailored to her baby's current age and developmental stage.
She wanted to keep building her AI knowledge while on leave.
She ended up building something genuinely useful for her family at the same time.

James in New York is a Senior Manager in the public sector exploring a move into big tech. He used Claude, ChatGPT, and Gemini together as a career consultant - refining his CV, aligning it to target roles, running mock interviews.
For one role, salary data online was pointing him toward a range well below his actual market value. Claude pushed the number up based on his background and experience. He gave that range to the recruiter. They confirmed it was right. He got the offer.

Two completely different situations.
Same pattern: specific input, useful output, real outcome.

If you have a workflow worth sharing, hit reply. I'll feature the best ones in future editions.