This is part 3 in my series on AI ROI. Part 1 was about turning hours saved into real results. Part 2 was about getting the same results with less AI usage. Today it is about which ideas deserve your budget in the first place.
She wanted to hire a junior accountant. Then she built GPTs instead.
"We start too big, because we feel AI is this superpower tool."
Marianne said this in one of our AI Finance Accelerator info sessions. She runs her own consulting company, helps companies through growth and change, and coaches in the my AI Finance Accelerator. She also worked at EY as an auditor.
She told us about one of our Accelerator participants, a fractional CFO.
They started with a very complex data set, and her custom GPT didn't work.
So Marianne and the group told her to try something simple.
Every month she had to clean one Excel file for that client. In her words: "It's just one Excel file, it's not complicated, but I really don't want to do it."
So she built a custom GPT for that job.
She was so excited that she sent Marianne her notes. She had tested it 5 times, and it worked!
She had wanted to hire a part-time junior accountant for exactly that job. Now the custom GPT does it, every month.
So she built 3 more.
Small idea. Real ROI.
Some of your AI ideas take 5 minutes. Others take 6 months.
|
|
Think about your team's list of AI ideas.
The list is probably big, but the ones that get built are probably small, no? An idea on a list has an ROI of zero. Plus, nobody ever got a bonus for a really long list 😉
Ideas are not the problem. The way you sort them is.
I've seen this with members of my AI Finance Club for a long time. Some ideas you can do in 5 minutes. Some need IT, clean data, three departments and six months.
Try to give all of these projects equal budget and they become Voué à l'échec (Doomed to fail).
Small ideas wait behind a committee. Big ones start without a plan and stall.
I go to the gym, and after my workout I like to ask ChatGPT what I should do next (I even call the challenges in my AI Finance Club – AI Workouts).
And a gym is a really good way to think about your AI ideas.
Nobody walks in and uses every machine, or does every exercise at the same time. You learn the correct movement on one first. Then the next. Each one works on its own to build a specific muscle group. Together, they make a circuit for your whole body.
Your AI ideas are the movements. And 3 tools help you decide which one to start with to improve your ROI, and create a bigger impact for your finance team – your biggest muscle groups 😉
You sort the ideas. You plan the big ones from the end. You build everything in small blocks.
Sort. Flip. Build.
Let’s go.
Sort, Flip, Build: my 3 steps to decide which AI ideas get your budget
Step 1: Sort.
|
|
Give each AI idea one of 4 zones, and a different decision for each
Ask each team lead for their AI ideas. Group them by process: close, FP&A, treasury, tax, controls etc
Now ask two questions about every idea.
What: do we agree on the result we want?
How: do we know the data, the tool and the steps?
The less clear the answers, the more method you need. Here is what each zone means for your decision.
- Simple. Both are clear. Drafting dunning letters, variance commentary from a clean P&L table, summarizing a contract. Do it this week. No project, no committee.
- Complicated. One is clear, the other not yet. You know you want every invoice booked to the right GL account, but you don't know how to connect AI to your ERP. Plan it, and get help from IT, a vendor or a colleague who has done it.
- Complex. Both are a bit unclear. An AI-supported month-end close. Everybody wants "faster". But faster how, with which data, and who reviews what? This one needs Step 2.
- Chaos. Nobody agrees, nobody knows. "Let's put AI everywhere in finance." Don't start a project here. Make the idea smaller, or shelve it.
One more filter.
Complexity is not important. Does this idea achieve a goal you really care about?
If you can't say which one, it gets no budget, however clever it is.
Tip – Running 10 km feels super hard. Putting your shoes on and going outside is not.
When you put an idea in Simple, those are your shoes. Stop planning and just do it. Keep the matrix for the rest.
15 minutes with it can save you months on the wrong project.
Step 2: Flip.
|
|
Budget for the goal. Don’t budget for the project.
This step is only for the complex ideas. The close. The forecast. The ones that need three departments.
Most teams plan from today. First this task, then that one, and the goal shows up somewhere at the end. Bent Flyvbjerg studied more than 16,000 big projects and most of them finished late and over budget.
His point: people start with the tasks, not with the goal.
So flip it. Ask the owner of each complex idea for an end goal.
Same than in the gym. You don't start with a list of exercises. You start with the goal, then you plan your training back from it.
It has two parts.
- What is true the day it works.
- And what you stop paying for, or where the freed-up time goes.
In part 1, I called those Sell and Fly. Fly means the hours go to better work. Sell means the hours replace something you pay for.
Say your end goal for the close looks like this.
- We close in 4 days instead of 8.
- You get commentary you trust on day 5.
- We stop booking the temporary help we add every quarter-end.
Now work back from it. What must be true just before that? And just before that? Until you reach today.
- You read the commentary on day 5.
- Before that, the controller reviews commentary drafted by AI.
- Before that, the variances are calculated automatically.
- Before that, the SAP export is clean and in the same format each month.
- So today: export one month and check the data like an auditor.
Your first step is super concrete. And when somebody comes with "can we also add a chatbot for the sales team?", you have an easy answer. Does it get you to your end goal? No? Then later.
Step 3: Build.
|
|
Build one machine, then the circuit.
In a gym circuit, you move from one machine to the next. Each machine trains one thing. And each time you add one, the circuit gets better.
AI in finance works the same way. Don't build the entire thing. Build blocks. Your variance commentary prompt, with the context, the format and your house style. The step that checks and cleans the export. The variance calculation. The email or PowerPoint page that goes out.
A good block follows three rules.
- It works alone. Like one machine, you can test it on its own.
- It is replicable. Anyone in the gym can use a machine without asking you how. Same for a colleague with your block.
- It is independent. If one machine is broken, you skip it and the circuit still runs. If one block breaks, you fix that one, not the whole process.
Here is the part I like most.
Your simple ideas from Step 1 are your first blocks. The dunning letter prompt. The commentary prompt. The data check. So when you start the complex project from Step 2, you don't start from zero. You assemble the blocks you already have.
That is what happened to our fractional CFO. She started with the heaviest weight, and it didn't work. Whereas her monthly Excel clean-up was just one machine. It worked alone. Then she built 3 more in a week.
I tell members to be "ambitious, but not too ambitious". Ask AI to automate everything and it gets so complicated that you never finish. So chunk the work.
On top, every block makes the next project cheaper.
Less time spent discussing, less time spent unpicking. More time saved, more tasks covered that don't require additional resources.
The best route to ROI with AI is one proven step at a time.
The One Thing to Remember
Ideas have no ROI.
Blocks that create incremental results do.
- Sort every idea into a zone.
- Only fund complex projects with clear end goals.
- Build one block at a time, then connect them into a circuit.
Right now, decide which block you will start with to build your AI finance circuit.
And download the PDF cheat sheet for this newsletter here to help you – AI in Finance Idea to ROI Cheat Sheet.pdf
|
|
You’ll be one step further towards creating a powerhouse finance team 😉
Best,
Your AI Finance Expert,
– Nicolas
P.S. The best way to learn this from me directly is my weekly masterclass. 60 minutes plus Q&A, where I answer your AI in finance questions. Join me here.
P.P.S Marianne's fractional CFO started with one Excel file. Here's how to build your first blocks the right way → How to Build Your AI CFO Clone in Claude (2026)