Rumbo · for finance teams
Make your finance team AI-native.
Train them on real finance work. Measure their AI capability. See whether they actually improve — and which of them can already spot an AI that is confidently wrong.
No sign-up. Review an AI-written board commentary and find what is wrong with it. Already have an account?
Your team has AI. Do you know who can actually use it?
You have given part of your finance function access to ChatGPT, Copilot or Claude. Most organisations stop there, and then cannot answer three questions: who uses it well, who trusts an output they should have challenged, and whether any of it is getting better.
The failure mode is not that people refuse to use AI. It is what happens when they do. An assistant produces a variance commentary that reads beautifully and attributes a cost rise to a cause that appears nowhere in the data. A cash note states operating cash of €62,800 when the statement shows an outflow of €29,400. Both are fluent, confident and wrong. Both get sent.
Access
is not capability
Completion
is not competence
Usage
is not judgement
A licence tells you who has the tool. A training completion rate tells you who sat through something. Neither tells you who would have caught the €62,800.
Four minutes that make the point better than we can
We built a short version of the thing we measure. You read an AI-generated board commentary and the numbers it was written from, and you say what is wrong with it. There are four errors. Two of them would change what a board decides.
No sign-up, no email wall. We show you the answers either way.
How a pilot works
- 1
Baseline, before anything else
Three tasks, about fifteen minutes, on a company nobody meets again. No hints, no coaching. This is the number every later figure is compared against, and taking it moves nothing — no points, no streak, no score.
- 2
Short work sessions, not courses
Thirty-five missions built from real finance situations: invoices with the Incoterm changed, reconciliations that do not close, AI drafts with errors planted in them. Five to fifteen minutes each, in a workspace with the files and an assistant.
- 3
A final assessment you open when you are ready
Same company as the baseline, a different quarter and different figures — so a change in score is the person changing, not the scenario. It stays shut until you open it.
- 4
A report you can take to your own leadership
One sentence at the top saying whether the team is getting better, then the funnel, the capability change, and time figures with their provenance attached.
What we will and will not claim
- Tasks completed, and how long each took.
- Assessment scores, before and after, under the same conditions.
- Which of seven skills moved, and by how much.
- Whether people catch AI that is confidently wrong.
- Hours saved on your team’s actual work. We cannot see it.
- A euro figure from an hourly rate we invented. You supply the rate or there are no euros.
- SOC 2, ISO 27001 or any certification we do not hold.
- Results from customers we do not yet have.
The report separates what was measured from what is estimated from what learners told us, on the page and in the export. A projection is labelled as a projection everywhere it appears.
What managers can and cannot see
- Who was invited, and who started.
- Team averages, once at least three people have evidence.
- Aggregate scores and skill movements.
- Workplace reports a learner chooses to submit — a task category and two durations, so a manager can confirm the estimate is reasonable. Never anything they wrote.
- Everything they wrote in a mission.
- Every message to the AI assistant.
- Their individual score. Managers see the team, not the person.
Averages are withheld below three learners so an individual cannot be recovered by subtraction. This is not a setting; it is how the reporting is built. Read the privacy policy.
Questions a CFO asks first
How is this different from an AI course?
A course teaches and tests recall. This gives people the work and measures what they produce — including whether they notice the error planted in the AI draft they were handed. The baseline and the final run on companies that appear nowhere in the training, so what is measured is whether the skill transfers, not whether the scenario was memorised.
How long before I see something?
The baseline is fifteen minutes and gives you a starting number for each person on day one. A change worth reporting needs roughly five completed missions per person; below that, the report marks the capability figure as preliminary.
Will the AI do the work for them?
The assistant computes, extracts and reconciles — that is the job. It will not decide what matters, write the deliverable, or tell anyone what is wrong with an AI draft they have been asked to check. That refusal is the assessment.
What happens to our data?
It stays in one database, under row-level security, with each learner able to read only their own rows. Learners can delete their account and their work from inside the product. We do not train models on your data. The full position, including what we are not certified for, is on the trust page.
What is the honest weakness?
Improvement is measured on our tasks, not on your month-end. We show a time difference against a reference we set ourselves, and we label it as an estimate every time it appears. If you need proof of hours saved on your own close, this pilot will not give you that — it will give you a measured capability change and a defensible basis for expecting one. The methodology sets out exactly how far each number can be pushed.
We are looking for our first design partners
Not customers — design partners. Twenty people, six weeks, hands-on onboarding and direct access to the founder. You influence what gets built, and you keep the capability report whether you continue or not.
You will not find customer logos on this page because there are none yet. What there is instead: a four-minute benchmark, a sample report showing exactly what six weeks produces, the methodology in full, and a trust page that opens with what we are not certified for.
