That’s a question most financial modellers and analysts will have asked themselves recently.
The latest round of AI agents are making bold claims, and since I make my living from building and auditing models, I regularly look at what’s new and whether the claims are justified.
One in particular has been making a lot of noise lately – posts about Shortcut.ai landed in my LinkedIn feed, so I took it for a test drive. Here’s what happened.
(I’ll also be running a live webinar on this on Wednesday 21st January 2026 at 12pm if you’d like to see the gory details.)
Shortcut.ai’s promise
Shortcut.ai claims to:
“Build complex financial models in minutes while understanding every calculation and insight.”
However (spoiler alert), I think it has a different concept from me of what complex means. Not least because it also claims:
“Typical time to build and interpret a financial model manually takes 2-4 hours.”
Whereas we tend to measure model builds in days or weeks.
Shortcut.ai presents as a webpage that looks very much like Excel.
I gave it a reasonably detailed prompt – here it is, typos included:
“Build me a 3 statement financial model for a small chain of 10 restaurants with a 4,4 5 timeline, driven by assumptions about Average Cover Price and using an inital margin for CoS on Drink and Food, with inflation assumptions driving the costs in the forecast. The model shoud include rent and turnover rent and split front of house and back of house staff costs. There is a head office with further staff costs. Some restaurants are yet to open and have capex associated with the openings.”
It asked me five clarification questions about currency, the opening schedule of restaurants and whether to forecast for each restaurant individually with consolidation (which I chose to do).
First impressions
My prompt had a couple of typos, but that didn’t seem to cause a problem – apart from perhaps the request for a 4–4–5 timeline. Whether adding the missing comma would make a difference I’m not sure, but most finance people would recognise 4–4–5 as a timeline for accounting periods of 4 weeks, 4 weeks and 5 weeks on a repeating pattern – very common in retail and restaurant businesses.
In around 10 minutes Shortcut gave me an annual model with four historical years and five forecast years – so it had interpreted my request slightly wrongly, but first impressions were that it looked quite good.
Now to scratch the surface and find out whether this model would stack up.
The first thing I noticed was that despite asking me which currency to use (I chose GBP £), the model was formatted as $ – apart from the header at the top of each sheet.
The model had five sheets:
- Assumptions
- Income statement
- Cash flow
- Balance sheet
- Restaurant detail
- Head office detail
I’ll use our SCILS framework (Separation, Consistency, Integrity, Linearity, Simplicity) to break down the model, then cover some other details I found using our analytical tools.
Separation
At first sight there was separation of inputs, calculations and reports – but on closer inspection this was not as thorough as it should be.
- A great deal of calculations were performed in the report sheets, rather than in supporting schedules. This makes formulae less transparent than breaking calculations down into small, visible steps.
- Hard-coded numbers (embedded assumptions) were used – for example to split a single head office overhead cost into reporting lines. This is a ‘top-down’ approach, but the split should have been an input, not an embedded assumption.
- Inflation was applied from the start of the timeline (i.e. four years ago), but this was embedded in a formula, so wasn’t obvious.
- Inflation was calculated using (1 + Rate)^N but instead of N referring to a counter row, it was a hard-coded number – different in each year. This is poor practice and means formulae vary between rows.
Consistency
Most of the formulae were copied across rows, although some were different in the first column. This is debated amongst modellers but many would consider it acceptable.
However, there were some inflation calculations containing different hard-coded numbers in each column – a big no-no.
Integrity
The balance sheet did balance – without any plug numbers – but other errors throughout the model make this a dangerously reassuring phenomenon.
Many people see a balance sheet check as a measure of integrity, and certainly it is a positive attribute – but in this instance it balances in spite of many other issues.
Linearity
There were no circular references – hoorah!
We like a model to:
- read like a book – left to right, top to bottom; or
- sometimes read like a report – still left to right but with some headlines at the top
Broadly speaking, the model complied with this.
Simplicity
The model was quite simple, but some of the formulae had cross-sheet references to large numbers of cells, which makes them more difficult to navigate and test.
The only functions used in the model were IF, SUM and MAX.
Maybe that’s a little too simple.
Structure
The calculations for restaurants were stacked up vertically, with a simple calculation for food and drink revenue and cost of sales. There was then another set of stacked calculations for rent, wages and overheads.
The income statement sheet referenced these as cross-sheet links to individual cells on the ‘Restaurant-Detail’ sheet – for example:
=’Restaurant-Detail’!C10+’Restaurant-Detail’!C22+’Restaurant-Detail’!C34+’Restaurant-Detail’!C46+’Restaurant-Detail’!C58+’Restaurant-Detail’!C70+’Restaurant-Detail’!C82+’Restaurant-Detail’!C94+’Restaurant-Detail’!C106+’Restaurant-Detail’!C118
This is definitely not how I would have built it, as it makes the model more labour-intensive for a human to work on.
I would either:
- have 10 sheets with identical structure and use 3D summing to consolidate, or
- if using a stacked structure, use SUMIFS or an indexing/lookup approach
At a minimum, I would have a summary on the same sheet so references are local, not cross-sheet.
The beating heart of most three-statement financial models is the corkscrew or control account. These were completely absent. There’s a reason they’re so widely used – they make it easy to see movements in balance sheet accounts. I’m surprised this nearly universal approach wasn’t used.
On a smaller point, the model used:
- Column A blank
- Titles in Column B
- Timeline starting in Column C
This leaves no spare columns for units or non-time-related drivers.
Formatting
Formatting was adequate but not outstanding:
- The final period of the assumptions sheet had no formatting
- No cell styles were used
- Some dark text appeared on dark backgrounds
- Negatives were correctly shown in brackets
- All monetary values used a $ format, despite selecting GBP
A units column would have avoided most of this.
Formula errors
This was perhaps the most important finding.
There were a lot of formula errors. A few formulas referred to blank cells (on the row above where they should have referred), meaning the values were incorrect. For example:
- Corporation tax used blank cells instead of the input tax rate
- Capex references were all offset by one row, meaning:
- Fit-out capex was spent every year for open restaurants instead of maintenance capex
- New restaurants had zero fit-out cost
This overstated capex – it should have been £7.5m, but the model showed £41.6m over 9 years.
It also meant the business remained loss-making instead of breaking even in 2027 – with cash burn around £5.5m per year instead of ~£1.6m, before becoming positive in 2027.
Further problems:
- Inputs for pre-opening costs weren’t used anywhere
- Inputs for equipment depreciation weren’t used
- Fit-out life of 10 years was incorrectly applied to maintenance capex
- Depreciation was combined incorrectly, with no mechanism to prevent over-depreciation
- Base rent calculations referenced only year one inputs – no absolute referencing
- Accrued expenses appeared as an asset not a liability
- Accruals were calculated as one-twelfth of staff costs, ignoring payable days and other expenses
This simply isn’t how accruals work.
The verdict
To be fair to Shortcut.ai, it probably has some appropriate use cases.
For example, it was quite good at populating assumptions with credible values with minimal prompting. This alone could be useful – such as when modelling in an unfamiliar sector – almost as a data-gathering assistant.
However, it does not cite its data sources.
It also has a nice Trace feature, similar to our Cell Tracker, which helps navigate formulae.
But this isn’t really what I was testing.
I wanted to know whether this is a replacement for an analyst – to build a simple model, like the sort a fractional CFO might want for a business trying to raise finance.
And the answer is an emphatic:
No.
At present, this tool is a sharp knife you could easily cut yourself with.
It produced a tidy-looking workbook – the sort of thing a banking analyst might produce – but as I scratched the surface, the veneer soon wore off.
The deeper I dug, the more problems I found – and after an hour or so the list had become very long.
I’ve been building and auditing models for a quarter of a century – I’m tuned in to finding errors and I have an armoury of tools to help me do so. That fractional CFO isn’t going to have those tools* or skills – and could easily overlook these issues.
Granted, the tool is in its infancy – but it suffers from overconfidence.
And that could lead to very expensive mistakes.
We’ll keep watching this space and may come back to Shortcut to see how it tackles other tasks. Meanwhile, humans – despite their flaws – are still a long way ahead of AI in building complex financial models.
And don’t forget, we’ll be diving deeper into this topic in our upcoming webinar: AI vs Analyst: Risks, Realities and What You Need to Know on Wednesday 21st January 2026 at 12pm
*Though they could do worse than to download our free nXt add-in
