Why faster is not always better

Published 26 September 2026

AI reads an annual report in seconds and sounds completely sure of itself. That is exactly what makes it so dangerous in finance. What can go wrong, where it already has, and why the check belongs before anything is written.

← Articles

Why AI seems so convincing

Ask an AI for the key figures of an annual report and a clean table appears within seconds. No typing, no searching, no tired eyes in the evening. Next to copying by hand, it looks not only faster but tidier and more thorough.

It really is faster. The tidy result is deceptive, though. A language model neither looks figures up nor copies them. It writes whatever answer is most likely, given everything it has read. Most of the time that is the figure in the report. Sometimes it is not, and the answer looks every bit as finished.

Confidence scores do not change that, and neither does any other signal that claims to rate how reliable a figure is. The model always gives the answer it considers most likely, so it is most certain of exactly what it says, right or wrong. A confidence score can never tell you that a figure is right.

What goes wrong

The mistakes are rarely obvious. Mostly they are plausible, which is exactly why they slip through:

The figure comes from the prior-year column.
Thousands turn into millions, or the other way round.
A line with a similar label is picked, such as one segment’s revenue instead of the group’s.
The figure comes from the notes instead of the balance sheet, income statement or cash flow statement, or from the wrong one of the three.
The figure appears nowhere in the report.

In a model, none of these stays small. One misread unit and the valuation is off by a factor of a thousand. One prior-year figure and growth turns into decline. And it is still your name under the result.

Even the biggest firms get caught

This is not a problem of small teams under time pressure. Since 2025, several of the largest advisory firms in the world have had to revise or withdraw reports because an AI made things up:

Deloitte Australia: a report for the Australian government cited research papers that do not exist and a court quote that was made up (Fortune).
KPMG Australia: in a report on agentic AI, only 5 of 45 citations led to real sources (City AM).
EY Canada: a report on loyalty programme fraud was withdrawn because more than 70 percent of its citations were fabricated, broken or wrong (National Technology).
PwC Middle East: several reports cited sources that were invented (International Accounting Bulletin).

These were citations, not balance sheets. But the mechanism is the same one that puts a wrong figure into a model: the result looked finished, and nobody checked it against the source before it went out. These firms have review processes and reputations to lose. It got through anyway.

Why checking afterwards does not work

The obvious answer is: then someone checks the result. In practice that rarely happens. An AI answer does not tell you where a figure came from, so checking it means finding every figure in the report yourself. That is the entire job you were trying to save, and when time runs short, it is the first thing to go.

Checking only works if it is much faster than doing the work yourself. And that only happens when the source sits right next to the value.

Why a cited source proves nothing

Some vendors counter that their AI is reliable because it cites a source for every figure. That sounds like proof, but it is not. A citation only shows where the AI claims to have read the figure. It does not show whether that is the right place in the report, or whether the source really says what the AI claims.

An automatic check that the figure really appears at the cited spot changes nothing either. If the model took the prior-year column or a single segment’s revenue, the figure is right there, and the check reports: correct. It confirms the wrong answer.

modelflow shows the source too, but not as a guarantee. It is there so you can see at a glance whether the AI took the right figure, and decide for yourself.

Hoping or knowing

No AI model reads a report without errors. Not ours, not others, and not future models either. That is not down to us but to the technical nature of current AI. So the question is who finds the wrong number, and when.

A tool that writes straight into your model makes wrong values your problem after the fact. You can hope the numbers are right, or you can know.

modelflow gives you the speed of AI-assisted extraction, under a minute for several hundred pages, and puts the checking where it costs least: visibly, per field and per page, before anything is written.

Every value appears on the report page it came from, with the figure highlighted.
If a value is right, one key is enough. If it is wrong, you click the right figure.
Nothing is accepted automatically, and no confidence score decides for you.
Only what you approved goes into the workbook, and the journal keeps the source page of every approved value.

A right value is a keystroke, a wrong number a click, a few seconds instead of hours of concentrated typing. And every figure in your model has been seen by a person. How modelflow compares with the other approaches is covered in How to fill Excel with figures from financial reports.