Even if you could 3D-print a car, you still wouldn’t be allowed to drive it on the road.
Not because the car can’t move. Because no one drove the car into a concrete wall.
The same goes for any financial software developed internally with the help of AI that feeds into statutory reporting. You will need to explain how AI thinks.
Of course, you shouldn’t take my word. You should ask your auditors, and when you see the happiness in their eyes, know it is because you showed them the invoicing path to become an Equity Partner.
This is not a warning against AI, but a warning against premature victory laps.
Producing an annual report is one achievement. Producing an annual report that auditors, regulators, lenders, investors, and the board can rely on is a different sport.
An AI system that can generate an annual report is impressive in the same way a car that can accelerate is impressive. Useful, yes. Sufficient, no.
The auditors are not primarily concerned with whether the vehicle moves. They are concerned about whether the steering wheel detaches at 200 km/hour while the Board is in the back seat discussing dividend policy.
Assume the CFO office succeeds by using AI to code a fully functional statutory reporting platform.
The system connects to the ERP, extracts trial balances, maps local charts of accounts into the group chart, performs eliminations, calculates currency translation, handles acquisitions, disposals, non-controlling interests, Zzzzzzz, deferred tax, cash flow statements, disclosures, local statutory formats, and XBRL tagging.
The numbers are right. The filing package is complete.
The CFO office has not 3D-printed a toy car. They have printed a real car. And it looks and feels like a Ferrari.
It starts. It accelerates. It brakes. It has cupholders and connects to Spotify, prompting the steering committee to call it “enterprise-ready.”
Then a Toyota Corolla pulls over, the most likely car to be owned by auditors because of its reliability and practicality. And lo and behold, the auditors have indeed arrived.
Then a Toyota Corolla pulls over, the most likely car to be owned by auditors because of its reliability and practicality. And lo and behold, the auditors have indeed arrived.
The question is not whether the car moves. The question is whether it is road legal.
The auditors will not say, “Wonderful, the AI produced the annual report, please show me where to sign.” They will ask where the numbers came from. “From the ERP” will not be the answer they are looking for. Auditors are basically forensic mechanics in conservative shoes. They will start removing panels.
You have now become a software vendor without sales
The first-year audit will almost certainly be more expensive, not less. The CFO office builds the machine to reduce costs, and the auditor responds by opening a second workstream called “understanding the machine,” which is audit-language for “we are going to invoice you for the privilege of explaining your own innovation back to you.”
The auditors will not merely sample the financial statements. That would be like crash-testing a car by licking the paint. They will feed the system duplicated trial balances, missing entities, changed ownership percentages, late top-side adjustments, unresolved intercompany differences, inconsistent FX rates, and ambiguous mappings.
They will want to see whether it fails loudly, blocks processing, generates exceptions, preserves evidence, and routes review to the right person.
Will that cost repeat every year?
Not fully, but it will not disappear either. Each year, auditors still test whether controls operated effectively. They examine changes to mappings, code, prompts, models, interfaces, access rights, accounting policies, ownership structures, FX tables, consolidation rules, and disclosure logic.
If the reporting machine is materially the same machine as last year, some audit effort can be reused.
If, however, the AI has been helpfully rewriting the engine every quarter, the auditor will not treat last year’s test as evidence. Last year, they tested a Ferrari. This year, you have arrived with something that looks like a motor home.
The expensive part of statutory reporting software is not arithmetic
Arithmetic is a cheap magic trick. The expensive part is institutional trust: interfaces, controls, permissions, audit trails, workflow, change management, data lineage, exception handling, reconciliation logic, security, evidence retention, regulatory filing support, business continuity, and scale. The CFO may discover that AI can build the visible cathedral quickly, but the foundations, fire exits, insurance policy, and load-bearing walls still need engineering.
AI may demolish the cost of building the first 70% of the application. Unfortunately, the remaining 30% contains 97% of the liability and all the audit questions.
Then comes governance. Who can change mappings? Who can change prompts? Who can approve the generated code? Who can deploy to production? Who can override eliminations? Who can reopen a closed period? Who can modify FX rates, ownership structures, materiality thresholds, or disclosure wording? Who can suppress exceptions?
A bad AI system treats all of this like airport luggage: same belt, same barcode, same bored confidence. A serious one knows that some bags go straight through, some need customs inspection, some contain liquids over 100 ml, and one suspiciously heavy suitcase marked “deferred tax asset” should probably be opened in a separate room.
What happens if the auditor doesn’t sign?
A statutory close is not merely a calculation. It is a procedural history. If the AI system cannot reconstruct that history, no auditor will sign. Responsibility remains with management. The CFO owns the reporting process. The board oversees it. The audit committee supervises it. The auditors test it. AI can accelerate work, but it cannot absorb liability.
So yes, the CFO office may eventually 3D-print the car. It may even be lighter, sharper, cheaper, and better tailored than the vehicles sold by enterprise software vendors.
But before anyone drives it onto the statutory highway, it needs crash tests.
If it passes, something important changes. The CFO office stops being a document factory and becomes a controlled computational representation of the enterprise. The annual report becomes the output of an interrogable financial model. Audit becomes less about chasing spreadsheets and more about testing a governed reporting machine.
To ensure that financial software has been crash tested, financial software vendors go through a comprehensive assurance process called SOC 1/ISAE 3402 related to financial reporting and/or SOC 2/ISAE 3000 related to GDPR. The expense runs in the tens to hundreds of thousands of US dollars. Every year. It is this assurance process or equivalent that internally AI developed systems potentially must go through and the cost will be added to the existing audit expenses.
AI in statutory reporting can become either an audit accelerator or an unprecedented increase in audit costs. The difference is not whether the output looks right. The difference is whether the system leaves evidence good enough that the auditor does not have to excavate the truth by hand.
Because financial statements are not prototypes. They are like vehicles other people rely on.
And if you are going to drive one on public roads, someone is going to inspect the brakes.
