How Cheetah Uses AI in Financial Reporting (and Where It Doesn't)

"Is this AI-generated? And if I run it again next month, will the numbers come out differently?"

Plenty of tools that use AI in financial reporting now put a model between your ledger and your report. Cheetah doesn't do that. While we might use AI to build the report "factory", the reports that are generated by Cheetah are deterministic, i.e. you get the exact same report if you run it with the same data multiple times. Every report that reaches you comes out of code we wrote, read, tested and can explain line by line. Your data is never used to train AI.

Here's where AI fits at Cheetah, and where it doesn't.

The factory and the report

Every Cheetah report has two parts. The factory is the code built around your business logic: which Xero accounts roll into which management line, how intercompany balances are eliminated, which exchange rate applies where. The report is what the factory produces each time it runs, on demand or on a schedule.

AI helps us build factories faster:

• Writing code. Pagination, sheet formatting and mapping tables are repetitive. A model drafts them quickly.
• Writing tests. Tie-outs to Xero's totals, leap years, 53-week years, a mid-year change of functional currency. Rigour that used to cost too much for a small shop is now cheap.
• Reading docs and research. A fast first pass through Xero's developer docs or an accounting standard, checked against the source before anything depends on it.

The hard part is still human: understanding how your business runs and handling the messy edge cases (more on that). If you'd rather build it yourself, here's what that involves.

Why the report is deterministic

When a factory runs, no model decides anything. Same data, same configuration, same period, same report.

Here's why that matters. Ask a model to build a consolidated P&L in March and it files Bank Charges under Finance Costs. In April it puts them under Administrative Expenses. The bottom line agrees, but your variance analysis now shows a $4,200 swing between two lines that never happened. We saw the same when we tested a general-purpose agent on a consolidation: fine for a one-off, but it works the report out afresh on every run.

The logic only changes when we change it with you, in a version we've reviewed.

We do use a model in one specific instance. When you set up unaudited financial statements, you can upload last year's signed statements and a model prefills the setup form: company name, directors, which line items and notes appear. You review every field before it's saved. It doesn't map a Xero account or calculate a number.

Why AI in financial reporting must not be a black box

We don't ship architecture or logic we don't understand and can't underwrite.

We deal with extremely sensitive data: group structures, payroll, margins, intercompany loans. We have a fiduciary duty to (1) ensure it's properly protected and handled, and (2) represent it properly in our reports.

AI makes the second easy to get wrong. Generated code can look right and still flip the sign on one account type or silently drop a class of transactions. So for every factory we ship:

• We can explain every function.
• The report reconciles to Xero, with the checks visible in the output.
• Judgment calls, like an FX convention or the treatment of an unmapped account, are decisions we made with you and wrote down.

If we can't explain it, it doesn't ship.

Lift the ceiling, don't lower the floor

AI should lift what we can achieve, not lower our standards. It's why a deliberately small team can take on a multi-entity consolidation with currency translation and eliminations. It isn't a licence to ship slop: plausible text that says nothing, numbers nobody can trace, code nobody understands.

That includes this blog. We use AI for research and to pressure-test drafts, never as wholesale output. Claims about Xero are checked against Xero's own documentation, and example figures are invented and labelled. Research informs our judgment. It doesn't replace it.

Your data is never used to train AI

We don't train models. Where a model handles customer material, like the prefill above, it's through Anthropic's commercial API, and Anthropic's stated policy is not to train on commercial inputs or outputs by default. We haven't opted in to anything that changes that, and our privacy policy says so.

The same thinking is why Cheetah no longer needs access to your Google Drive: ask for the least you need, and say plainly what happens to it.

Four questions to ask any AI reporting tool

• Same inputs, same report? If a model generates the output on each run, expect drift.
• Can someone explain every number? Which accounts, which rule, which rate.
• Does it check itself against the ledger? Tie-outs should be visible in the output.
• Is my data used for training? Ask which provider, and on what terms.

Our answers: yes, yes, yes, no. If you'd like a factory built around your business logic, let's talk.