Our AI is not allowed to do arithmetic. That is the feature

Language models are confidently wrong about numbers. Here is the architectural rule that keeps an invented figure from ever reaching your tax return.

S

The Salli team

Dec 2025 · 6 min read

A solid charcoal papercraft calculator beside a thin, hollow, empty paper speech bubble

Ask a language model what your tax bill is and it will tell you. Fluently, in the right format, with the right currency symbol — and sometimes wrong by six figures, with no signal that anything went astray.

That is a fun quirk when the question is trivia. It is unacceptable when the answer goes on a return you sign. So in Salli, the AI is structurally not allowed to do arithmetic on your money. Not discouraged by a prompt — prevented by where the code lives.

This isn't a hypothetical risk

In March 2026, a CNBC reporter asked ChatGPT for help with a real tax return (ESPP sales, a genuinely tricky area) and got a fast, well-formatted, confident answer. At one point the model told them something was "almost certainly" correct when it actually needed verification. The advice was delivered breezily enough that the reporter was nearly ready to file on it. J.T. Eagan, a clinical assistant professor of accounting at Purdue University quoted in the piece, warned that this exact pattern, plausible, confident, and wrong, is what catches people who aren't already tax experts, because they have no way to tell a hallucinated answer from a correct one just by reading it.

That's not a one-off. A widely-cited small-claims case, Moffatt v. Air Canada, saw a Canadian tribunal rule the airline liable after its website chatbot invented a bereavement-fare policy that didn't exist and told a grieving customer he could claim a refund under it. Air Canada argued the chatbot was "a separate legal entity" responsible for its own words. The tribunal rejected that outright: the company is responsible for what its AI tells customers, full stop. It's a strong precedent for exactly the kind of product Salli is: a financial tool where an AI's confident-sounding mistake has real, billable consequences for the person relying on it.

The temptation, and why we didn't take it

It would be easy to let an AI "just compute" your tax from a description of your income: type in your salary and deductions, get a number back. It would feel magical, right up until it hallucinated a rate band, missed a credit, or silently misapplied last year's threshold instead of this year's, and did so as confidently as it would state something true. Bloomberg Tax's analysis of AI hallucination risk in tax work flags exactly this: models can fabricate citations and produce wrong answers with no visible hedge, because sounding certain and being correct are two entirely different capabilities in a language model.

What the AI actually does instead

Salli's AI reads your documents, drafts entries for your approval, and explains results in plain language. Every number it shows you (your tax liability, your APIT credit, your final bill) was produced by a deterministic engine running a versioned rule pack that is golden-tested against the IRD's own worked examples, never generated by the model itself. The AI's job is to parse, draft, and explain. The engine's job is to compute. Those are deliberately two different systems, and the boundary between them never moves, however good the underlying model gets.

Why this is safer, not slower

A deterministic engine gives the same answer to the same inputs, every time, and records exactly which rule produced each figure. That's what makes a number reproducible years later if the IRD ever asks how it was calculated, and what makes it something you can actually stand behind, not just something that sounded plausible when you read it the first time.

The Air Canada case is the sharpest illustration of why this matters: a company was held legally responsible for a number its AI made up. Salli's architecture makes that specific failure mode structurally impossible for the numbers that count, because the AI is never the thing doing the counting.