Reef TRHEvery "AI for accountants" tool demos the same way. Clean sample invoices, a tidy chart of accounts, a handful of transactions that sort themselves while the room nods along. It looks like the future. Then the firm loads a real client's month, and the future quietly comes apart.
The instinct in that room is simple: buy the tool with the best demo. It is also how firms end up paying for AI that cannot survive their own books.
The demo runs on the 10% that was always easy
A demo works because it runs on data that does not exist in real life. Real books are a supplier who writes "inv" on the back of a delivery note, a client who sends the same receipt twice, an expense that only makes sense because of a decision someone made three years ago. The demo never shows those, because those are the actual job. The clean part was always going to work; it is the nine-tenths underneath that decides whether the thing survives.
A generic tool has to guess, and it guesses silently
A shrink-wrapped tool has to assume an average firm, so it books things the average way. But no firm is average. Each has its own clients, its own habits, its own reasons for putting a cost where it puts it. When a generic tool meets that, it guesses, and it does it without telling you. It files a one-off equipment purchase as a routine subscription because that is what the pattern usually looks like, and the mistake rides quietly into the report. A wrong number that looks confident is worse than no number at all, which is the whole problem with an AI you cannot check.
What actually holds on real books
The systems that survive a real month are not the ones with the best demo. They are the ones built around the mess. They read documents in whatever state they arrive, they learn a specific firm's way of booking from that firm's own history, and they hold anything they are unsure about for a person to decide instead of guessing. That is the difference between software that performs on a stage and software that runs a client's actual month, which is exactly what our monthly-reporting build was designed to do.
So when you watch the next demo, do not ask whether it works. It will; that is what demos are for. Ask what happens when you hand it the messiest client you have, unsorted, mislabelled, half-photographed. The answer to that question is the only one that matters, and it is the one the demo is built not to show you.
Common questions
Why do AI accounting tools work in the demo but fail on our books?
The demo runs on clean, curated sample data. Real books are full of exceptions, odd formats, duplicates, firm-specific categories, that the tool never had to handle. The demo tests the easy 10%; your month is the other 90%.
How do I tell a real tool from a good demo?
Ask to run it on your own messiest client's month, not the vendor's sample. Watch one thing above all: when it is unsure, does it flag the item for a person, or does it guess and move on?
Is a custom build always better than an off-the-shelf tool?
Not always. What matters is whether the tool can fit your firm's own way of booking, which it can only do by learning from your history. Judge it on adaptation, not on the label.