A non philosophical AI use case in tax

A group has 150 entities feeding into the same corporate tax process. Sort them properly and most fall into one of three categories 1 - straightforward, sitting on the ERP with a trial balance that already ties to the financial statements 2 - external, coming from a practice rather than the group's own systems, usually as a single scanned PDF of a physical file, several hundred pages, typed ledgers, handwritten notes, stamps, no structure to speak of 3 - a third category that doesn't have a polite name, so call it: multi-input, multi-adjustment, multi-source lunacy, five or six disconnected things landing on one person's desk at once, signed PDF accounts, a trial balance in Word, a spreadsheet of journals to bridge local books to the audited number, a thread of audit query emails, meeting transcripts where the real adjustments got agreed. Read one of the scanned files by hand, at a pace where you're actually extracting figures rather than skimming, and you are looking at the better part of two working days before you've reconciled anything. Across 150 entities, that is weeks of a qualified person's time spent finding numbers, not judging them. I am on a panel this week where the conversation will spend most of its time on the harder questions such as how much judgement AI should be trusted with, where accountability sits when it gets something wrong, what the profession loses if it moves too fast. Those questions deserve the room they'll get but I want to put one narrower thing next to them, not instead of them: a specific piece of the job that does not touch judgement at all, and that AI already does well. I have not built this exact version for corporate tax but severals similar to it in other areas, so take the two days as a realistic estimate rather than a measured result. But the design is not degree level complicated, and there's no reason it would take longer than an afternoon to get a first version running with even a passing knowledge of co-pilot or chatGPT. Set the destination first: one standard trial balance format, a list of adjustments with a source tagged to each one, a flag on anything that couldn't be read with confidence. Every entity gets mapped into that same shape. The straightforward ones barely need anything: pull the ERP trial balance, pull the financial statements, confirm they agree, flag it if they don't. The scanned, practice-sourced ones need OCR that can cope with handwriting and stamps, not just typed text, a pass that recognises the single file is really dozens of documents bound together and splits them before extracting anything, then a trial balance rebuilt from whatever ledger pages actually contain account balances, since there's no clean source to start from. The multi-source ones need each input handled on its own terms: the PDF accounts and the Word trial balance extracted into tables, the journal spreadsheet taken across largely as is, the audit emails and meeting transcripts read for what was actually agreed rather than every option that got floated and dropped. Every number carries where it came from, a document, a page, an email and every entity gets reconciled on its own terms. And anything extracted with real uncertainty goes in front of a person before the pack is called final, with more scrutiny where the source material was messier to begin with. None of that is a tax decision really, it's the work that happens before a tax decision is possible, and across 150 entities it's currently the part most likely to eat weeks of good people's time, because a manual reconciliation across five mismatched sources, repeated entity after entity, is hard to retrace six months later when someone asks where a number came from. That's the whole idea, its not that AI should be trusted with judgement. That it can already take the weeks out of getting 150 entities to a point where judgement is possible, leaving the actual scarce hours for the actual scarce thing. I spend a lot of time talking about knowing your data and processes intimately before getting into transformation but this is a relatively simple idea of hoe to deal with the ambiguity of not owning the output of the underlying processes and how you might harness new tools to deal with it.