generally unacceptable accounting practices

more notes from the mental prison

American readers! “The Unaccountability Machine” will soon be available in your continent again/properly! It’s available from University of Chicago Press or Amazon (the UCP page has the correct cover - the Amazon one has picked up the graphics for the UK paperback version which is also imminent).

---

The very first post on this ‘stack was called “Every Accounting System Is A Mental Prison (from which we must periodically try to escape”), and in the FT at the weekend I had a little article trying to make a few suggestions to the escape committee. Long term readers will perhaps have recognised the underlying point I was trying to make – that if we are going to take AI seriously, we have to think about actually changing the way we do things to take advantage of a new technology for organising information, not just sending more emails at a slightly ridiculous cost.

I ran out of space (and any reasonable estimate of reader patience) before being able to get too far into specifics, so I thought I’d expand here – partly to get it off my mind and into print, but also because thinking about how you could use AI to change the way company accounts are compiled is a good starting point for thinking about “what are these things we call accounts?”, which in my view is a very important thing for people to have in their consciousness.

Before doing so, a bit of an apology – David Gerard pulled me up for doing something I’d pontificated against before, and saying “LLMs” when I wasn’t actually wanting to refer to that specific architecture. I do probably think that something like an LLM would be one component of my ideal accountancy AI (because part of what I want to do is extract implicit rules and structures from a large corpus of users of accounts). But quite obviously, you can’t really be tolerating hallucinations in this context; the numbers have to add up. And although human auditors also sometimes have brainworms and make things up, that’s not much of an excuse.

Taking the raw material (ledgers and daybooks – records of individual transactions) and aggregating it into an income statement and balance sheet is basically a classification task, though, so I think it’s exactly the sort of thing that some sort of neural net ought to be good at. And by tweaking the way that the network classifies things, you could do some fun things with the journey from the ledger to the financial report:

Comparability: The way things work at the moment, every company’s accountants carry out this fundamental task – aggregating the transactions in the ledger into an income statement and balance sheet – in the way that they think suits best. In the famous language of the Companies Act, they want to provide a “true and fair view”. There are loads of choices and compromises to make along the way, and each set of accountants are likely to make different, but equally justifiable decisions. The job of being a financial analyst is, in large part, that of understanding what decisions have been made, and doing your best to adjust the accounts of a group of companies so as to make them more comparable.

What would be nicer, might be if every time you wanted to compare the performance of half a dozen companies, you could use a system that had direct access to the ledgers of all six, and could draw up six sets of financial statements, making choices specifically with the goal of maximising comparability. There would have to be quite a bit of institutional design around this – obviously, the raw ledgers can’t be made directly accessible to the general public as there is a lot of extremely confidential information there. But a Generally Accepted AI could be used by everybody, leaving the auditors freed up to do the thing that they are hardly ever able to do these days – check that the actual ledgers are recording reality.

Timing. As I say in the FT piece, “The most misleading numbers in any annual report are often the dates at the top of each column — they imply, often comically wrongly, that 12 months is the relevant period over which performance should be assessed”. This is a real bee in my bonnet, and it’s a genuinely difficult problem to solve. You want the accounts to cover a meaningful period of time with respect to the actual production process, but you also want them to be at least a bit relevant at the moment when you’re doing the analysis. Another classification task of the sort that neural nets are very good at is finding underlying structure and cyclical patterns, and making suggestions about different time perspectives.

Even more wackily, you could combine this sort of classification request with a comparability project. As well as having different timescales, similar companies can also be out of phase – at any given moment, they are at different points in their capital and product development cycles. Genuine comparability might involve shifting the window back and forth in time to assess and benchmark the results when cleaned up for differences in the timing of major decisions.

Purpose. And of course, the other big compromise in a set of accounts is that they are used for all sorts of different purposes. Creditors want to concentrate on some things and shareholders on others. Regulators have so little overlap with either group that they actually do commission and require entirely different reporting, which is the occasion for huge amounts of trade association whining in the financial sector. If we got away from the idea that “the accounts” had to be one single document used by everybody, we could spend some time thinking about what we actually wanted to know about the business and how to present it.

Even the sacred convention of double-entry book-keeping itself might not necessarily be appropriate to every purpose. The double entry principle isn’t a fundamental law of capitalism – it’s a tool of convenience, to help accountants avoid arithmetic mistakes. It actually causes a lot of problems itself, because it’s very bad at dealing with things like brands and software assets, where the value isn’t always created in specific identifiable transactions. To my mind, it seems a bit odd to be worrying a lot about the possibility of AI hallucinations, in a context where we’re prepared to tolerate absolute fiascos like the current state of accounting for intangible assets.

But … more or less everyone I’ve tried out this idea on, absolutely hates it (the exceptions being AI guys with no accountancy background). The problem is that it would all involve actually handing over responsibility and actually delegating important decisions to the AI. My view is that if we’re not going to do that, what’s the point of it? (And “none” is certainly a valid response here – a lot of this is a thought experiment in trying to make all the “transforming our economy” rhetoric concrete and seeing if it still seems anything like as attractive).

Most people, though, actually like that single “set of accounts”, because it’s very central to a mental model that’s developed over centuries, and out of which you have to say the industrial world hasn’t done too badly. The illusion of a single, shared objective reality that is nonetheless sufficiently filtered and attenuated for a normal business school graduate to understand it in a few hours if they apply themselves – that’s quite an important thing to believe in. Substituting it for a sort of postmodern, Platonic model in which the reality of business is always hidden away, but different interpretations get presented to you based on a machine’s interpretation of what you want to see, which in turn is based on an algorithmic consensus of other users … that’s quite a big red pill to swallow.

------

[I should say, although this sounds like the most appalling grovelling, that both the piece as published in the FT and the thoughts above were greatly influenced and improved by Tony Tassell. He hated the idea as much as everyone else does, but printed it anyway which is great, and challenged it in interesting ways that I had to think a lot to deal with].