The Tick Mark Becoming a Test
Accountants have always had a model and forgot it. Every accountant already believes fiercely in invariants; but they just don't use that specific word. But they understand the idea. The whole pedagogy is this: you already know what should be true. Now say it out loud, in a form the machine can check. That is the tick mark becoming a test.
Because it is a test, techniques like Lean, Six Sigma, Total Quality Management, and Agile can be applied. And because of that, reliable, repeatable, scalable industrial processes can be created.
The simplicity trap. A spreadsheet feels simpler than a model, and that is true. But that simplicity comes from the limited expressive power of the spreadsheet. A model can represent far more than a spreadsheet ever could. It carries information, structure, and metadata that a spreadsheet discards, and it makes explicit the assumptions a spreadsheet leaves unstated. The additional complexity is not accidental: it exists because the model is holding more.
Simplicity favors the accountant who wants to finish the task quickly, without considering how the work lands on the next person in the chain. Richness favors the accountant who wants accuracy, transparency, and accountability across the whole chain. A "shortcut" is a matter of perspective. Viewed holistically, the model; slightly more complex at the start; makes the full workflow dramatically easier for the system as a whole. And there is a new partner in that chain now: artificial intelligence. Be conscious of your shortcuts. Considered all things, they may not be shortcuts at all.
Resistance is not irrational. Spreadsheets are democratic, flexible, and forgiving; models demand rigor up front and punish sloppiness. The incentive structure favors documents as well: you bill for producing them, and audit standards are written around them. The technical shift is the easy part. Rewriting what "audit evidence" means when the assertions are executable is the hard part; and that is mostly a regulatory and professional-identity problem, not a technology problem.
The real risk is not that accountants fail to learn modeling. The real risk is that a model built with spreadsheet-grade rigor is worse than a spreadsheet: wrong at scale, with AI operating on it confidently and systematically wrong.
The discipline that spreadsheets accidentally imposed; you can see every cell, however ugly; must be reimposed deliberately: version control, tests, review of logic rather than review of outputs, and explicit assumptions. That is a genuine skill gap: dimensional modeling, constraint specification, controls-as-code, data lineage.
How to make the switch. Start with a rename, not a curriculum. The worst way to teach modeling is to teach data modeling, dimensional schemas, or a query language. The right way is to point out that everything accountants, auditors, and analysts already do has a model underneath it; they have simply been expressing it on the medium of paper and then electronic paper:
- A business event is a situation that occurs within an enterprise.
- A journal entry is a typed and classified transaction record with a validity rule projected from those business events.
- A trial balance is a query over those typed, classified journal entry records.
- A reconciliation is a join plus a residual assertion ("the difference must be zero, or named").
- Consolidation eliminations are transformation rules, not a worksheet marked up in red ink.
- Financial statements are views over the underlying state and changes in that state; projections, not artifacts.
- A tie-out is a proof obligation.
Accountants, auditors, and analysts just need the right software. And yes, that is a challenge. In fact, this is one of the biggest unsolved opportunities in business software: the proper balance of flexibility and control.
The AI bridge is what makes this tractable now. Twenty years ago this switch would have required accountants to become programmers, and it would have failed: too much syntax, too little payoff. Today the accountant does the two things they are actually qualified to do: (a) state what must be true, in their own words, and (b) review whether the machine's formulation captured it. The AI writes the logic, the query, the test. The accountant reviews assertions, not code. That collapses the skill gap from "learn to program" to "learn to be precise about rules"; a skill accountants already have and use every day on their clients' policies.
Concretely, my approach with someone would be something like this: take one workpaper they know cold; a bank reconciliation, a flux analysis, an impairment test; and have them narrate the rules while I formalize them. Then we break it on purpose. A spreadsheet tie-out still foots because a hardcoded cell quietly absorbed an error; the model version throws an exception. That demonstration converts people faster than any argument. Fear of the model ("it exposes my mistakes") flips to trust in it ("it catches mistakes before the reviewer does") in about one afternoon.
What is actually hard; and I would be lying if I said it was mostly conceptual:
- Documents feel finished; models feel provisional. A signed PDF is psychologically safe. A live model is always open, always refutable. Reviewing logic is also more uncomfortable than reviewing numbers: you cannot tick a number and move on. This is an identity shift, and identity shifts do not happen through instruction.
- Exposure. A spreadsheet's opacity protects the preparer. A declared model makes assumptions visible and blameable. Some of the resistance is self-interested and should not be dressed up as conservatism.
- Institutional lag. Audit standards, liability rules, and billing all reward document production. Technology moves on a six-month cycle; standards move on a ten-year cycle. So there will be a long hybrid period in which models live inside firms while the deliverable is still a PDF; which is fine, as long as the PDF is a render of the model and not a re-typed copy of it. The tell that someone has not really switched: hardcoded overrides with no provenance. Spreadsheet habits survive inside model-shaped tools.
Difficulty, honestly. The conceptual switch takes a focused month of practice, not a degree. For technical accountants, auditors in methodology roles, and FP&A people, it is faster — a few weeks; they are already halfway there. For the median preparer, expect a year of uneasy practice before reviewing logic feels normal. For the profession: five to ten years, uneven, with the gap opening between firms that internalize it and firms that bolt AI onto document workflows and wonder why the errors got faster.
The single most useful thing I can do in a conversation is refuse to talk in documents. When someone hands me a spreadsheet, I decompose it; data, logic, presentation; and ask for the invariants. And never, ever, use the term "tag". Do that often enough, and the mental switch stops being a concept and becomes a habit.
Additional Information:
Comments
Post a Comment