The demo was real. The conclusion was wrong.
Somewhere in the past year, you watched AI do something genuinely impressive with a spreadsheet.
Maybe Copilot wrote a nested formula in three seconds that would once have cost you twenty minutes and a forum thread. Maybe you dropped a CSV into ChatGPT and it summarised your quarter in perfect prose. And somewhere in that moment, a conclusion quietly formed — in your head, or worse, in a board meeting:
We don't need to fix our spreadsheet problem. AI is about to fix it for us.
That conclusion is wrong — and it's wrong in a way that will quietly cost the businesses that act on it. Not because AI is over-hyped. Because the fix is aimed at the wrong layer.
What you think AI will do for the spreadsheet
The promise, as sold, comes in three parts. AI will write the formulas, so the one person who understands the file stops being a bottleneck. It will answer questions about your data, so you can ask "which jobs lost money last month?" in plain English. And it will automate the updates, so the tracker finally stops being three weeks stale.
All three features are real. All three make the spreadsheet better at being a spreadsheet. And none of them touch the reason the spreadsheet is a problem.
Last week we made the case that your load-bearing spreadsheet is really your operating manual — the only place your actual operating model has ever been written down, in a language only one or two people can read. Keep that frame, because AI runs straight into three walls it creates.
Wall one: AI inherits your errors at machine speed
Audits of real-world spreadsheets consistently find errors in roughly nine out of ten — that's the research baseline, not a horror-story outlier. Your load-bearing file has been accumulating quiet mistakes for years.
Here's what an AI does with that: it treats every one of them as ground truth.
The model doesn't know that cell G14 has been wrong since 2023, that the March tab uses the old margin formula, or that the "totals" row stopped including freight when someone re-sorted the sheet. It reads what's there and answers with total fluency. "Garbage in, garbage out" is fifty years old — what's new is that the garbage now comes back beautifully phrased, instantly, with confidence.
Remember what actually sank JPMorgan's spreadsheet-driven risk model: not the broken formula, but the silence — an error that sat unnoticed for months. AI doesn't break that silence. It makes it more articulate.
Wall two: the spreadsheet doesn't contain your operation
The operating manual in your spreadsheet has a missing volume: everything nobody ever wrote down.
The exception your operations manager applies without thinking. The client you always phone before invoicing. The pricing rule that changed in March — in practice, but never in the file. The reason job type C is quietly declined even though the sheet says it's profitable.
AI can only read what has been encoded. Ask it "can we take this job?" and it will compute an answer from the cells — not from the constraint every human in the room knows about. The critical 20% of your operation — the part that holds your margin and your risk — isn't in the file at all. No model, however large, can read what was never written.
Wall three: operations don't run on answers — they run on enforcement
A chat interface gives you an answer when you ask. An operation needs the rule to fire when nobody asks: the compliance deadline that escalates itself, the below-margin quote that gets blocked before it goes out, the client who gets flagged when their behaviour changes.
We have a line at 4What that carries our whole worldview: in a world with all the answers, the question becomes important. Knowing which question to ask, and when — that isn't a detail of the operation. That is the operation. Bolt a chat window onto a spreadsheet and you become the full-time scheduler of every question that matters. Miss one, and the answer machine sits silent while the deadline passes.
What AI does when nobody encoded the rules
There's a documented preview of where this goes. In early 2024, a Canadian tribunal ordered Air Canada to pay damages after its website chatbot confidently explained a bereavement refund policy that didn't exist. The airline argued the chatbot was effectively responsible for its own statements. The tribunal disagreed: your AI, your liability.
That's an AI doing exactly what AI does when the rules aren't encoded: improvising, fluently, on your letterhead. Now imagine the same behaviour attached not to a refund query, but to your quoting, your compliance dates, or your trust accounting.
The sequence that actually works
Here's the part the demo doesn't show — and the reason the title of this post says "not in the way you think" rather than "not at all".
AI genuinely is about to transform how operations run. But it performs in proportion to the quality of the operating knowledge it sits on. Which puts the real work one layer down.
Operational Encoding is the discipline of researching how an industry actually operates — its regulations, formulas, workflows and unwritten rules — and encoding that operating knowledge into purpose-built systems, before any software gets built. I coined the term — Schalk van der Merwe, at 4What Digital — and "before any software" was always the point. It matters doubly for AI.
Because once the operation is encoded, every wall above comes down:
- The rules exist in executable form — so AI can enforce and monitor, not just chat. The below-margin quote gets blocked by a rule, not caught by a prompt.
- The data is validated against source systems — so an answer can be verified, not just believed. The nine-in-ten error baseline stops being the foundation.
- The exceptions are modelled — so the edge cases that make language models hallucinate become defined paths instead.
- Every action leaves an audit trail — so when AI acts, you can see what it did and why. That's the difference between an operator and an improviser.
On that foundation, AI stops being a demo and starts being staff: compliance checks that run themselves, margin erosion flagged the week it starts, client communications drafted from context that's actually true.
The order is the whole game. Encode first, then automate. Reverse the order and you're automating the chaos.
The three-question AI-readiness test
Before spending anything on "AI for operations", ask:
- If you asked an AI "which of our jobs lost money last quarter?" — does the data it would read actually contain the answer? Costs allocated, hours captured, rules applied consistently?
- If it gave you an answer, could anyone verify it against a source of truth? Or would you be trusting the same file with the nine-in-ten error baseline?
- If it acted for you tomorrow — sent the quote, approved the discount — is there a written rule it would be following? Or would it be improvising, like the chatbot that invented a refund policy?
Three noes doesn't mean you have an AI problem. It means you have an encoding problem — and that's better news than it sounds, because encoding is a solvable, systematic discipline. It's the one we work in.
Where this leaves you
AI is not the escape hatch from the spreadsheet problem. It's the strongest argument ever made for finally fixing it — because every month your operation stays un-encoded is a month of compounding AI leverage handed to whoever in your industry encodes first.
If the demo has already convinced someone on your leadership team that Copilot makes the problem go away, this is the post to send them. And if you want the foundations, start with what Operational Encoding is and why your business can't quit spreadsheets.
FAQ: AI, spreadsheets and Operational Encoding
Can AI like ChatGPT or Copilot fix a business-critical spreadsheet?
Not in the way most people expect. AI can write formulas and summarise data, but it can only read what is in the file — so it inherits the spreadsheet's existing errors at machine speed, and it cannot see the unwritten rules and exceptions that live in people's heads. It makes the spreadsheet better at being a spreadsheet; it doesn't fix the operation underneath.
What is Operational Encoding?
Operational Encoding is the practice of researching how an industry actually operates — its regulations, formulas, workflows and unwritten rules — and encoding that operating knowledge into purpose-built systems. The term was coined by Schalk van der Merwe at 4What Digital. The methodology runs in three moves: decode the existing operation, encode it into a system built for that industry, and let the platform learn from live operational data.
What has to happen before AI can safely run parts of an operation?
The operation has to be encoded first: rules written in an executable form, data validated against source systems, exceptions modelled instead of ignored, and every action leaving an audit trail. Once that exists, AI stops being a chat window and becomes an operator — enforcing rules, monitoring thresholds and drafting work from context it can actually trust.
Is AI in spreadsheets useless, then?
No — as a productivity aid it's genuinely useful. Formula help, quick summaries and one-off analysis are real wins. The mistake is treating a language model bolted onto a spreadsheet as an operations layer. Answers on demand are not the same thing as rules that enforce themselves.
Schalk van der Merwe is co-founder of 4What Digital, where he leads Operational Encoding for operations-heavy businesses. Reach him at schalk@4whatmarketing.com or visit 4whatdigital.com/operations.