Operational Encoding: Encode the Industry, Not the Company

Everyone is encoding companies

Palantir encodes a company. Its ontology turns one organisation's data and logic into a digital twin - a private model of that one customer, and no one else. Celonis encodes a company: it mines a single business's event logs to show how that business's processes really flow. Every vertical-AI startup you can name encodes a company too - or, more precisely, an account. This client's workflow. This client's data. This client's edge cases.

They are all excellent at it. And they are all doing the same thing: encoding one operator, one customer at a time.

There is a move one level up that almost nobody is making. Don't encode the company. Encode the industry - the regulations every firm in it must meet, the formulas the whole sector runs on, the workflow shape they all share - once, and deploy it to every operator in the vertical.

That's the whole idea. And it has a name.

What Operational Encoding is

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, co-founder of 4What Digital - in 2026. Read it again and put the weight on one word: industry. Not the company. Not the account. The industry. That single choice of altitude is what separates Operational Encoding from everything it gets compared to.

The unit of encoding is the industry, not the company

Here is the claim the rest of this post defends: the right unit of encoding is the industry, not the company.

It sounds like a small distinction. It's the whole game, for one reason - an industry is a real, encodable object, and most of it is shared. Every insurance broker in the country answers to the same regulator, runs the same core formulas, and moves a client through roughly the same lifecycle. The regulations exist before any single broker types them into anything. The compliance thresholds don't change when you walk across the road to a competitor. The great bulk of "how this industry runs" is common to every operator in it - and common things can be encoded once.

Encode a company and you have built one digital twin: powerful, bespoke, and sealed inside that customer's walls. The next customer starts from zero - their data, their build, their bill. Encode an industry and you have built the machinery every operator in it can run on, with only the last, thinnest layer - this firm's own quirks - specific to each one. One is a custom job repeated N times. The other is a single asset that deploys N times.

That's the difference between deep and compounding. Palantir and Celonis go deep on one account. Operational Encoding compounds across a market.

Why company-level encoding hits a ceiling

There is nothing wrong with encoding a company. For a single large enterprise with genuinely unique operations, a bespoke digital twin is exactly the right tool. But as a way to serve a whole industry, it has a ceiling built into it.

The knowledge doesn't travel. Everything you learn encoding one bank is locked to that bank's data and that bank's contract. The next bank down the road - same regulator, same instruments, same reporting obligations - is a fresh engagement that starts again from its own logs. The shared majority gets re-derived, re-billed, and re-built every single time. That isn't a moat; it's a treadmill with good margins.

It also means the vendor's real asset stays surprisingly thin. What compounds for a company-level encoder is delivery capability - they get faster at doing the same bespoke work. What never compounds is an owned model of the industry itself, because they never built one. They built a hundred models of a hundred customers.

Encode the industry and the arrow reverses. Every operator you add sharpens the one shared model - and the shared model is the thing you own.

The mechanism: three moves

The discipline runs in three moves. They're the same three whether you're encoding one operation or a whole sector - the altitude changes what "decode" points at, not the sequence.

Decode. Research the industry the way a practitioner lives it - the regulations, the formulas, the workflows, and the unwritten rules that never make it into any manual. Existing spreadsheets across the sector are primary source material, because that's where the real operating model already lives, firm by firm.

Encode. Build the system around what was decoded, so it holds those rules natively and enforces them automatically - and build the shared machinery once, at the level of the industry, not once per customer.

Learn. The platform runs on live operational data from every operator on it, and gets sharper the longer it runs - surfacing patterns no single firm's spreadsheet ever could, because no single firm can see the whole sector at once.

"But you still build software" - yes, second

One objection, cleared quickly. None of this means software doesn't get built. It does - it's just the second move, not the first. You encode, then you build on top of what you encoded.

I've argued that sequence at the level of a single business already: buy the tool, bolt on the AI, commission the custom build, and the operation still ends up back in the spreadsheet, because nobody decoded it first. This post is the level above that. Same rule - decode before you build - applied to an entire industry instead of one firm. If the sequence is new to you, start there. If it isn't, the point here isn't the order. It's the altitude.

What it is not

Four things Operational Encoding gets mistaken for - and the altitude tells them all apart.

It is not a company-level digital twin. Palantir's ontology and Celonis's process model encode one organisation from its own data and live inside its walls. Operational Encoding encodes the industry - before any single client's data exists - and deploys to all of them. Company-deep versus industry-wide.

It is not a vertical AI agent. An agent is a worker that does one task in one industry - claims, prior-auth, invoicing. It needs an encoded industry underneath it to be correct. The agent is disposable and gets rebuilt as models improve; the encoded industry is the asset that outlives it. (More on that fight later in the series.)

It is not consulting. A consultancy researches the operation and leaves; the knowledge lands in a slide deck that goes stale. Operational Encoding produces a running system, and the model it builds is owned and redeployed - not shipped once and abandoned.

It is not fine-tuning a model on industry data. That's a downstream technique. Encoding is what decides what's worth learning in the first place - which regulations must be enforced, which formulas must be exact, which workflow the operator actually lives. A fine-tune with no encoded industry underneath it is a confident intern with no supervisor.

Ten industries, not ten clients

This isn't a thought experiment, and the proof is in the noun. We've done the decode work - practitioner-level, regulation by regulation - across ten industries:

Law · Accounting · Insurance Broking · Investment Advisory · Property · Life Insurance · Data Consultancies · Venues & Hospitality · Film Production · Retail Supply.

Not ten clients. Ten industries - each one an operating model that deploys to every firm in it. What stays behind the gate is the how: the specific formulas, the data models, the research methodology. The diagnosis is public; the machinery isn't. You share the pain freely, prove you can solve it, and gate the method - because the method is the product.

The economics follow

Encoding once and deploying many is also what makes the economics work - but that's a fight of its own, and I take it on directly in the next post: why so much of what's sold today as "AI revenue" is really consulting in a trenchcoat, and why an owned industry model is the only thing that breaks the trap. The point here sits upstream of the money. Get the unit right - the industry, not the account - and the economics take care of themselves.

Why name it at all

Because a thing without a name can't be owned, defended, or bought. The market has agreed the moat has moved - a16z says it moved to proprietary data, Bessemer to "systems of action," Foundation Capital to "service-as-software" - and every one of those framings still encodes at the level of a company. None of them names the altitude. Naming it is how "encode the industry, not the company" stops being a nuance and becomes a position you can plant a flag on.

So here's the flag. Whoever encodes an industry first owns the layer every operator in it runs on - not one customer's twin, but the whole vertical's operating model. That's the asset. Everything downstream - the platform, the agents, the AI - gets built on top of it and rebuilt as the tools change. The encoding is the part that lasts.

We encode industries before we build software. This is post one of a series that takes the claim into the fights it starts - with disguised-consulting AI revenue, with vertical-agent hype, with the company-level digital twin head-on, and with the demand side of the same idea. For the original, ground-level diagnosis of why operations get stuck in the first place, start with the founding essay. For the full definition, origin and methodology in one place, it lives on the Operational Encoding page.

FAQ: Operational Encoding, defined

What is Operational Encoding?

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. The term was coined by Schalk van der Merwe, co-founder of 4What Digital, in 2026. It runs in three moves: decode the operation as a practitioner lives it, encode it into a system built for that industry, and learn from live operational data.

How is Operational Encoding different from Palantir ontology or process mining like Celonis?

Palantir encodes one company's data into a digital twin, and Celonis mines one company's event logs to map its processes. Both are company-level: they start from a single client's data and stay inside that client's walls, so the next customer starts again from zero. Operational Encoding is industry-level - it encodes how the whole vertical operates (the regulations, formulas and workflow shape every operator shares) once, before any single client's data exists, and deploys that to every operator in the industry.

How is Operational Encoding different from building vertical AI or AI agents?

A vertical AI agent is a worker built to do one task in one industry - claims, prior-auth, invoicing. Operational Encoding is the layer underneath it: the encoded model of how the industry actually runs that the agent needs in order to be correct. The agent is disposable and gets rebuilt as models improve; the encoded industry is the asset that compounds. You don't build the agent until you've encoded the vertical it's supposed to work in.

Who coined the term Operational Encoding?

Operational Encoding was coined by Schalk van der Merwe, co-founder of 4What Digital, in 2026, to name the discipline of encoding an industry's operating knowledge into purpose-built systems before software gets built.

What does "encode the industry, not the company" mean?

It means the reusable asset is a model of how the whole vertical operates - the regulations, formulas and workflow shape every operator shares - rather than a bespoke model of one client's data. Company-level tools like Palantir and Celonis encode a single organisation and stay inside its walls, so the shared knowledge gets re-derived with every new customer. Operational Encoding encodes the industry once and deploys it to every operator in it, so the model compounds across the market instead of restarting each time.

Schalk van der Merwe is co-founder of 4What Digital, where he coined Operational Encoding and leads it across ten industries. Reach him at schalk@4whatmarketing.com or visit 4whatdigital.com/operational-encoding.