What does "AI as an operating model" actually mean?

Ask ten companies about their AI strategy and nine will show you features. A summariser here, a chatbot there, a copilot in the sales tool. Ask how the company runs differently because of AI and the room goes quiet.

That gap is the whole subject of this site, so it's worth defining properly.

AI as an operating model means AI is built into how your company senses, decides, and acts — the loops that run the business — rather than added to what your product does. The test is simple: if you switched it off, would your product lose a capability, or would your company stop working the way it works?

If it's the first, you have AI features. If it's the second, AI is part of your operating model. Most companies have the first and believe they have the second.

A feature changes the product. An operating model changes the company.

An AI feature is a capability: summarise this document, draft this reply, score this lead. Features are visible, sellable, and easy to point at in a board deck. They're also easy to copy — whatever feature you shipped this quarter, your competitor ships next quarter, usually on the same foundation models.

An operating model is the set of loops a company runs on: how signal from the market gets in, how it becomes a decision, how the decision becomes shipped work, and how the result feeds back. It's mostly invisible — cadences, handoffs, meeting rhythms, who decides what. Nobody screenshots it for the board deck. But it's where the compounding is, because a rewired loop gets faster and smarter every cycle, and it can't be copied by a competitor reading your release notes.

The distinction matters because the two get funded and measured as if they were the same thing. They aren't. A feature is a product bet. An operating model is a company bet.

The three loops AI actually rewires

Strip any company's operating model down and there are three loops that matter. This is the lens I use in my own work — developed through practice, broken and rebuilt more than once.

1. The signal loop — how reality gets in. Customer conversations, support themes, usage data, competitor moves, sales objections. In most companies this arrives pre-filtered: summarised by one team, reinterpreted by another, then presented monthly as three bullet points. Each filter is well-meaning, and each one drifts the picture further from the market. AI changes the economics here completely: it can ingest the raw feed — all of it — and cluster, connect, and track it continuously. The design rule I hold to: no human pre-filtering on the way in. Humans interpret; they shouldn't gatekeep.

2. The synthesis loop — how signal becomes a decision-ready picture. This is the layer most companies staff with meetings. Status meetings, alignment meetings, pre-meetings for meetings — most of them exist to move information between heads, not to decide anything. AI does this job better: turning the clustered signal into themes, trends, and framed options, refreshed on a rhythm instead of assembled heroically before each quarterly review. When the synthesis loop works, a strange thing happens to the calendar: the meetings that survive are the ones where something gets decided.

3. The judgment loop — where humans decide. This is the loop AI should not absorb, and the one lazy designs blur. "Human in the loop" usually means a human rubber-stamping everywhere, which is just friction wearing a safety vest. The better design puts humans at the genuine judgment points — the calls that need accountability, taste, or a view of consequences the data can't carry — and gives them a decision-ready picture when they get there. Fewer decision points, better-framed decisions, humans doing the part only humans can do.

What it looks like in practice

Two versions of the same company.

The feature version. Support gets an AI tool that drafts replies. Tickets close 20% faster. The team works exactly the way it always has; the roadmap is still built from a quarterly survey and the loudest customer in the room. Real value — genuinely — but nothing about how the company senses or decides has changed.

The operating-model version. Support conversations, sales calls, and product usage flow into one synthesis loop. Every week the product team sees a decision-ready picture: what's rising, what it connects to, what options it suggests. The monthly "voice of customer" deck no longer exists because it's obsolete on arrival. Roadmap decisions happen against evidence that's days old, not quarters old. The support team still has the drafting tool — but now the company works differently.

Same models, same vendors, wildly different outcome. The difference was never the technology. It was whether anyone redesigned the loops around it.

Is this just automation with better branding?

No — and the difference is worth being precise about. Automation takes a fixed process and executes it faster. An AI operating model changes what the process is: which information reaches whom, which meetings exist, where decisions happen, what a role spends its week on. Automation optimises the current loops. An operating model redesign asks whether those loops should exist at all. Most of the value I've seen sits in the second question, which is exactly the question a tool purchase never forces you to ask.

Why most AI transformation stalls here

The pattern is consistent: the model was fine, the workflow around it was never redesigned. A company buys capable tools, drops them into unchanged loops, and gets a 10% efficiency gain plus a licence bill. The pilot "proves" AI is overhyped, and everyone goes back to the quarterly deck. The failure wasn't technical. It was treating an operating-model problem as a procurement decision.

This is also why the work is genuinely hard — and why it compounds for the companies that do it. Redesigning loops means changing meetings people are attached to, decision rights people fought for, and information flows people quietly control. That's operating work, not IT work. It's also why a competitor can't copy it by buying the same tools.

How to start (without a transformation programme)

Don't start with a strategy document. Pick one loop — signal is usually the right first choice, because the pain is obvious and the politics are mild — and rebuild it end to end: raw signal in, AI synthesis in the middle, humans deciding at the end. Run it on a weekly rhythm for six weeks. Kill the meetings and reports it makes redundant — this is the step most people skip, and it's the step that makes it an operating-model change rather than another tool.

Then let the result argue for the next loop.

---

Everything above comes from running these loops myself — and breaking them often enough to keep the writing honest. That's what the rest of this site is: field notes, not theory.

Building something and wrestling with this? Tell me what you're building — or start with how I work.