AI Operating Model for Growing Businesses: Three Layers, Three Owners

An AI operating model is how a business organises its AI: what it knows, how work gets done and who is responsible. For a growing business, it fits in three layers. The bottom layer is the company's knowledge, the middle is written workflows, and the top is the agents that run them. Each layer needs a named owner.

What is an AI operating model?

An AI operating model describes how an organisation runs AI day to day: where the knowledge comes from, how work moves from request to approved output, who owns each part, and how results are measured. Large companies build centres of excellence for this. A growing business needs something lighter and more concrete.

Consultancy definitions usually cover people, process, governance and technology, and compare structures such as centralised, federated or hub-and-spoke teams. Those structures assume a dedicated AI function. Most businesses of 10 to 200 people do not have one, and should not need one.

A more useful frame for them is the three layers of a Company OS. Each layer answers one question. What does our AI know? How is each job done? What actually does the work?

What are the three layers?

There are three layers: data, workflows and agents. Data means the company brain, the canonical facts about the business. Workflows are how each job is done, written down. Agents are the AI that carries out those workflows, reading the brain first and stopping for a person before anything goes out.

The order matters. Agents built on empty data and unwritten workflows produce fast, generic work. Build from the bottom up, even if you start with only one workflow.

What goes wrong at each layer?

Each layer has its own failure. Data fails when facts go stale or conflict. Workflows fail when methods live in people's heads or solve the wrong problem. Agents fail when agents have too much access or nobody checks their output. Knowing which layer failed tells you what to fix.

Data layer: stale or conflicting facts. The AI quotes last year's pricing because nobody updated the file. The fix is ownership and regular sweeps of new calls, tickets and emails.

Workflow layer: the wrong job, or no method. According to RAND's 2024 study, the most common root cause of AI project failure is that the problem is misunderstood or miscommunicated. In practice, that looks like an agent built for a job nobody defined.

Agent layer: too much reach, too little review. An agent with write access to the CRM and no approval point can do real damage quickly. Start read only and keep a person approving anything published, sent or spent.

The cost of getting this wrong is visible in the data. According to MIT NANDA's 2025 study, 95% of organisations it studied saw no measurable return from generative AI, and the authors blamed systems that do not adapt to context or improve.

Who should own each layer?

Each layer needs one named owner. The data layer is owned set by set, such as the product lead for product facts and the sales lead for objections. Each workflow is owned by the person who does that job best. The agent layer is owned by whoever is responsible for AI overall, including access and approval rules.

Accountability is also where official guidance starts. According to the National AI Centre's Guidance for AI Adoption, the first of six essential practices is to decide who is accountable. The last is to maintain human control.

How is this different from an enterprise AI operating model?

An enterprise model organises teams around AI: a central function, embedded specialists and governance boards. A growing business organises AI around the work it already does. There is no new department; there are three layers, a handful of owners, and approval points where they matter.

That does not make it less rigorous. Identical questions get answered, just closer to the work: who decides what the AI may say, who keeps it current, and who approves what goes out. The answers are named people, not committees.

For the data layer in depth, see the guide to what a company brain holds.

Skein builds these three layers inside growing B2B businesses, with owners named from the start; the Company OS page shows the architecture. That suits a business of roughly 10 to 200 people without a dedicated AI team. A large organisation with an established data function will usually need the fuller enterprise model.

Frequently asked questions

What is an AI operating model?

An AI operating model is how a business organises AI day to day: where its knowledge comes from, how work moves from request to approved output, who owns each part and how results are measured. For a growing business, it can be described in three layers: data, workflows and agents.

What are the layers of an AI operating model?

For a growing business, three layers work well. First, a data layer holds canonical facts about the company. Next, a workflow layer holds written methods for each job. Finally, an agent layer runs those workflows, reading the data first and stopping for a person's approval before anything is published, sent or spent.

Who should own AI in a small business?

Ownership works best split by layer. Each set of facts has an owner, such as the product lead for product facts. Each workflow is owned by its best practitioner. One person owns the agents overall, including access and approval rules, so accountability is never shared vaguely.

Do small businesses need an AI centre of excellence?

Rarely. A centre of excellence suits large organisations with dedicated AI teams. A business of 10 to 200 people gets the same benefits from three clearly owned layers, written workflows and fixed approval points, without creating a new department or hiring specialists.

How do you build an AI operating model?

Build from the bottom up. Start the data layer with a few canonical sets, such as product facts and objections. Write down one frequent workflow. Then add an agent for it, read only at first, with a named approver. Name an owner for each layer and expand one workflow at a time.

Related reading