Every organization serious about AI ultimately asks the same structural questions: who owns it? It’s not who owns the permit, it’s a procurement question. Who owns the capabilities, standards, priorities, and decisions about what gets built and what gets discontinued.
The answer is your AI operating model, and in practice it takes one of three forms: centralized, federated, or hub and spoke. I’ve seen all three in client engagements over the past two years, and I’ll provide the conclusions first. The centralized model creates bottlenecks. Mixed models create chaos. Hub and spoke is the only form of scale I’ve seen, and the rest of this post will discuss why.
This is the first post in a series on AI operating models. Here we deal with the shape. Future posts will cover what’s inside: governance and decision-making rights, capability levels, ways of working, measurements, and the operating rhythms that keep it all alive.
Why the shape of your AI operating model matters
Most AI launches are not technological failures. They failed in terms of structure. The license expires, a communications rollout occurs, a few weeks of enthusiastic first wave experimentation, and then usage plateaus because no one knows what will happen next. There is no path from a great idea in one team to a repeating pattern across twelve teams. No one is deciding which use cases are worth investing in. There are no quality thresholds, so half the organization tacitly concludes that the output cannot be trusted.
The shape of the operating model determines whether things happen as planned or not at all. So before you debate about tools, models or agents, first define the form. These are the three options.
First form: centralized model
In a centralized model, one team owns everything about the AI. Typically these companies are badged as AI centers of excellence or are within IT. Every use case, every swift library, every agent, every approval flows through this one team. Business unit requests; central team delivers.
Why this is interesting
Centralization is the right choice for organizations with a strong governance culture, and centralization does have real strengths. Standards are consistent because one team sets them. Risks are easier to manage because there is a single point of control. Rare skills are collected rather than diluted. During the first 90 days of an AI program, centralized teams often emerge as winners.
Why fail on a large scale
Then the requests came. A central team of six people cannot serve an organization of six thousand people. The backlog grows, business units wait months for use cases they could prototype in a week, and two predictable things happen. First, queues kill momentum; people who wait eight weeks for answers stop asking. Second, the emergence of the shadow of AI. Teams began solving their own problems with whatever tools they could reach, beyond the standards carefully created by the central team. You will experience the worst: congestion and uncontrolled expansion, at the same time.
The deeper problem is proximity. The central team does not know its work. Financial close, claims processing, field engineering workflows: the highest value use cases lie in the details of how the work is actually done, and the central team is structurally too far away from those details to find them.
Second form: combined model
The combined model is the opposite approach. Each business unit drives its own AI adoption. Each function buys, builds, and organizes itself. Sometimes this is a deliberate choice; more often it is what you get by default when no one has made a choice at all.
Why this is interesting
Federation solves the proximity problem instantly. The people closest to the work choose the use cases, so relevance is high and no one waits in line. The startup energy is real: true innovation emerges quickly, and teams that capitalize on it reap the benefits quickly. If you assessed the model at week six, you would call it a success.
Why fail on a large scale
Assess at the twelfth month, and the picture is different. Each unit has reinvented the same wheel. Five teams had created five versions of the same meeting summary workflow, but none of them were shared. Fast quality varies widely because there are no common standards and no quality limits. Governance is inconsistent, and in regulated sectors, this is not a bad thing; it is a responsibility. And because nothing is measured the same way twice, no one can tell the board what the organization is actually getting from its spending.
Joint AI adoption results in islands of capability and seas of duplication. Lessons learned in one unit never move around. You pay for the same learning curve over and over again, and the gap between your best team and your median team gets wider every quarter.
Form three: hub and spoke model
Hub-and-spoke leverages the strengths of both and is purposely designed to mitigate their failure modes. A small center has things that need to be consistent: standards, governance, fast and pattern libraries, quality standards, measurements, and an acceptance process that decides which use cases get investment. The spokes are the business units, each of which has leaders who are close to the work, discovering use cases, adapting key patterns into their functions, and passing what they learn to the hub.
An important design point is what the hub doesn’t do. It doesn’t provide every use case; that’s the centralized trap. This allows spokespeople to convey, and harvest what they learn so that patterns proven in one function become the starting point for every other function. Compound values instead of duplicating.
Why this works when others fail
Set the three forms against the two failure modes, and the logic is clear:
- Congestion problem. Delivery capacity depends on the radius, so it is tailored to the organization, not to the number of employees in a central team. The hub is small by design and remains small.
- Clutter problem. Standards, governance and measurement are at the center, so consistency is structural rather than aspirational. One standard of quality, one way of calculating value, one library that everyone uses.
- Proximity issues. Champions sit on the job. Use cases are found in existing values, unpredictable from the center.
- Duplication problem. A feedback loop from spoke to hub turns one team’s solution into everyone’s pattern. That circle is the single biggest difference between organizations that merge and organizations that don’t.
It’s worth noting that hub-and-spoke is not free. This demands real investment in the network of champions, and requires a rhythm of operations, a regular rhythm of decision making, review and reinvestment, or the hub will splinter into bulletins and the spokes will fall back to the federation by default. The form is required; that’s not enough. That’s what the rest of this series covers.
Choosing in practice
If you recognize your organization in the centralized description, the step to take is not to disband the central team; it is to reuse them. Stop sending, start activating. If you recognize the combined description, the step is to set up a small hub and start harvesting what your best team has built, because the raw materials for your pattern library already exist; it’s just stuck in a silo.
Whatever the goal, the goal is the same. In every interaction where I have seen the value of AI truly extend beyond the initial enthusiasts, there is an underlying hub-and-spoke structure. I haven’t seen a counterexample, and I have.
In conclusion
The shape of your AI operating model is a decision, and not building the model is itself a decision, a decision that defaults you to federation and its costs. The centralized model buys consistency at the expense of bottlenecks. The combined model buys speed at the expense of chaos and duplication. Hubs and spokes are the only shapes that give you consistency and closeness at the same time, and it is this shape that this series will build on.
Get the white paper
This post discusses model form. complete white paper, Operating Model for Enterprise AI Enablementgoes further: six components in a hub and spoke structure, a four-stage maturity model to position your organization squarely, and a data and permissions readiness checklist to execute before wide rollout. Download the white paper and use it as a working reference for the rest of this series.
Useful Links
AI Empowerment Operating Model | Gethin Ellis
Why Your AI Output Is Inconsistent (And Nine Frameworks That Fix It)
Microsoft Foundry vs AWS Bedrock vs Vertex AI: Which Wins?
PakarPBN
A Private Blog Network (PBN) is a collection of websites that are controlled by a single individual or organization and used primarily to build backlinks to a “money site” in order to influence its ranking in search engines such as Google. The core idea behind a PBN is based on the importance of backlinks in Google’s ranking algorithm. Since Google views backlinks as signals of authority and trust, some website owners attempt to artificially create these signals through a controlled network of sites.
In a typical PBN setup, the owner acquires expired or aged domains that already have existing authority, backlinks, and history. These domains are rebuilt with new content and hosted separately, often using different IP addresses, hosting providers, themes, and ownership details to make them appear unrelated. Within the content published on these sites, links are strategically placed that point to the main website the owner wants to rank higher. By doing this, the owner attempts to pass link equity (also known as “link juice”) from the PBN sites to the target website.
The purpose of a PBN is to give the impression that the target website is naturally earning links from multiple independent sources. If done effectively, this can temporarily improve keyword rankings, increase organic visibility, and drive more traffic from search results.
