How to Implement an LLM in a Firm

Bringing an LLM into your firm is a lot like deciding how you’ll eat. We’ve already covered the ingredients you will need to make an AI strategy real. Now comes the architectural choice: do you completely outsource the cooking to a restaurant, hire your own private chef who learns your tastes or co-hire a chef in a shared cloud kitchen?

So how will you balance speed, safety, cost and culture to choose the architecture that fits?

Each executive sees a different facet of the LLM - cost, security, compliance, adoption or speed of delivery and each perspective is valid. The real challenge is harmonising them into a single, coherent strategy.

So which option best fits your firm’s appetite?

This is the enterprise version of ChatGPT, Claude, Gemini - the ready-to-serve option.

The upside is control. Every dish is prepared to your specifications with ingredients that never leave your home. Done well, this becomes a strategic asset - tailoring meals that no restaurant could replicate.

The biggest downside is time to market. This demands significant investment in cutting-edge hardware hardware (like NVIDIA GPUs), expertise to manage data systems and highly skilled staff - all of which adds up quickly. Set-up can take months or even years and once built, you are hook for every upgrade, every cleanup, every experiment.

The opportunity is leadership: to build something unique, deeply woven into workflows, defensible as intellectual property and hard for rivals to copy. The threat is equally sharp: talent wars, rapid obsolescence, high resource costs and the risk of investing heavily before the first meal ever hits the table.

This is co-hosted enterprise solutions in Cloud. For example, Open AI APIs co-hosted in Microsoft Azure. Here you don’t run a full kitchen yourself but you share access to a professionally managed one. You co-hire a skilled chef who uses their ovens, equipment and safety standards - while still letting you bring in your own recipes and ingredients.

The strength is balance. You can customise outputs with your own data, while leaning on big vendor’s infrastructure for scale, compliance and security. You can start faster than a full build but with more control than outsourcing everything.

Weaknesses include partial dependency. Costs can scale unpredictably. Integration work to connect the shared kitchen to your existing processes often involves complex orchestration and data synchronisation challenges, limited transparency into how the underlying model “cooks” results.

The opportunity lies in leverage from rich vendor ecosystems making adoption smoother. The threat is lock-in: once your recipes depend on one provider’s kit, it’s hard to leave.

Then there are tools like Microsoft Copilot - think Korean BBQ. The grill and the ingredients are provided but you need to cook the meal yourself at the table.

The upside is accessibility: these tools are built into apps people already use and can deliver quick productivity gains.

The catch? These meals are only as good as the diners’ skills - without training, some come out perfect, others burnt.

This is a halfway solution. Useful tactically but not a substitute for a deliberate LLM architecture.

Do you want the speed of a restaurant kitchen, the control of a custom-built private kitchen or the balance of a shared kitchen?

Maybe the real question isn’t about kitchens at all - it’s about the kind of organisation you want to become. Do you want ultimate control, keeping everything in-house? Are you comfortable with standard, ready-made meals if they’re safe and cost-efficient? Or do you prefer co-hiring a chef in a shared kitchen, balancing speed with some control? Your choice reflects your priorities, risk appetite and long-term vision.

Proprietary kitchens take years to build and many firms won’t see a competitive advantage worth the cost. At the same time, letting data leave your walls may feel risky but with safeguards it’s often manageable. The real choice is where your firm can truly differentiate and how long you are willing to wait for it.