Local AI or cloud AI?

Cloud, when the material is allowed to go there.

A hosted model is the straightforward place to start when the data in the job is allowed to be processed there, and a hosted service is acceptable to the people who own the risk. That includes tools many teams already have: Claude, OpenAI, and Copilot.

For a stricter tenancy, Microsoft Azure AI Foundry and AWS Bedrock are managed options that still run in the cloud, under an account you control. You are accepting that provider's uptime, their retention settings, and their terms. Those settings have to be checked against the work.

Local, when it has to stay.

Local means the model runs on hardware you operate. The job does not send that material to a model provider unless you later choose to. That matters when the material cannot leave the premises, or when the team wants the system to keep working without depending on one vendor's API.

Keen Concepts can use models on hardware you already own. One option we supply is the AI Box: a compact machine, assembled in Australia, configured and supported. The specification is on the situations page. The wider service is private and local AI.

Local is more equipment to house, power and maintain. A model you host is still software that needs updates. Ongoing maintenance is part of the engagement for that reason.

You can start in one place and move.

A workflow designed around a clear job can move later. A cloud pilot is often the right first step. Moving that same job onto hardware you own does not have to mean throwing the process away, if the process was scoped properly. The reverse is also available: a local pilot can later call a cloud model for work that is allowed to leave.

Some designs are mixed from the start. The sensitive record stays local. A separate step, with material that is allowed to move, calls a hosted model. Write that split into the design.

Questions that decide it.

• What exactly would be sent: the document, a summary, or only a question?
• Who is the provider, and what do they store?
• Is there a contractual reason the data must stay in a named environment?
• Does the job need to run if that provider is unavailable?
• Who will apply updates six months from now?

If the answers are mixed, the design can be mixed. The point of the consultancy is to make that split on purpose, then build the automation around it. Tell us what cannot move.