Practical answers to the questions businesses actually ask about putting AI to work.
Latest answer
Create a workspace for each team, assign models, capabilities and skills to each one, and separate what's available from what's deployed — so every team gets the AI its work justifies.
Choose which AI providers your data may reach and on what terms, run the most sensitive work on model endpoints inside your own environment, and put your corporate identity in front of all of it.
Make the platform the constant and the models interchangeable: connect providers through API keys, AWS Bedrock, the subscription you already pay for, or local endpoints — so adopting next quarter's best model is an admin settings change, not a migration.
If AI works, dependency follows — manage it like any other critical dependency: switchable providers so no vendor is a single point of failure, observable usage and cost, and a person approving consequential outbound actions.
Author or import your security team's own skills into a governed catalog with a safety scan on import, run analysis in sandboxes with pre-installed libraries, hold service credentials centrally so they're injected at runtime — never through the model — and let the team build small governed apps, all deployed to their workspace only.
Let people create small working business apps by describing what they need, publish them to a governed catalog, and run every app in a sandbox that can only call the actions it was explicitly allowed.
Put the workspace in the cloud and make every device a window onto it — start a task at the desk, pick it up on the phone exactly where you left off, with every device visible and revocable from the web.
Make AI act as the signed-in user, never a super-user: Microsoft 365 access runs through each person's own sign-in, on-prem SharePoint Server is read via Kerberos delegation under existing permissions, and identity comes from Microsoft Entra — fix permissions once and AI inherits the fix.
You don't have to pick a winner: consolidate — one governed workspace where multiple providers' models sit in the same picker, the ChatGPT subscription you already pay for connects in, and the work happens where your Microsoft 365 tools already are.
Log every AI request at the platform layer — usage, cost, and logs per model, key, and endpoint — attribute activity to named users, and require a recorded human approval before consequential output leaves the business.
Draw the line by consequence — anything that leaves the organisation is prepared by AI but reviewed, edited and explicitly approved by a person in an editable panel, so the interface enforces the rule instead of a policy document.
Don't take the models away — connect the ChatGPT subscription your organisation already pays for so staff keep the models they chose, inside a governed workspace, then widen model choice and publish the team's best prompts so the sanctioned option wins by preference, not prohibition.
Run the language model on an OpenAI-compatible endpoint inside your own network, choose the embedding model behind your knowledge sources, and organise the files into governed, refreshed document stores — so the chain from document to answer runs on infrastructure you control.
Bring AI to where the work already lives: connect email, calendar, files and team chat with each person under their own sign-in, add the business apps your teams run on, and hold credentials centrally so secrets are injected at runtime — never through the model.
Bans push AI use onto phones and personal accounts. Give staff a governed workspace with models as capable as the ones they're sneaking off to use, signed in with their work identity, with usage visible — when the safe path is also the best path, the paste problem mostly solves itself.
Route routine work to cheaper models, reserve frontier models for the problems that earn them, run some models on infrastructure you own, and measure cost per model and endpoint — so the invoice becomes a set of choices you control.
Match the model to the job: route routine work to cheaper models, run local models for high-volume low-stakes tasks, and watch spend per model and per team — so frontier prices are paid only when the work earns them.
