I use AI every day. Claude for most of my writing and a fair bit of my code, ChatGPT when I want a second opinion, Copilot because it’s sat there in Teams whether I asked for it or not. That’s not unusual any more, and if you’re reading this you’re probably the same.

The question I keep getting asked is this: “when do you stop using those tools and start looking at Microsoft Foundry?”

Short answer; when the AI stops being something you use and starts being something the business depends on.

two boxers facing each others and touching gloves before a fight in a well lit practise gym

Experimenting is easy, trusting it is the hard part

Consumer AI tools are brilliant at helping one person get one job done faster. If that’s genuinely what you need, you don’t need Foundry. Buy the licences, teach people how to prompt properly, and get on with it. I’d rather tell a client that than sell them a platform they’ll never use.

It changes the moment AI touches a process. When it needs access to your data, runs without anyone watching, or produces an answer a customer actually sees, you’ve got a different problem on your hands. Now you need to know what it touched, who asked it, and why it said what it said.

2 scientitsts i white lab costs and goggles with a variety of vials and flasks filled with coloured liquids and they are mixing between them

What Microsoft Foundry actually is

I usually describe it as Fabric for AI. One place to build, deploy and run things, instead of a model over here, a vector store over there, and a spreadsheet of API keys that nobody wants to admit owning. Hundreds of models, secure connections to your data, agents, monitoring and governance, all in the same platform.

It isn’t a replacement for Claude or ChatGPT. Those still sit open on my second monitor.

The microwave and the commercial kitchen

The analogy I keep coming back to: consumer AI is a microwave, Foundry is a commercial kitchen.

Nothing wrong with a microwave. I use one most days. But if you’re feeding two hundred covers a night you need consistency, hygiene standards, someone accountable when it goes wrong, and a record of what came out of the kitchen and when. Enterprise AI is the same. When it’s supporting customers or making decisions, “it usually gets it right” isn’t an answer you can give an auditor.

a chef in white shirt and black chef apron, with many bowls in front of him full of food and twisting a pepper mill onto the food

Where the value actually shows up

Nobody I work with has ever asked for AI because it’s exciting. They ask because something is slow, or expensive, or being done by a person who’d rather be doing almost anything else.

The things that tend to matter in practice:

  • Agents that connect properly to the systems you already run. A Zendesk triage agent and a HubSpot contact enrichment agent in production. Neither of those was ever going to happen in a chat window.
  • Picking the right model per job. Something small and cheap for classification, something heavier for the reasoning work. Not being tied to one provider is worth more than people realise, particularly when prices move.
  • Seeing usage and cost in one place, ideally before finance comes asking.
  • Governance and audit trails from day one, rather than bolted on in a panic when the first security questionnaire lands.

That’s usually the difference between a proof of concept everyone was impressed by in March and something still running in production in December.

My rule of thumb

If I’m writing an email, summarising a document or getting unstuck on some Python, I’ll reach for Claude or ChatGPT without thinking about it.

If it needs to run at 3am, talk to a line of business system, or survive a security review, it goes in Foundry.

Most businesses need both, and there’s no prize for picking a side. The mistake I see most often is asking the microwave to run the restaurant, and then concluding that AI doesn’t work.

If you’re somewhere in the middle of that decision, I’m always happy to talk it through.

Tags: , , ,