Microsoft AI Foundry: a practical guide for UK SMBs

Microsoft AI Foundry is the platform Microsoft uses to build, deploy and govern AI on Azure. It was formerly called Azure AI Foundry; Microsoft renamed it in 2025 and the old name still turns up in slide decks and partner conversations. For a UK SMB, it is the engine room for AI you build yourself, not a tool you hand to staff. This guide is written for finance directors, managing directors and operations directors, not developers.

Lanmark is a Microsoft Direct Cloud Solution Provider, which means we hold the licensing and support relationship directly with Microsoft rather than through an intermediary. We build on Azure AI stacks and Microsoft AI Foundry on that basis.

What Microsoft AI Foundry is, and what it is not

Microsoft 365 Copilot is a user-facing product that sits in front of people. Microsoft AI Foundry is the infrastructure that lets an organisation build its own AI products.

It contains four things that matter to most SMBs:

• A catalogue of foundation models from several providers, each available behind a Microsoft commercial agreement so the data does not leave the Microsoft trust boundary.
• A workspace for grounding those models on your own data. The technical name is retrieval-augmented generation (RAG). In plain English, the model looks up the answer in your files before it writes a reply.
• A workspace for building AI agents that plan and carry out multi-step tasks.
• A governance layer giving you content safety, prompt security, observability and audit logging across all of the above.

It is not a productivity tool you hand to staff, not a website chatbot you can buy off the shelf, and not an alternative to Microsoft 365.

Microsoft AI Foundry compared with Microsoft 365 Copilot

The most useful distinction: Microsoft 365 Copilot is consumed by your users. Microsoft AI Foundry is consumed by your business systems.

Copilot helps your team draft, summarise and answer questions inside Word, Outlook, Excel and Teams. It is licensed per user at a published UK list price and bought as a productivity tool. Foundry is licensed by consumption: a charge per million tokens of model usage plus storage and other Azure consumption. It is bought as a platform, by a project sponsor, to build something specific.

Most SMBs need Copilot first and Foundry second. Some need only Copilot for a long time. A smaller group, where the value of AI sits in answering questions over their own data or automating multi-step tasks, will need Foundry sooner. Our guide to Microsoft 365 Copilot Business for UK SMBs covers the Copilot side.

When a UK SMB actually needs Microsoft AI Foundry

If the AI you want involves your own documents, your own records or a multi-step business process, you are likely on the Foundry side of the line. If you simply want staff to be faster in the apps they already use, you are on the Copilot side.

The four most common SMB use cases that justify Foundry today:

• Knowledge agent: a grounded internal chatbot that answers questions over your own policies, contracts and SharePoint files.
• Customer agent: an assistant that reads product documentation, knowledge base articles and structured data and responds accurately.
• Triage agent: a back-office assistant that reads incoming emails or documents, classifies them, drafts a reply and queues an action for human approval.
• Process agent: a specialist assistant for a regulated process where a generic Copilot is not controlled enough, for example a compliance check or a financial review.

If none of those resembles what you are trying to do, you probably do not need Foundry yet. Copilot, possibly Copilot Studio, and good Microsoft 365 hygiene will take you a long way before the platform underneath becomes the bottleneck.

What you can build with Microsoft AI Foundry

Each build starts with a foundation model from the catalogue: a current general-purpose model where reasoning quality matters, and a smaller, cheaper model for high-volume, cost-sensitive tasks. The model is grounded against your own data using a vector index, prompt-engineered for the task, wrapped in a content-safety layer that filters harmful inputs and outputs, and exposed to your own application or to a Copilot Studio agent.

Foundry also supports light fine-tuning, where the model is taught the tone or domain language of your business. Most SMBs do not need this in year one. Grounding the model on good data, plus careful prompting, outperforms naive fine-tuning on most operational tasks.

The output is usually invisible to users. They see a chatbot, a recommendation in their CRM, an enriched record in their ERP, or a draft reply in Outlook.

Microsoft AI Foundry pricing, in plain English

Foundry is not a per-user licence. It is consumption-based, and the consumption sits inside your Azure subscription. There are three meaningful cost lines:

• Model usage, charged per million input and output tokens and varying by model.
• Vector index and storage, in Azure AI Search or Azure Cosmos DB, which for an SMB-scale knowledge base is a modest monthly line.
• Content safety, monitoring and orchestration, which is largely included but can add a small premium at higher tiers.

As an indicative planning range rather than a quotation, a project of the scale described here would typically sit in the hundreds to low thousands of pounds per month in Azure consumption once live, plus the partner build cost, usually structured as a fixed-scope four to eight week pilot. The figure moves with data volume, query rate and the model you choose, and Microsoft revises its own pricing regularly. A costed estimate once the use case is known is more useful than any range published on a website.

Governance, security and the UK regulatory context

For most UK SMBs, governance is the reason Foundry gets approved over a generic build on a consumer AI service. The platform sits inside your Microsoft tenant, runs on Azure, respects Entra ID identity, integrates with Microsoft Purview classification and data loss prevention, and offers region-locked data residency. UK South and UK West regions are available for both compute and storage, which matters for FCA-regulated firms, NHS suppliers and any organisation whose clients ask due diligence questions.

It also provides content safety filters, prompt injection protections and the audit trail you need to show that AI is being used responsibly.

The platform is only as good as its setup. The most common governance failure is not technical: it is allowing a Foundry build to read SharePoint content that the user was never supposed to see. An AI agent inherits whatever permissions you give it, and most SharePoint estates have accumulated sharing decisions nobody has reviewed in years. Review the SharePoint sharing model, Purview classification and Entra ID Conditional Access policies before any AI workload goes live.

How Microsoft AI Foundry fits the AI Capability Ladder

Lanmark uses the AI Capability Ladder to keep the AI conversation in business language. Foundry sits at rungs two to four:

• Knowledge AI: Foundry plus a grounded knowledge base.
• Agentic AI: Foundry plus orchestration, with people approving the steps that matter.
• Action-taking AI: Foundry integrated with your line-of-business systems, with audit and rollback designed in from the start.

Most SMBs are at rung one, productivity AI, which mostly means Copilot. The right first Foundry project is almost always a single grounded knowledge agent on a single high-value data set, kept deliberately narrow so that the build, the cost and the governance stay controllable.

How Lanmark approaches Microsoft AI Foundry

Our engagement model is deliberately small at the start:

1. AI Readiness Review. A focused conversation that places your business on the ladder and establishes whether Foundry is the right next step or Copilot adoption alone is.
2. Fixed-scope pilot. Four to eight weeks on a single use case, with a written success measure agreed before we start.
3. Rolling adoption. If the pilot lands, add use cases at the rate the business can absorb them.

Four to eight weeks is realistic for a single grounded knowledge agent, assuming the data is reasonably clean and the use case is narrow.

If you are weighing up whether Foundry is right for you, see our AI consultancy work and generative AI consultancy page, then get in touch or call 020 7123 4910.