Multi-Model Is the New
Default Architecture

Microsoft is evaluating a self-hosted version of DeepSeek V4 in Copilot Cowork. Not as an experiment — as the next logical step in a multi-model strategy that has long been industry standard. For mid-sized businesses, that means: whoever relies on just one model is relying on the wrong one.

According to Axios (06/16/2026), Microsoft is evaluating the Chinese open-source model DeepSeek V4 as a cheaper option alongside the existing Anthropic and OpenAI models in Copilot Cowork — the agentic Microsoft 365 layer with usage-based billing that launched worldwide on June 16. The goal: companies choose the right model per use case, not per annual contract. Charles Lamanna, EVP Copilot, explains the move by pointing out that flat-rate pricing isn't sustainable — heavy users trigger hundreds of tasks per week and drive up infrastructure costs disproportionately. A decision on final integration is expected in the coming weeks.

The message behind it: Microsoft decides per use case which model runs. You should too.

What multi-model means for mid-sized businesses ▲

Anyone who thinks multi-model is an enterprise luxury is wrong. The a16z CIO Survey Q1 2026 shows: 81 percent of the Global 2000 CIOs surveyed by a16z (3rd annual a16z CIO Survey, 01/30/2026, n=100 VP/C-level) already use three or more model families in testing or production — a year ago it was 68 percent. Gartner puts the average at 4.2 models in production. And the trend is clear: no serious company still bets everything on a single foundation model.

Market context: The AI market shares in May 2026 show why multi-model is rational: ChatGPT holds 46.4%, Gemini 27.7%, Claude 10.3% — for the first time since the AI boom, no single provider controls the majority. Betting on one provider means betting on a concentration risk that's structurally avoidable.

What does that mean for a mid-sized business with 15 to 200 employees? Every use case has its own requirements profile:

• Phone agent (Ben) needs low latency and strong speech processing. Response time determines whether a call succeeds — here speed matters more than maximum reasoning depth.
• Email agent (Eva) processes sensitive customer data and needs a model with strong analytical ability and GDPR-compliant data processing. Context fidelity across long correspondence threads is a must.
• Dunning agent (Tom) works with structured financial data and needs reliability with numbers and legal wording. Hallucinations here wouldn't just be embarrassing — they'd be expensive.

A single model for all three? That's like having your best salesperson also do the bookkeeping — just because they happen to be there.

At Amazing Agents, we build multi-model-capable agents. You choose the model per use case, not per contract.

Sovereignty begins before model choice ▲

DeepSeek V4-Pro comes with impressive numbers: 1.6 trillion parameters, an MIT license, and a 1-million-token context window. Microsoft plans full self-hosting in EU Azure regions — with its own compliance controls and bias safeguards.

But self-hosting in an EU cloud region doesn't automatically mean "GDPR-compliant" or "EU-sovereign." The reality is more complicated:

• EDPB Opinion 28/2024 makes clear: LLMs generally cannot be classified as "anonymous." All GDPR obligations — legal basis, data protection impact assessment, information duties — still apply.
• BSI C3A criteria for cloud sovereignty aren't automatically met by a US cloud. The CLOUD Act still applies.
• Bias travels with the weights. Studies on DeepSeek R1 document 91.2 percent pro-China bias on geopolitical questions. No comparable peer-reviewed study exists yet for V4. Microsoft says safeguards have been added — external verification is still pending.

Important context: the DeepSeek consumer app has been investigated, and in some cases blocked, by several German data protection authorities. Self-hosting the open-weight models in Azure isn't affected by that — but the debate shows how sensitively the topic is perceived in DACH.

Even so, the right order for mid-sized businesses is clear:

1. Classify your data — which categories do you process? Art. 9 GDPR (health, biometrics), professional secrets under Section 203 of the German Criminal Code (StGB)?
2. Assign the model — sensitive data belongs in a European frontier model or a self-hosted setup; general office tasks can go to a cheaper model.
3. Document compliance — data protection impact assessment, data processing agreement, technical and organizational measures.

We recommend this order to every client. Not because we're lawyers — we're not — but because choosing a model without data classification is flying blind.

What you can do today ▲

Multi-model isn't a vision of the future. It's the architecture that Microsoft, 81 percent of the Global 2000, and mid-sized businesses are already building — or should be.

The decisive audit question: which data is allowed to reach which model?

If you can't answer that question in under 30 seconds, it's time for a conversation.

Frequently asked questions about multi-model strategy ▲

What is Microsoft Copilot Cowork?

Copilot Cowork is Microsoft's agentic M365 layer, launched worldwide on June 16, 2026. It enables usage-based billing per task — without a classic flat-rate license — and integrates multiple AI models depending on the use case.

Why is Microsoft evaluating DeepSeek V4 for Copilot?

Charles Lamanna (EVP Copilot) explains the move: flat-rate pricing isn't sustainable — heavy users trigger hundreds of tasks per week and drive up infrastructure costs disproportionately. DeepSeek V4 is meant to serve as a cheaper alternative for suitable use cases.

What does multi-model mean for SMBs with 5-20 employees?

Every agent gets the model that fits its requirements: phone agents need speed, email agents need context fidelity, dunning agents need reliability with numbers. Anyone who doesn't think about these separately is optimizing for the average — and gets none of the use cases quite right. 81 percent of the Global 2000 CIOs surveyed by a16z (3rd annual CIO Survey, 01/30/2026, n=100 VP/C-level) already use three or more model families — mid-sized businesses follow this architecture with a 12-18 month lag. The market context shows why that's rational: no provider holds the majority anymore.

Is DeepSeek V4 in Azure automatically GDPR-compliant?

No. EDPB Opinion 28/2024 makes clear: LLMs generally cannot be classified as "anonymous." A legal basis, a DPIA, and information duties still apply. The CLOUD Act continues to apply — even to Azure EU regions.

Does DeepSeek R1 have a bias problem?

Studies document 91.2 percent pro-China bias on geopolitical questions for DeepSeek R1. There's no peer-reviewed study yet for V4. Microsoft says safeguards have been added — external verification is still pending.

How many models do global companies use on average?

According to Gartner, an average of 4.2 models in production. 81 percent of Global 2000 companies already use three or more model families (a16z CIO Survey, Q1 2026).

What steps come before choosing a model?

First classify your data (Art. 9 GDPR, Section 203 StGB), then assign the model, then document compliance (DPIA, data processing agreement, technical and organizational measures). Without these three steps, any model decision is flying blind.