Apple Intelligence is already running on your team’s Macs and iPhones — a 3-billion parameter on-device model that handles writing assistance, summarisation, and system-level tasks without an internet connection. That sounds like sovereign AI. It isn’t. Apple Intelligence uses a hybrid architecture: straightforward tasks run on-device, but more demanding requests route to Apple’s Private Cloud Compute (PCC) servers. Your data leaves the device. Not to OpenAI — to Apple’s cloud infrastructure. For personal productivity tasks, that’s a reasonable trade-off. For marketing agencies processing client data, it’s not a compliance solution. Open-weight local models run via Ollama give you genuine data sovereignty. Apple Intelligence does not.

What Apple Intelligence Actually Is#

Apple Intelligence is not a single model. It’s a tiered system with three layers, each with different data handling:

• AFM-on-device: Apple’s Foundation Model running natively on M-series Macs and A18 Pro iPhones. Approximately 3 billion parameters, using mixed 2-bit and 4-bit quantization to fit within 8–12 GB of unified memory. Handles writing tools, summaries, notification prioritisation, and basic generative tasks. Genuinely on-device — no data transmitted.
• Private Cloud Compute (PCC): Apple’s proprietary cloud inference cluster for tasks too complex for the on-device model. More capable responses, longer context handling, and advanced reasoning go here. Data leaves the device and is processed on Apple’s servers — though Apple states PCC requests are not retained or logged.
• ChatGPT integration: For certain requests, Apple Intelligence routes to OpenAI’s GPT-4o. This is opt-in, and Apple prompts users before the first transfer — but it exists, and it means third-party cloud inference is part of the Apple Intelligence architecture.

The on-device model is genuinely impressive for its size. The compliance problem for agencies is that you cannot determine, per query, which layer Apple Intelligence will use. Complex prompts, long documents, and requests requiring extended reasoning route to PCC automatically. There’s no user-visible indicator that distinguishes a fully local response from a PCC-assisted one.

The Compliance Problem: “Mostly Local” Isn’t Sovereign#

For personal tasks — rewriting an email, summarising your calendar, generating a headline — the Apple Intelligence hybrid model is excellent. The privacy trade-off is reasonable and Apple’s PCC architecture is genuinely more privacy-preserving than standard cloud AI.

For marketing agencies processing client data, the standard is different. GDPR Article 28 requires a signed Data Processing Agreement with every third party that processes personal data on your behalf. Apple does offer privacy documentation and its PCC architecture is designed to minimise data exposure — but the legal question remains: does processing client personal data through Apple Intelligence, where some requests route to PCC, require a formal DPA and transfer impact assessment?

The answer depends on your jurisdiction, your clients’ data, and your DPA documentation — and most agencies have never considered the question. That ambiguity is itself a compliance risk. An open-weight local model running via Ollama eliminates the ambiguity entirely: the inference happens on your hardware, no data leaves your network, no third-party processor is involved.

You may be interested in: Sovereign AI for Marketing Agencies: Keep Client Data Inside Your Building

What Apple Intelligence Is Good For (In an Agency Context)#

This isn’t an argument against using Apple Intelligence. It’s an argument for using it on the right tasks.

Apple Intelligence excels at personal productivity work where the data involved is your own — not your clients’:

• Rewriting and proofreading your own emails and documents
• Summarising meeting notes from your internal team
• Generating image concepts for internal brainstorming
• Priority inbox management for your personal email
• Siri shortcuts and system-level automation on Mac and iPhone

The 2026 Siri roadmap — “Siri LLM” for natural dialogue and “Siri Chatbot” for agentic tasks — will expand what Apple Intelligence can do at the system level, potentially including a $15/month Apple Intelligence Pro subscription for advanced features, according to TokenRing’s roadmap analysis. These are genuine productivity gains for individuals.

None of them solve the agency’s need to query client data privately. That’s a different problem requiring a different tool.

What Open-Weight Local Models Give You That Apple Intelligence Cannot#

Running Qwen2.5-32B or Llama 3.3-70B via Ollama on a Mac Mini M4 Pro gives an agency four capabilities Apple Intelligence structurally cannot provide:

Guaranteed on-device inference, always. There is no PCC routing. There is no ChatGPT fallback. Every query — regardless of complexity or context length — is answered by the model running on your hardware. Not “usually local.” Always local. That guarantee is what compliance requires.

Client data isolation by account. You can run separate Ollama instances or separate RAG pipelines for each client. Client A’s data never mingles with Client B’s inference context, architecturally. Apple Intelligence has no equivalent per-client isolation mechanism.

Queryable first-party data. Pairing Ollama with a RAG pipeline over your BigQuery exports, WooCommerce records, or GA4 data turns your local model into an analytics assistant for client data specifically. Apple Intelligence cannot query your structured marketing databases — it handles documents and system tasks, not structured data analysis.

Model choice and fine-tuning. Open-weight models can be swapped, updated, or fine-tuned on your agency’s own data. Apple Intelligence runs the model Apple chose, at the capability level Apple configured. You have no visibility into model versions, training data, or update schedules.

You may be interested in: How to Query First-Party Marketing Data with a Local LLM Without Cloud Risk

The Practical Deployment Decision#

The answer for most marketing agencies isn’t Apple Intelligence or open-weight models. It’s both, for different purposes.

Use Case	Apple Intelligence	Open-Weight Local (Ollama)
Personal email rewriting	✅ Excellent	Overkill
Client data analysis	❌ Not suitable	✅ Correct tool
Campaign attribution queries	❌ Cannot query structured data	✅ Via RAG pipeline
Internal meeting summaries	✅ Good	Works but unnecessary
Client brief drafting	⚠️ PCC routing risk	✅ Fully local
GDPR-safe client data processing	❌ Ambiguous	✅ Clear
Per-client data isolation	❌ Not available	✅ Architectural

Use Apple Intelligence for the personal productivity layer it was designed for. Deploy Ollama with Qwen2.5-32B on a Mac Mini M4 Pro for anything involving client data, structured analytics, or tasks where compliance certainty matters. The two tools are not in competition — they operate at different layers of an agency’s AI stack.

Where Transmute Engine Fits#

Neither Apple Intelligence nor an open-weight local model can produce good analytics from incomplete data. If your WooCommerce event tracking is browser-side — subject to ad blockers, Safari ITP cookie restrictions, and checkout page timing failures — the data your local model reasons over has structural gaps. Confident answers from incomplete records are worse than no answers, because they direct decisions in the wrong direction.

The Transmute Engine™ captures WooCommerce events server-side from PHP hooks — bypassing ad blockers entirely, preserving the full attribution chain, and routing clean first-party data to BigQuery. When that BigQuery data feeds your local Ollama RAG pipeline, you get complete records analysed by private inference. Apple Intelligence cannot query it. A properly deployed open-weight local model can. The infrastructure decision and the model decision are connected.

Key Takeaways#

• Apple Intelligence uses a 3-billion parameter on-device model (AFM) for simple tasks, but routes complex requests to Private Cloud Compute — it is not fully local and is not a sovereign AI solution
• For personal productivity tasks using your own data, Apple Intelligence is excellent; for client data in a marketing agency, the PCC routing creates compliance ambiguity that open-weight local models eliminate
• Open-weight models via Ollama (Qwen2.5-32B, Llama 3.3-70B) guarantee fully local inference on every query, support per-client data isolation, and can query structured marketing databases via RAG
• The practical answer for most agencies is both: Apple Intelligence for the personal productivity layer, Ollama for client data and analytics
• Complete data is the prerequisite for useful local AI analytics — server-side event collection via Transmute Engine ensures the model reasons from complete WooCommerce records

Apple Intelligence is a well-designed productivity tool for the individual. It was not designed for a marketing agency’s compliance requirements. Knowing the difference — and deploying accordingly — is the decision that separates agencies running AI responsibly from agencies running AI conveniently.