Mobile AI Features That Never Send User Data to a Server: What Is Possible on iOS and Android in 2026

Your CISO wants AI that never sends user data to a server. The product team wants AI that actually works. In 2026, these requirements are no longer in conflict for most enterprise mobile use cases. Here is exactly what on-device AI can do, on what hardware, at what accuracy level — based on production systems, not benchmarks.

What on-device means in 2026

"On-device" means the model weights are stored locally on the device. All inference — the process of running user input through the model to produce output — happens on the device processor. No API call. No network request. No data leaving the device during AI use.

This was a theoretical statement for most enterprise use cases three years ago. In 2026, it is a practical one. The combination of capable open-source models (Llama 3, Phi-4, Gemma 2, Mistral 7B, Whisper), efficient inference frameworks (llama.cpp, Core ML, QNN, MNN), and hardware that has caught up with the model requirements (A15+ NPU on iOS, Snapdragon 8 Gen 1+ on Android) means that production-quality AI runs locally on current devices.

Wednesday built Off Grid to prove this with production users, not benchmarks. 50,000+ users run text AI, voice transcription, image generation, and vision features in Off Grid with no cloud inference calls. The capabilities described in this article are what those users experience daily.

Text AI on-device

Text AI covers the largest category of enterprise mobile AI use cases. On-device 7B parameter models handle:

Documentation assistance. A field technician describes a repair verbally and in short notes. On-device AI structures those notes into a formatted work order. A clinician speaks rough observations and the AI organises them into a structured clinical note format. Accuracy on enterprise text structuring tasks: 88-94%.

Summarisation. A sales rep reviews 20 customer interactions before a quarterly review call. On-device AI summarises each interaction into two sentences. A manager needs a summary of the week's field reports. On-device AI processes the reports and produces a summary. Context window limits apply — 4,000-8,000 tokens per call covers most individual document summarisation tasks.

Classification. Support tickets classified by urgency and category. Customer feedback classified by sentiment and topic. Work orders classified by job type and priority. Classification tasks are among the strongest on-device AI use cases — accuracy of 90-95% is achievable with appropriate model selection.

Named entity extraction. Extracting key information from unstructured text: customer names from calls, part numbers from field notes, medication names from clinical text, dates and amounts from financial documents. On-device models handle this well for common entity types.

Short-form drafting. Generating first drafts of standard emails, reports, or notifications from structured inputs. Response length and quality are appropriate for enterprise internal communication; on-device models are not suitable for polished long-form content generation.

Conversational Q&A over provided context. A technician asks "what is the maintenance interval for this equipment?" and the app retrieves the relevant manual section and passes it to the on-device model as context. The model answers the question using only the provided context, not general knowledge. This is the on-device RAG pattern and it works well for knowledge bases under 500MB.

Voice transcription on-device

On-device Whisper is the standard for enterprise voice transcription features that must not send audio to a server.

Whisper achieves 95%+ word accuracy on clear English speech at 1.5-3x real-time processing speed on current flagship hardware. A 5-minute dictation is transcribed in 100-200 seconds.

For enterprise use cases:

• Field service documentation: strong accuracy in moderate noise environments. Background machinery noise reduces accuracy; extreme industrial noise environments require testing with representative audio.
• Clinical documentation: strong accuracy on standard medical terminology with the medium or large model variant. Very specialised terminology (rare surgical procedures, uncommon drug names) may benefit from a domain-adapted Whisper fine-tune.
• Sales and customer interaction: strong accuracy on phone-quality audio, clear English, and standard business vocabulary.
• Legal dictation: strong accuracy on standard legal vocabulary. Unusual case citations or rare jurisdictional terminology may require domain adaptation.

Wednesday shipped on-device Whisper in Off Grid. The implementation handles variable noise environments, devices without NPU acceleration (using CPU inference at slower speed), and language variation. It is a production implementation across 50,000+ users, not a demonstration.

Image and vision AI on-device

Image AI runs on-device using hardware-accelerated backends: Core ML on iOS (Metal GPU), QNN on Snapdragon 8 Gen 1+ (NPU), and MNN on ARM64 Android (CPU, with SIMD optimisation).

Image generation. Diffusion models (LCM, SDXL Turbo) generate images in 10-30 seconds on current flagship hardware. Quality is suitable for product visualisation, simple illustrations, and content creation. Not suitable for photorealistic generation or complex compositions requiring very high resolution. Wednesday shipped three production image generation backends in Off Grid — the only known production implementation across all three hardware backends from a single React Native app.

Image classification. Classifying images into predefined categories — equipment condition ratings, product quality tiers, damage assessments, plant species identification — runs on MobileNet, EfficientNet, and similar lightweight architectures with 94-98% accuracy on well-defined classification tasks. This is among the most mature on-device AI capability.

Object detection. Identifying and locating specific objects in images — parts, products, text, barcodes — runs on YOLOv8 and similar architectures at real-time speed on current devices. Enterprise use cases: quality inspection, inventory management, equipment identification in the field.

Vision-language models. Models that can answer questions about an image — "what is wrong with this equipment?" or "what does this document say?" — now run on-device with 7B-class vision-language models. Accuracy is lower than GPT-4o vision for complex scenes but suitable for structured enterprise visual inspection tasks. Wednesday shipped vision-language features in Off Grid.

Face detection (local only). Detecting whether a face is present in an image, without identification or matching against any database, runs on-device without any privacy concern. This is the only face-related AI that should be on-device in enterprise apps — face recognition against external databases requires careful compliance review regardless of where inference runs.

Document analysis on-device

Document analysis covers features that extract, understand, or answer questions about documents loaded into the app.

Document Q&A. Load a PDF, specification, policy document, or manual. Ask questions about it. On-device embedding converts the document to a local vector index; on-device inference answers questions using retrieved passages as context. Works for documents up to approximately 500MB. Wednesday shipped document Q&A in Off Grid using this pattern.

Form and table extraction. Extracting structured data from scanned forms, tables, and documents. Combination of on-device OCR and on-device text processing. Accuracy: 88-93% on well-structured forms; lower on handwritten or degraded originals.

Document classification. Identifying the type, category, or routing of a document without reading its full content. Suitable for enterprise document management features where documents need to be sorted or routed automatically.

Translation. On-device translation for common language pairs (English-Spanish, English-French, English-Portuguese, English-Mandarin, and others in the top 20 language pairs) using NLLB and similar models. Quality is production-grade for standard business text. Less common language pairs have lower accuracy.

What on-device cannot do in 2026

Honesty about the limitations matters. These are the enterprise use cases where on-device AI is not the right answer today.

Real-time knowledge. On-device models have a training cutoff. They do not know about events, regulatory changes, or product updates that occurred after the training data was collected. A customer service AI that needs to answer questions about current product pricing or a compliance assistant that must reflect this week's regulation cannot be purely on-device.

Very long document analysis. Processing a 200-page contract or a full year of financial statements in a single inference call requires a 100,000+ token context window. On-device 7B models typically handle 4,000-8,000 tokens. For very large document tasks, cloud APIs with 128,000+ token context windows are required.

Complex multi-step reasoning. Tasks that require many steps of logical inference — "given these 10 constraints, determine the optimal allocation" — where current on-device 7B models lag GPT-4o class cloud models significantly. Most enterprise mobile tasks do not require this level of reasoning, but complex analytical tasks do.

High-accuracy multi-speaker real-time transcription. Identifying who said what in a live multi-person conversation, in real time, with high accuracy, is not reliably achievable on-device in 2026. Cloud APIs with speaker diarisation and streaming transcription remain ahead for this specific capability.

Full capability table

Capability	On-device feasible	Accuracy	Best for	Not suitable for
Text structuring and documentation	Yes	88-94%	Field notes, clinical documentation	Very long documents
Text summarisation	Yes (under 6,000 words)	Strong	Individual documents, reports	Book-length analysis
Text classification	Yes	90-95%	Ticket routing, sentiment	Ambiguous, multi-class edge cases
Named entity extraction	Yes	88-93%	Common entities (names, dates, numbers)	Highly specialised terminology
Voice transcription	Yes	95%+	English, clear speech	Extreme noise, rare vocabulary
Image classification	Yes	94-98%	Defined category sets	Open-world classification
Object detection	Yes	Strong	Known object categories	Unconstrained open-world
Image generation	Yes (flagship, 10-30s)	Suitable for illustration	Product viz, simple content	Photorealistic, high-res
Vision-language Q&A	Yes	Moderate	Structured inspection	Complex scene description
Document Q&A	Yes (under 500MB)	Strong	Manuals, policies, specifications	Very large knowledge bases
Translation	Yes (top 20 pairs)	Production-grade	Standard business text	Rare language pairs
Real-time knowledge	No	—	—	News, live data, current events
100K+ token context	No	—	—	Very long document analysis
Multi-speaker real-time transcription	Not reliably	—	—	Live meeting captioning

How Wednesday has shipped all of this in production

Off Grid is not a proof of concept. It is a production product with 50,000+ users, 1,700+ GitHub stars, and publicly auditable architecture.

Text AI: llama.cpp on CPU with Core ML and QNN acceleration where available. Multi-turn conversation, document Q&A, and text structuring all running locally.

Voice AI: on-device Whisper across the full device compatibility matrix. Handles variable noise environments and non-NPU devices via CPU inference fallback.

Image AI: three backend implementations — MNN for ARM64 Android, QNN with NPU for Snapdragon 8 Gen 1+, Core ML for iOS. All three running production image generation for 50,000+ users.

Vision AI: vision-language models for image understanding and Q&A, running on-device with the same backend infrastructure as image generation.

This is the reference implementation for enterprise teams whose CISO needs to understand what on-device AI is capable of, what it requires to ship, and how it handles the device matrix. It is not a vendor's capability claim. It is a public product with verifiable user numbers.