This piece was written for enterprise technology leaders and originally published on the Wednesday Solutions mobile development blog. Wednesday is a mobile development staffing agency that helps US mid-market enterprises ship reliable iOS, Android, and cross-platform apps — with AI-augmented workflows built in.
A plain-language capability table covering what on-device AI can do offline, what still requires a server, and what each capability costs to add to an enterprise app.
Forty-three percent of enterprise mobile users work in locations with unreliable or no connectivity at least once a week. That number comes from field operations, hospital networks, construction sites, and office basements where Wi-Fi drops and cellular is blocked. If your AI features stop working the moment the connection drops, you have not built AI into your app. You have built a dependency on someone else's server.
On-device AI runs entirely on the phone or tablet. No request leaves the device. No connection is required. The AI processes text, voice, images, and documents using the device processor and a model stored in local memory. This is not a workaround. It is how the most privacy-sensitive and connectivity-constrained enterprise use cases get solved.
This guide covers what on-device AI can do today, what it cannot, device requirements by capability, and realistic cost ranges for adding each capability to an existing enterprise app.
Key findings
Voice transcription, text generation, document scanning, and image classification all work fully offline on modern iOS and Android devices.
Real-time language translation, complex multi-step reasoning, and web search require a server connection and cannot be replicated on-device at equivalent quality today.
Adding a single on-device AI capability to an existing app typically costs between $40,000 and $90,000 depending on complexity and integration depth.
Wednesday's Off Grid project shipped all four core capabilities (text, voice, image, vision) on iOS and Android from a single app with 50,000+ users and zero cloud AI dependency.
Why offline AI matters now
Your board's mandate to "add AI" usually means one of two things. Either they want the app to get smarter and more useful, or they want to reduce manual work for users. Both goals are achievable without a cloud AI dependency, and in many cases the on-device path is faster to approve and faster to ship.
There are three drivers pushing enterprise teams toward on-device AI specifically.
First, compliance. Any data sent to a third-party AI service is data leaving your control. For healthcare, financial services, legal, and government applications, that triggers review processes that can add months to a launch timeline. On-device processing eliminates the data transfer entirely.
Second, reliability. Field operations, clinical settings, and manufacturing floors have poor connectivity. An AI feature that drops out when the signal drops is worse than no AI feature at all. It trains users to distrust it.
Third, cost. Cloud AI charges per request. An app with millions of active users running AI features can generate six-figure monthly API bills. On-device AI has zero marginal cost per inference once the model is on the device.
What on-device AI can do today
The table below covers the capabilities your team is most likely to ask about. "On-device" means the capability works with no network connection. "Device minimum" is the oldest device that produces acceptable results.
| Capability | On-device | Notes | Device minimum |
|---|---|---|---|
| Voice transcription | Yes | Whisper models, 3-5% word error rate in English | iPhone 12 / Android 2021 flagship |
| Document scanning and OCR | Yes | Extracts structured text from photos of documents | iPhone 11 / Android 2020 flagship |
| Image classification | Yes | Labels images from a fixed category list | iPhone 11 / Android 2020 flagship |
| Text summarization | Yes | Summarizes documents up to ~10,000 words | iPhone 13 / Android 2022 flagship |
| Form auto-fill from photo | Yes | Reads a form image and populates fields | iPhone 12 / Android 2021 flagship |
| Short text generation | Yes | Replies, descriptions, notes under 500 words | iPhone 13 / Android 2022 flagship |
| Language detection | Yes | Identifies the language of a text passage | iPhone 11 / Android 2020 flagship |
| On-device image generation | Yes (slow) | 30-90 seconds per image on 2023 devices | iPhone 14 Pro / Android 2023 flagship |
| Named entity extraction | Yes | Extracts names, dates, amounts from text | iPhone 12 / Android 2021 flagship |
| Sentiment classification | Yes | Positive/neutral/negative at sentence level | iPhone 11 / Android 2020 flagship |
Every capability in the table above runs with the device in airplane mode. No API key. No monthly bill. No data leaving the device.
What still requires a server
On-device AI has real limits. The table below is equally important.
| Capability | Requires server | Why |
|---|---|---|
| Web search and real-time data | Yes | The model has no knowledge of events after its training cutoff and no access to the internet |
| Complex multi-step reasoning | Often | 3B-7B parameter models produce noticeably weaker results on tasks requiring long chains of logic |
| Real-time language translation (50+ languages) | Often | High-quality translation for rare language pairs needs larger models that don't fit comfortably on-device |
| Large document analysis (100+ pages) | Often | Context window limits on smaller models affect quality on very long documents |
| High-resolution image generation | Yes | Generating high-quality images at 1024px and above takes minutes on-device vs seconds in the cloud |
| Custom model training or fine-tuning | Yes | Training always happens server-side; only inference runs on-device |
The practical rule: on-device AI is excellent for well-defined, focused tasks. When the task requires open-ended reasoning over large amounts of information, a server connection gives better results.
Device requirements by capability
Not every user will have a 2023 flagship. Your device minimum decision affects what percentage of your user base gets the full AI experience.
Voice transcription with Whisper small runs on iPhone 12 and equivalent Android. That covers roughly 85% of enterprise iOS users and 70% of enterprise Android users based on typical enterprise device refresh cycles.
Text generation with a 3B parameter model requires an iPhone 13 or equivalent. That is about 75% of enterprise iOS users today. The gap closes every year as devices age out of enterprise fleets.
Image generation is the most demanding. Producing a single image takes 30-90 seconds on an iPhone 14 Pro. On most Android devices it takes longer. This capability is worth adding only when your use case truly requires it.
For apps where your user base skews toward newer devices (financial services, healthcare with clinical staff devices, managed enterprise deployments), the device requirements are rarely a barrier. For consumer-facing apps, they matter more.
Cost to add each capability
These ranges assume an existing native iOS and Android app with a modern architecture. Greenfield apps or apps with significant technical debt cost more. Ranges cover design, engineering, testing, and release.
| Capability | Engineering cost range | Timeline |
|---|---|---|
| Voice transcription (Whisper) | $35,000 - $55,000 | 4-6 weeks |
| Document scanning and OCR | $30,000 - $50,000 | 4-6 weeks |
| Image classification (custom categories) | $40,000 - $70,000 | 5-8 weeks |
| Text summarization | $45,000 - $75,000 | 6-8 weeks |
| Short text generation | $55,000 - $90,000 | 7-10 weeks |
| Full on-device AI suite (text + voice + vision) | $150,000 - $250,000 | 14-20 weeks |
The full suite cost is not simply the sum of individual capabilities. Shared infrastructure (model loading, memory management, on-device storage) is built once and used across all features, which reduces the marginal cost of each additional capability.
The Off Grid reference point
Wednesday built Off Grid as an open-source proof of concept that these capabilities work at scale. Off Grid runs on iOS, Android, and macOS from a single app. It includes:
- Text generation via llama.cpp
- Image generation via MNN/QNN/Core ML
- Voice transcription via Whisper
- Vision (image understanding and description)
Zero cloud dependency. Zero ongoing API cost. The project has 50,000+ users and 1,700+ GitHub stars. It is not a demo. It is a working application that Wednesday's engineering team built and maintains.
When a client asks whether on-device AI is real or a marketing claim, Off Grid is the answer. The source code is public and the app is in the App Store.
See case studies at mobile.wednesday.is/work
How to pick what to build first
Start with the capability that solves a problem your users have today, not the most technically impressive feature on the list.
The fastest path to a shipped on-device AI feature is voice transcription. The infrastructure is well-understood, the Whisper model is proven, and the use case is clear in almost every enterprise context. A field technician filing a report by voice, a clinician logging a patient note, a sales rep capturing a meeting summary. One capability, clear value, four to six weeks to ship.
The second-fastest is document scanning with OCR. Most enterprise apps involve some form of paperwork. Scanning a document and extracting the data into a structured form removes a manual step that users dislike. The implementation is straightforward and the device requirements are low.
Text generation and summarization are worth adding once voice and document scanning are live. They require more careful design because the output is generative and needs guardrails. Budget an extra two weeks for prompt design and output validation.
The decision framework is simple: pick the capability that removes the most manual work for your users, confirm it works on the devices your users carry, and get it shipped before attempting the next one.
Want to go deeper? The full version — with related tools, case studies, and decision frameworks — lives at mobile.wednesday.is/writing/mobile-ai-without-internet-what-is-possible-cost-2026.
Top comments (0)