Contents7
“It should have AI in it.” The request comes up in many app projects, often before anyone knows where to put it. Well placed, AI removes a tedious entry or a sorting job nobody likes. Badly placed, it adds a cost to every click, a wrong answer delivered with confidence, and one more legal obligation. The right question is not “do we need AI?” but “which of the user’s actions gets shorter?”.
In short
AI helps in an app when it reads, sorts, summarises or translates what the user gives it.
What must be exact (an amount, a date, an access right) is still computed by the server, with rules.
A person approves what the AI suggests as soon as the result is published or commits someone.
The model is chosen per task: in the cloud, or directly on the phone for some simple tasks.
Every call is paid for, and a conversational assistant must identify itself as such since 2 August 2026.
Where AI really helps
The uses that hold up in production have one thing in common: the AI transforms material the user provides, and the result can be checked at a glance.
- Reading a document or a photo. A receipt, a delivery note, a business card, a rating plate: the AI extracts the useful fields, and the user confirms instead of typing.
- Sorting what comes in. A customer request classified by topic and urgency, routed to the right person.
- Summarising. The thread of a long file, a client’s history before a meeting, a transcribed meeting.
- Drafting. A standard reply, a description, a report, which the person reviews before sending.
- Translating. A listing, a message, an interface, for an audience that does not speak the team’s language.
- Searching your documents. Answering a question by citing the passage of the internal document it rests on, as our article on RAG for business documents describes.
From the field
A photographed receipt becomes a bookkeeping entry
Our bookkeeping entry agent, in internal use on Telegram, receives the photo of a receipt and extracts the supplier, amount, VAT, date and category. Duplicates, amount consistency and a confidence threshold are checked before anything is written, and the only human step is approving the entry with one button. It is described on our products page.
Where it has no business
A language model produces the most likely continuation of a text. That is a strength for reading and writing, a weakness for anything that must be exact every time.
- Calculations and rules. A total, a due date, an access right are computed by code, on the server. AI can read the amount on a receipt; it does not decide which VAT rate applies.
- Decisions that commit someone. Rejecting an application, scoring a candidate, setting a price: the decision stays human, and the law regulates automated decisions, as our article on automated decisions explains.
- Anything sent without review. A text published or sent to a client in the company’s name goes through a person.
Watch out
A client learned this the hard way
On a client’s website, a model tasked with writing listings invented three property developments that did not exist, and they were published. The pages were taken down and reported to search engines as removed. Since then, every listing written by AI waits for human approval before going live.
At Palmora Property, a real estate agency in Phuket that we work with, this is the rule: new listings are written in the house format, then wait for human approval, as described in the Palmora case study.
In the cloud or on the phone
Two families of models are available to an app, and they complement each other.
A cloud model, through an API
The most capable models, from several providers (OpenAI, Anthropic, Google, Mistral, or hosted open models)
Every call is paid for, and the data leaves the phone
The provider can change without touching the app
A model on the phone
No cost per call, and it works without a network
The data stays on the device
A smaller model, for simple tasks, and missing from older phones
Both major mobile platforms now open their on-device model to apps. Apple presents its Foundation Models framework as access to the on-device language model behind Apple Intelligence, with features available offline and inference free of cost for developers, according to its September 2025 announcement. On Android, Gemini Nano runs on the device to summarise, proofread, rewrite or describe an image, but may be unavailable on the user’s phone, warns the Android documentation, which then recommends falling back to a cloud model.
For a business app, the combination is often right: the phone for small tasks and offline use, the cloud for what needs a capable model. If the data is sensitive, there are also models you run on your own servers, described in the four levels of private AI.
What each call costs
An AI feature in the cloud has not only a development cost: it has a cost every time it is used, which grows with the app’s success. Three habits keep it in check.
- The right model per task.
Sorting a request does not need the most expensive model. Choose the model task by task, and keep that choice changeable.
- Never pay twice.
A text already translated or summarised is cached and does not go back to the model.
- A cap and an alert.
A counter per user and per day stops a loop or abuse from draining the budget.
This is what we apply at Palmora Property: listings are translated into six languages by a model chosen for its cost, about six times cheaper than the standard model, and a cache avoids paying twice for the same translation. The full calculation is in the cost of AI in business.
One security rule comes on top: the model’s access key never lives in the app installed on the phone, where it could be read. The call goes through the server, which checks who is asking for what, as app security details.
The rules to know
Since 2 August 2026, Article 50 of the AI Act applies: an AI system that interacts with people must tell them they are dealing with an AI, unless it is obvious, according to the Commission’s AI Act Service Desk. Data sent to the model remains subject to the GDPR: where does it go, and is it reused by the provider? The full picture is in the AI Act and GDPR for a small business. An app open to minors calls for extra care, which we discuss in AI companions and teenagers.
Where to start
Start from your users’ most tedious action, not from the fashionable model. If it is data entry from a document, AI reading with human confirmation saves time from the first version. Then measure the correction rate: it tells you whether the feature helps or needs rework.
The feature goes into the requirements document, is tested on a clickable prototype, and weighs on the price of the app as well as its maintenance budget. Tools that generate a whole app raise other questions, covered in vibe coding and low-code or code. The rest of the journey (the back end with Supabase, Firebase or your own server, PWA or native app, offline, field teams, client portal, custom or off-the-shelf, notifications, subscriptions, developer account, taking over an abandoned app) is gathered in the section. And schools have their own debate: ban AI or teach it.
Our app development prices each AI feature with its monthly usage cost, from the scoping stage.
Frequently asked questions
Which AI features are most useful in a business app?
Those that shorten a repetitive action: reading a document or a photo to pre-fill a form, sorting incoming requests, summarising a file, translating, or finding an answer in your internal documents. In each case, the user sees the result and can correct it.
How much does an AI feature in an app cost?
There is the development, and a usage cost on every call when the model is in the cloud. That cost depends on the model chosen, the length of the texts and the number of users. Choosing the model per task, caching results and setting a cap per user make it predictable.
Can AI be used without sending data to the cloud?
Yes, in two ways: a model on the phone, like those Apple and Google open to apps, for simple tasks; or a model hosted on your own servers. The first works offline but stays limited, the second needs infrastructure to run.
Should users be told they are talking to an AI?
Yes. Since 2 August 2026, Article 50 of the AI Act requires a system that interacts with people to tell them it is an AI, unless that is obvious from the context. A clear label in the interface is enough, and it also avoids misunderstandings with your users.
Sources
- Apple Newsroom, Foundation Models framework, September 2025 (new tab)apple.com
- Android documentation, Gemini Nano (new tab)developer.android.com
- European Commission AI Act Service Desk, Article 50 (new tab)ai-act-service-desk.ec.europa.eu
European Commission AI Act Service Desk, Article 50
Read on ai-act-service-desk.ec.europa.eu(new tab)Jim · founder of Brain Plus AI
Also in web and mobile apps
Stepping back22 min
Vibe coding: AI codes fast, the product stays human
AI has removed real technical barriers, but it has not removed what makes a product. What is true and what is false in "anyone can do anything with computers", with examples in both directions.
Understand7 min
App security: accounts, sign-in and data
The most common flaw in an app is not a guessed password, it is a forgotten access rule that lets one user read another’s data. Here is what really protects your users’ accounts and data.
Choose and price6 min
How much an app costs: what makes the price vary
A price range says nothing about your project: what makes an app quote vary is a handful of specific decisions, plus a monthly running cost to be priced separately. Here they are.