GBCORP BLOG · AI
AI App Development: An Ordinary App With One Extraordinary Step
Most AI apps are ordinary apps with one model-driven step that does real work. The step is the hard part, and the app around it is the part GBCORP already has.
AI app development produces an app in which a model does a job a person would otherwise do: answers from the business’s own content, reads a document into a form, sorts and drafts, finds by meaning, generates a summary. GBCORP builds the app on the shared engine that powers its white-label range, so sign-in, payments, notifications, consoles and store compliance exist on day one, and the work goes into the model step: scoped, evaluated, reviewed and priced as a release.
What the model step looks like in an app
| Step | What it does | Where |
|---|---|---|
| Assistant | Answers from the business’s content, cites the source, hands over to a person | Support, sales, member apps |
| Capture | Photographs a receipt, an ID or a form and fills the record | Finance, onboarding, field work |
| Sort and draft | Classifies messages and drafts the reply for approval | Service desks, inboxes |
| Search | Finds products or documents by meaning | Catalogues, portals |
| Generate | Descriptions, summaries, reports from data | Retail, reporting |
Behind an interface, with a review screen
The app calls a service GBCORP builds; the service calls the model. A better or cheaper model is a change to the service, not the app. At first a person confirms outputs on a review screen; those confirmations become the evaluation set, and accuracy per field is measured weekly against it. Review is removed where errors are harmless and kept where they are not.
Store rules for AI features
App stores have rules for generated content: moderation, reporting, age ratings and clear labelling. GBCORP’s engine carries user-generated-content moderation, reporting and blocking already, and AI features are built to the same rules, so a submission is not rejected for the feature that was the point of the app.
Cost per request, visible
Model usage is metered and reported weekly beside accuracy. Prompts are trimmed and cached where the model allows, smaller models are used where the evaluation set shows they suffice, and a step whose cost rises with volume is redesigned before it becomes a surprise. The business sees the cost of the step and its value on one page.
Checklist before building an AI app
- The one job the model step does, in a sentence.
- Fifty to two hundred real examples with their correct answers.
- Who reviews outputs at first.
- What the app does when the model is wrong or unavailable.
- Fields that must be masked before leaving the project.
- The number the app should move.
The numbers to expect
Scoping in one week. A first release with the model step and its review screen in four to eight weeks at a fixed price. Accuracy measured weekly. Model usage at cost; support at a known monthly figure. Customised or more complex functionality is quoted separately before any work starts.
A worked example
A membership organisation wants an app in which members ask questions about benefits, events and rules and get a correct answer at once. On the shared engine, the members’ app with sign-in, notifications and the console is configured in days; the assistant step is built behind an interface, indexed on the organisation’s own documents, and shown with a review screen to the staff for four weeks. Answers cite the document they came from and hand over to a person when unsure. By week eight the review is removed for benefits and events and kept for rules. Support tickets fall, and the console shows which questions have no good document yet.
When the model is wrong or away
An AI app is designed for the model’s failure. Every output carries a confidence and a source; low confidence goes to a person. A model outage leaves the app usable: the assistant says it cannot answer now and offers the contact route; extraction falls back to manual entry. Nothing in the business stops because a provider had a bad hour, and the log shows what happened.
Privacy inside the app
Users are told what the AI feature does and what it sends, in the app, before first use. Sensitive fields are masked before a request leaves the project. Conversations are stored in the client’s own project, retained for the period the client sets, and deletable by the user. Privacy labels in the stores describe the feature accurately, which is also what passes review.
Measuring the app, not the demo
Accuracy per field against the evaluation set. Review rate and how it falls. Cost per request. Questions the assistant could not answer, which is the list of documents to write. Time saved per task, measured before and after. These sit in the console from the first release, and the quarterly review reads them.
From one step to the next
The first model step is the easy one. The second reuses the service, the evaluation habit and the review screen on a harder process. An app that started with an assistant adds capture, then sorting, each as a priced release, each measured on its own page.
Ownership of the app and the data
The app, its service, prompts, evaluation set and logs live in the client’s own cloud project. Store accounts and the model provider account are in the client’s name. On custom work, the code is delivered. No credentials are retained after a hand-over unless a support contract requires them.
Questions we hear first
Can you add AI to our existing app?
Yes. The step is added as a release, with the same review and measurement, whether or not GBCORP built the app.
Which model do you use?
The one that scores best on your evaluation set at the cost you accept; it can be swapped later.
Does it work in several languages?
Yes. The step works in the user’s language and the app is multilingual with right-to-left layouts as a setting.
Who owns the app and the data?
The business: code on custom work, cloud project, model provider account and store accounts in its name.
Where to start
The GBCORP AI Development and Automation product is the priced starting point; the GBCORP AI development and machine learning hub sets out the method. For the technical view, read AI application development explained.
GBCORP operates across 17 countries on 4 continents through six pillars: AI & IT, Innovation, Sustainability, Infrastructure, Premium Business, and EV's & Solutions.
