GBCORP BLOG · AI
AI Development Company: What GBCORP Builds and How It Measures It
An AI development company should be judged on one thing: whether the systems it builds do measurable work in production. GBCORP publishes the method by which it makes that measurable, so a business can hold it to it.
GBCORP is an AI development company in the working sense: it builds systems in which a model does a job inside a business process, and it proves the job is being done with a number measured every week. The systems run on the same engine as the group’s apps and platforms, which is why an AI step can be added to an existing app or built into a new one at a fixed price per release, with the data in the client’s own project and the model behind an interface that can be swapped.
What GBCORP builds
| System | What it does | Where it pays |
|---|---|---|
| Assistants | Answer from the business’s own content, cite the source, hand over to a person | Support, sales, internal knowledge |
| Extraction | Invoices, forms, IDs and receipts read into records | Finance, onboarding, logistics |
| Classification and routing | Messages, tickets and orders sorted and drafted | Service desks, inboxes |
| Search | Catalogues and archives searched by meaning | Retail, portals, records |
| Agents | Multi-step tasks under rules, with a person approving actions | Operations, back office |
| Automation | Model steps inside workflows, with review and measurement | Any repeated process |
The evaluation set runs the project
The scoping week collects fifty to two hundred real examples with their correct answers. Every candidate model and every prompt change is scored against them. The weekly report shows accuracy per field, cost per request and the review rate. The set grows from the review screen as people confirm outputs. Without it, an AI project is an opinion; with it, the business can see whether it is working and stop or expand it from evidence.
The review step
A person confirms outputs at first. Where errors are harmless, the review is removed once accuracy holds; where errors cost money or trust, it stays, and the person still saves most of the time. An agent that takes actions keeps an approval step for anything that moves money or sends a message outside the business. GBCORP does not remove a review to make a demonstration look better.
Data handling and the model behind the interface
Prompts and documents go to the model provider only for the request, under terms that exclude training. Sensitive fields are masked before they leave. Everything sent and received is logged in the client’s own cloud project. The app calls a service GBCORP builds, and the service calls the model, so a better or cheaper model is a change to the service and not to the app.
Checklist for judging an AI company
- Ask to see an evaluation set and the weekly accuracy of a live system.
- Ask where the review step sits and when it is removed.
- Ask where data goes and whether it is used for training.
- Ask for the cost per request on the same report as accuracy.
- Ask whether the model can be swapped without a rebuild.
- Ask which AI systems the company runs itself.
The numbers to expect
Scoping in one week. A first system live with its review step in four to eight weeks at a fixed price. Accuracy measured weekly. Model usage at cost; a support plan at a known monthly figure covering evaluation runs and model updates. Customised or more complex functionality is quoted separately before any work starts.
A worked example
A property management company with ten thousand tenant messages a month asks whether AI can help. The scoping week collects two hundred messages with the correct category and the reply a manager would send. The first release classifies each message, drafts a reply from the company’s own policies and shows both on a review screen. Accuracy on category is high from week one; drafts are edited a third of the time in week one and a tenth by week six. Review stays for anything involving money or notice periods. Managers handle exceptions; the queue is cleared daily for the first time.
Agents, and where a person stays
An agent is a model step that plans and takes several actions to complete a task: look up an order, check a policy, draft a message, update a record. GBCORP builds agents with the actions listed and permissioned, with an approval step for any action that moves money, changes a legal state or sends a message outside the business, and with a log of every action taken. Agents that run without those controls are demonstrations, not systems.
Costs that are visible
Model usage is metered per request and reported weekly beside accuracy, so the business sees the cost of the step and its value on one page. Prompts are trimmed and cached where the model allows; smaller models are used where the evaluation set shows they suffice. A step whose cost rises with volume is redesigned before it becomes a surprise. The support plan covers this monitoring.
Multilingual by default
Assistants answer in the customer’s language from content written in one. Extraction reads several scripts. Classification works across languages. Interfaces are multilingual with right-to-left layouts as a setting. A business serving several markets builds one system.
From one step to a programme
The first step proves the method on an easy process. The second uses the same service, evaluation habit and review screen on a harder one. Within a year, a business has a handful of measured steps and a weekly page showing each one’s accuracy and cost. GBCORP runs its own operations this way, and offers the same programme to clients.
Ownership of the system
The service, prompts, evaluation set and logs live in the client’s own cloud project and are the client’s. The model provider account is in the client’s name with usage visible. 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
Do you train your own models?
Rarely. Commercial models behind an interface, fed the business’s own content, cover most needs; a trained model is built when the evaluation set shows it is needed.
Is our data safe?
It stays in your own cloud project, is masked before it leaves for a request, and is excluded from training under the terms GBCORP configures.
Can you add AI to an app you did not build?
Yes. The step is added as a service beside the existing system, with the same review and measurement.
What does an AI project cost?
A fixed price per release after scoping, with the priced starting point on the GBCORP e-shop, plus model usage at cost.
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 combined view, read an AI and software development company explained.
GBCORP operates across 17 countries on 4 continents through six pillars: AI & IT, Innovation, Sustainability, Infrastructure, Premium Business, and EV's & Solutions.
