AI Integration Services: Adding a Model to the Software You Run

AI integration services add a model-driven step to software that already exists: an assistant inside the customer portal, extraction inside the invoice queue, classification inside the ticket system, search inside the catalogue. GBCORP builds the step against an interface rather than a vendor, puts a review screen in front of it, measures it weekly and prices it as a release. The existing software keeps working throughout, because the integration is added beside it rather than inside its core.

Where a model step fits

An interface, not a vendor

The app calls a small service GBCORP builds, and that service calls the model. When a better or cheaper model appears, the service changes and the app does not. Prompts, examples and output formats live in the service with version numbers, so a change in behaviour is a release with a date, not a mystery.

The review screen and the evaluation set

At first, a person confirms each output on a review screen; the confirmed outputs become the evaluation set. Accuracy per field is measured against it weekly. Where errors are harmless, the review is removed; where they are not, it stays and the person still saves most of the time. Cost per request is reported beside accuracy, so the business sees the whole price of the step.

Data handling

Documents and prompts go to the model provider only for the request, under terms that exclude training. Sensitive fields are masked before they leave. The service logs what was sent and what came back, in the client’s own cloud project, so an audit is possible and the data never sits in a developer’s account.

Checklist before an integration

• The one step to add first, and the system it lives in.
• Fifty to two hundred real examples with their correct answers.
• Who reviews outputs at first, and for how long.
• Fields that must be masked.
• The accuracy at which the review can be removed.
• Access to the existing system’s API or database, read-only at first.

The numbers to expect

Scoping in one week. The first step live with its review screen in three to six weeks at a fixed price. Accuracy measured weekly; review removed on safe fields within a quarter. Model usage passed through at cost; a support plan at a known monthly figure. Customised or more complex functionality is quoted separately before any work starts.

A worked example

A distributor receives four hundred supplier invoices a week as PDF attachments and types each into its accounting system. The integration adds a service that reads each attachment into a draft record, shows it on a review screen beside the original, and posts it on confirmation. In week one, accuracy on the supplier name and total is high and on line items lower; the review stays on line items. By week eight, line-item accuracy has crossed the agreed threshold on the three largest suppliers, and review is removed for them. The typing is gone; the reviewer handles exceptions.

Integrating with systems that have no API

Many business systems expose no API, or one that is read-only. The service works from exports, mailboxes and scheduled database reads where it must, and writes back through the paths the system allows: an import file, a form, an e-mail to a queue. GBCORP maps the available paths in the scoping week and chooses the one that keeps the existing system untouched. A system with no API is slower to integrate, not impossible.

Keeping the step honest over time

Models change, suppliers change their invoice layouts, and customers ask new questions. The weekly evaluation catches drift as a falling number before anyone notices in the queue. The evaluation set grows from the review screen, so it stays representative. A model swap is tested against the set before release. The support plan covers the evaluation runs, so the step stays as accurate in year two as in week eight.

Choosing the first step

The best first integration is the step with the most repetition, the clearest right answer and the lowest cost of error: reading a form, sorting a queue, answering the questions the support desk answers every day. A step with a subjective answer or a high cost of error is a later project, once the method is proven on an easy one.

Multilingual by default

An assistant answers in the language the customer wrote in, from content written in one. Extraction reads documents in several scripts. Classification works on messages regardless of language. The interface is multilingual with right-to-left layouts as a setting, so a business serving several markets integrates once.

Ownership of the integration

The service, its prompts, its evaluation set and its 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 exit, the service keeps running; the client can hand it to any team, because it is built against an interface and documented.

Questions we hear first

Will it break our current system?

No. The step is added as a separate service the system calls; the system keeps working if the service is switched off.

Which model do you use?

The one that scores best on your evaluation set at the cost you accept; it can be swapped later without changing the app.

Can it read our documents?

Yes. Content the business owns is indexed in its own project and the assistant answers from it, citing the source.

Do we need data scientists?

No. GBCORP builds and measures the step; the business supplies examples and reviews outputs.

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 automation side, read AI automation services explained.

GBCORP operates across 17 countries on 4 continents through six pillars: AI & IT, Innovation, Sustainability, Infrastructure, Premium Business, and EV's & Solutions.