Custom AI Solutions: When Off-the-Shelf AI Stops Fitting

Custom AI solutions are AI steps built around one company’s data, process and measure of success. A general assistant can draft an email; it cannot reliably extract the twelve fields your finance team needs from your suppliers’ invoices, classify your support tickets by your categories, or answer staff questions from your policy documents with a citation. Custom AI does those things, measured against your own examples, inside your own systems. GBCORP builds custom AI as small, measurable steps with review, and prices each one before work starts.

Signs that off-the-shelf no longer fits

Typical custom AI steps

Extraction: fields from invoices, forms, contracts or delivery notes into the system, with a person reviewing uncertain values. Classification: tickets, emails or documents sorted by your categories and routed. Answering: staff or customer questions answered from your documents, citing the source. Drafting: replies, summaries or reports from your data in your tone, approved before sending. Each starts with a hundred real examples and the correct answers.

Measured, reviewed, costed

Every custom step has an evaluation set run weekly, a review rule for uncertain output, a log of what a person changed, and a cost per task. Review is reduced only where the figures support it. A step that cannot beat the manual process on the evaluation set is stopped, and the company has spent a scoping week rather than a year.

Checklist before commissioning custom AI

• The one task to start with, and how many times a week it is done.
• A hundred real examples with the correct results.
• The system the result must land in.
• What must be reviewed by a person, and by whom.
• The data the step may and may not use.
• The figure that proves it worked: hours, errors or response time.

The numbers to expect

Scoping in one week, priced on its own, ending in an evaluation set and a fixed quotation. A first custom step live in three to six weeks. Running costs reported monthly. Later steps as releases at fixed prices. Customised or more complex functionality is quoted separately before any work starts.

A worked example

Consider a freight forwarder whose team types data from shipping documents into its operations system. The scoping week collects a hundred documents with the correct fields and sets the measure: fields correct without a person changing them. The custom step reads each document, fills the fields, and sends uncertain ones to a person. In the first weeks every result is reviewed; as the evaluation figures hold, review is reduced to uncertain fields only. Typing time falls, and every change a person makes improves the next version.

Custom does not mean starting from zero

Custom AI uses established models and tools; what is custom is the instructions, the examples, the rules, the review and the integration. That keeps cost and time reasonable. Training a model is considered only when the evaluation set shows that established models cannot reach the required accuracy, which is rare for the first steps most businesses need.

Ownership of the AI work

The evaluation set, the instructions, the review rules and the logs belong to the client and sit in the client’s accounts. If the client changes supplier, the work goes with it. The evaluation set is the most valuable part: it is the definition of a correct answer for that business, and it outlives any model.

Running costs in plain figures

Each step reports cost per task and per month beside the hours it saves. Simple tasks use cheaper models; repeated answers are cached. A monthly limit stops runaway costs until someone approves more.

Data protection

Model access is configured under providers’ business terms, personal data is masked where the step does not need it, and logs sit in the client’s own account. The data a step may read is listed in writing.

From one step to several

The parts built for the first step, model access, evaluation, logs and cost limits, are reused by the second. Each later step is faster and cheaper, and they share one weekly report.

When not to use AI

Where a fixed rule does the job, GBCORP writes the rule. Where the volume is a few documents a week, a person is cheaper. Where no one can say what a correct answer is, the scoping week ends with that finding rather than a build.

Questions we hear first

Do we need a lot of data?

Usually not. A hundred well-chosen examples are enough to build and measure most first steps.

Will you train a model from scratch?

Rarely. Most custom steps use established models with your documents, rules and examples; training is chosen only when it measurably helps.

What if the AI is wrong?

Uncertain output goes to a person, every correction is logged, and the evaluation set shows whether accuracy holds.

Can it run inside our systems?

Yes. Results land in your system through its interface, with permissions you define.

Where to start

The GBCORP AI Development and Automation product is the priced starting point; the GBCORP AI development hub sets out the method. For what is and is not in the price, read AI development services explained.

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