GBCORP BLOG · AI
Custom AI Development: The Build, Week by Week
Custom AI development is less mysterious than it sounds. It follows a sequence of weeks with a clear output each week, and a business can see at every step whether the AI is good enough to continue.
Custom AI development builds an AI step around one business task: reading documents, answering questions from the business’s own content, classifying requests, or drafting replies. GBCORP runs it as a sequence of weeks, each with a visible output and a decision. The business sees accuracy measured against its own examples before it commits to the next stage, and the project can stop early, cheaply, if the evidence says so.
The build, week by week
| Stage | What happens | Output |
|---|---|---|
| Scoping week | The task, a hundred real examples with correct results, the measure, the data rules | A scope, an evaluation set and a fixed quotation |
| Weeks one and two | A first version tested against the examples | An accuracy figure; continue, adjust or stop |
| Weeks three and four | Integration with the business’s system and a review screen | A working step inside the real process |
| Week five | Pilot with a person reviewing every result | Corrections logged; accuracy on live work |
| Week six | Launch with a review rule for uncertain results | The step running, with a weekly report |
| Every week after | Evaluation set re-run; costs and corrections reported | Review reduced only where figures support it |
The people the business provides
An expert who knows the task and can say what a correct result is. A reviewer who checks output in the pilot and after. An owner who reads the weekly report and decides whether the step expands or stops. None needs technical knowledge; each needs a few hours a week in the first month.
Where a project should stop
After weeks one and two, if the first version is far from the accuracy the task needs. After the pilot, if corrections show a pattern that better instructions cannot fix. Stopping at either point costs a few weeks rather than a year, and the evaluation set and findings remain the business’s to use later.
Checklist before starting custom AI development
- Choose one task and count how often it happens.
- Collect a hundred real examples with correct results.
- Name the expert, the reviewer and the owner.
- List the data the step may and may not use.
- Agree the accuracy the task needs.
- Agree the figure that shows value: hours saved or errors avoided.
The numbers to expect
Scoping in one week, priced on its own. A first AI step live in about six weeks, within the usual three to six weeks of build. Running costs reported monthly. Later steps faster, reusing the same parts. Customised or more complex functionality is quoted separately before any work starts.
A worked example
Consider a property-management company that answers hundreds of tenant emails a week, many about the same twenty topics. In the scoping week, staff collect a hundred emails with the replies they sent. In weeks one and two, a first version drafts replies from the company’s policies and scores well on most topics, but poorly on deposits, where the rules are complicated. Deposits are excluded. Integration with the email system and a review screen follow, the pilot reviews every draft, and the step launches with drafts sent only after a person approves them. The deposit topic waits for a better rules document.
The evaluation set as the contract
The hundred examples with correct results are more than a test; they are the agreed definition of good work. The quotation refers to them, progress is reported against them, and the client keeps them. If the business later changes provider or model, the evaluation set shows at once whether the new option is better or worse.
Instructions, rules and versions
A custom AI step is mostly instructions and rules written for the business’s task, plus examples. Every change to them is versioned and tested against the evaluation set before release, like any other software change. A business can see which version produced which answer, and roll back if a change makes things worse.
Costs, week by week
Each stage has a fixed price agreed in the scoping week. Running costs are estimated per task during the first version and reported monthly after launch, with a limit that stops unexpected spend.
Data protection
The data a step may read is listed in writing. Personal details are removed where the task does not need them, models are used under business terms, and logs sit in the business’s own account.
From one step to several
The second AI step reuses the first one’s parts: model access, evaluation, review screen and logs. It moves through the same weeks faster, and both steps appear in one weekly report.
Ownership terms
Evaluation sets, instructions, rules, integration code and logs sit in the business’s accounts. Model provider terms can be in the business’s name. Running costs are paid directly, with no mark-up from GBCORP.
After launch
The weekly report shows accuracy, corrections, volume and cost. When figures hold for several weeks, review is reduced for the categories that earned it. When they drop, the cause is found and fixed before review is reduced again.
Questions we hear first
Can it go faster?
Simple steps sometimes launch in three weeks; the pilot is never skipped, because it is where accuracy is proved.
What if we do not have a hundred examples?
The scoping week collects them from your records; most businesses have more than they think.
Do you train a model?
Rarely. Most steps use established models with your instructions, examples and rules.
Who owns the result?
You do: the evaluation set, instructions, review rules and logs sit in your accounts.
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 how GBCORP measures what it builds, read what an AI development company should measure.
GBCORP operates across 17 countries on 4 continents through six pillars: AI & IT, Innovation, Sustainability, Infrastructure, Premium Business, and EV's & Solutions.
