GBCORP BLOG · AI
AI Agent Development: Tools, Limits and a Review Screen
Building an AI agent is mostly deciding what it may do, in what order it earns each permission, and how a person sees what it did. The model is the smallest part.
AI agent development produces a model step that acts: it reads a request, uses tools to look things up and do things, and finishes a job or hands it to a person. GBCORP builds agents inside the software it already runs for a business, with the same method as any AI step, an evaluation set, a review screen, measured accuracy and cost, plus the parts that are specific to agents: tools, limits and logging.
How an agent is built, step by step
| Step | What happens | When |
|---|---|---|
| Define the job | One job, fifty to two hundred real examples with correct outcomes | Week 1 |
| Grant read tools | Look up records, search documents, check availability | Weeks 2–3 |
| Add the review screen | Every proposed action shown to a person with the context | Week 3 |
| Measure | Accuracy, escalation rate and cost per job, weekly | From week 4 |
| Grant write tools | Send, book, create, one at a time, with limits, as accuracy holds | Weeks 6–12 |
| Remove review where safe | Routine categories run alone; exceptions still reach a person | From the numbers |
Tools and the order they are granted
Read tools first, because a wrong lookup costs nothing. Then drafting tools, because a person still sends. Then write tools with limits: send to this list, book within these rules, create but never delete, amounts below this figure. Each tool has a description the agent reads, a schema it must follow and a log entry for every call.
Limits and hand-over
Confidence below a threshold, a request outside the job, a sum above a limit, a customer asking for a person, an unusual pattern: each hands the job to a human with everything the agent saw attached. A good agent hands over often at first and less as the evaluation set grows.
Logging for audit
Every model call and every tool call is logged: what the agent saw, what it decided, what it did, and what it cost. A person can replay any job. This is what makes an agent acceptable to a security team and a regulator, and it is built before the first write tool.
Checklist before development
- One job, with examples and correct outcomes.
- The tools, read and write, in the order they will be granted.
- The limits and hand-over rules, written down.
- Where the review screen sits and who staffs it.
- What must never be automated.
- The number that proves it: time to resolve, accuracy, escalation rate.
The numbers to expect
A first agent with read tools and a review screen in four to eight weeks at a fixed price. Write tools added as releases, priced before they start. Weekly reports on accuracy, escalation and cost per job. A support plan covering evaluation runs, model updates and cost reporting. Customised or more complex functionality is quoted separately before any work starts.
A worked example
A travel agency receives booking changes by message all day. The agent is built in three stages. Stage one, read tools: it looks up the booking, the fare rules and availability, and drafts a reply for an agent to send. Stage two, after four weeks of measured accuracy, a write tool with limits: it may hold a new flight within the same fare class and below a fare difference, and must hand over anything else. Stage three, after a further month, routine changes complete without review, with the log showing every step. The agents handle exceptions and complex itineraries, and response time falls from hours to minutes.
Designing tools the agent can use safely
Each tool has a plain description, a strict schema for its inputs, and a result the agent can read. Write tools carry their limits inside them: a send tool that only sends to the customer on the case, a booking tool that refuses fares above a figure, a create tool with no delete. The agent cannot exceed a tool’s limits because the tool enforces them, not the prompt. GBCORP designs tools this way so a mistake in reasoning cannot become a mistake in the world.
Cost per job and when to stop
An agent calls a model several times per job, so cost per job is tracked from the first week and shown beside accuracy and escalation rate. Where cost per job exceeds the value of the job, the agent is narrowed or replaced by a rule for that category. A pause switch per category lets the business stop an agent instantly, and the evaluation set runs on every change so a model swap or a prompt change is measured before it is live.
Agents inside existing software
An agent is most useful where the work already is: the support console, the booking screen, the document queue. GBCORP adds agents as releases inside the client’s existing app or console on its engine, with the same accounts, roles and audit history, so the person reviewing the agent’s work is in the screen they already use.
Where the data goes
The agent runs in a cloud project in the client’s name. Prompts and documents go to a model provider only for the request and are not used for training under the terms GBCORP configures. Sensitive fields are masked before they leave. Every tool call and model call is logged, so a customer’s security team gets a plain answer.
Pricing per release
A scoping week priced on its own. A fixed price for the first agent with read tools and a review screen. Each write tool or new category as a priced release. Model usage passed through at cost and shown per job. A support plan at a known monthly figure. Nothing priced by the hour.
Questions we hear first
Can the agent use our systems?
Where they have an API, yes, as tools with limits; where they do not, GBCORP says so.
What stops it doing something harmful?
Tools granted one at a time, limits in the console, the review screen, and a pause switch per category.
How much does an agent cost to run?
Model usage at cost, shown per job in the console beside accuracy.
Which models?
Commercial models behind an interface, swappable; no client data used for training.
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 to ask a supplier, read choosing an AI agents company.
GBCORP operates across 17 countries on 4 continents through six pillars: AI & IT, Innovation, Sustainability, Infrastructure, Business Lounge, and EV's & Solutions.
