AI agents
Task agents: one job each, done precisely
The useful agents are narrow. One reads incoming orders and creates them for review. One checks a supplier invoice against the purchase order and the receipt. One routes a leave request to the right approver with the balance attached. Each has a brief, a set of tools and limits, and a record of everything it did.
- Built for one task
- Tool calls with permissions
- Measured on your cases
In short
A task agent is an AI agent built for a single, well-defined job in your operations. It holds context across the steps of that job, calls your CRM, ERP, payroll or messaging systems through the permissions of a named role, stops at checkpoints where a person must decide, and logs every action. It is measured against your own cases before it goes live, and a person stays in the loop wherever it matters.
What we build
Reading agents that turn an order from a WhatsApp message, an email or a photo into a draft record for a person to confirm. Checking agents that match an invoice to its purchase order and receipt, or a payment to its invoice, and flag what does not add up. Routing agents that classify a request, pull the account history and send it to the right person with a drafted reply.
Staff-facing agents that answer leave, attendance and payslip questions from AurumOne in the employee’s language. Each one works inside the permissions of a role, so it can do no more than the person it assists.
- Order and enquiry reading with review
- Invoice, receipt and payment checking
- Request classification and routing
- Staff questions on leave, attendance and pay
- Internal lookups across CRM, ERP and payroll
How we keep it governed
An agent gets a role, not the keys. It acts through the same permissions a staff member would have, every tool call is recorded, and actions above a threshold wait for approval. Where an answer must be right, the agent is grounded in your documents and data and shows its source.
Before launch we build an evaluation set from a hundred of your real cases and measure the agent against it. It goes live when it beats the baseline on your data, not on a public benchmark, and it is measured again as your business changes.
- Least-privilege access through a role
- Approval checkpoints by amount or type
- Trace logs of prompts, tools and outputs
- Evaluation on your own cases before go-live
- Rollback for any deployed change
What it connects to
Your messaging channels, your CRM and order system, your ERP or accounting package, AurumOne or your HR system, and the language models of your choice, running through cloud APIs or a private deployment.
- WhatsApp Business and email
- CRM, ERP and accounting
- AurumOne and HR systems
- Cloud or privately deployed models
Where it shows up
Automation playbooks this service runs in
Task Agents
Common questions
Will the agent make things up?
It can, which is why ours are grounded and limited. Answers come from your documents and records with the source shown, actions are restricted to a role’s permissions, and anything the agent is not confident about goes to a person. We measure error rates on your own cases before launch and keep measuring afterwards.
Which AI models do you use?
Whichever fits the task and your data rules. Some tasks suit a large cloud model, some a smaller one, and some need a model deployed privately so data never leaves your environment. We choose per task and can switch as models improve.
Where does our data go?
It stays in your systems. The agent reads what its role permits, the logs are yours, and we agree in writing which models may see which data. Private deployment is available when the data must not leave your infrastructure.
Related