AI Agents · Custom enterprise AI agents
Custom AI agents, engineered around your processes.
Some work is too specific for an off-the-shelf bot. A custom agent gets one defined job, access to exactly the tools it needs, clear limits, and a person in the loop wherever a mistake would matter.
The problem
Where the time goes today.
- Work that crosses three or four systems, so staff spend the day copying between them.
- The same operational questions answered by hand every day: where is it, what is owed, what happened.
- Requests that need reading, a lookup and an action, which rule-based automation cannot handle.
- Off-the-shelf chatbots that do not know your data and cannot do anything in your systems.
How it works
What the agent does, step by step.
- Define the job and the limitsOne clear responsibility, the people it serves, what it may never do, and who approves what.
- Connect approved toolsRead access first. Each tool is a narrow, named capability with its own permission, never open access to a database.
- Ground it in your dataAnswers come from your records and documents through retrieval or fixed queries, not from the model’s memory.
- Add approvals and loggingSensitive actions wait for a person. Every request, tool call and result is recorded.
- Test on real examplesWe evaluate on your own past requests before launch, including the awkward ones.
- Monitor and improveAfter launch someone reviews logs, failures and cost, and the agent is tuned.
Integrations
What it connects to.
“Used in our projects” means we have built and run that kind of connection. Anything else needs a feasibility review before we commit.
| System | Status | Note |
|---|---|---|
| REST APIs | Used in our projects | The usual route into modern systems. |
| PostgreSQL and MySQL | Used in our projects | Through approved queries, not free-form SQL. |
| Existing custom software | Used in our projects | We build and maintain such systems ourselves. |
| Third-party ERP or CRM | Proposed: feasibility review | Depends on the vendor’s API and your licence. |
| Email and calendar | Proposed: feasibility review | Draft-and-propose first; sending is a separate decision. |
In practice
Examples of the work.
Illustrative exchanges with fictional data. They show intended behaviour, not recorded client conversations.
Order status agent
Staff: “Which orders for Customer X are late?”
Agent: checks the user’s role, runs the approved “late orders” query, and lists the orders with their stage and due date.
Proposed action: draft an apology message to the customer, for a person to approve.
Month-end preparation helper
Staff: “What is still unreconciled for October?”
Agent: runs the approved reconciliation query and lists the differences with their sources.
It cannot post entries; it can open a task for the accountant.
Security and control
What keeps it safe.
- Role-based access: the agent can only do what the person asking could do.
- Narrow, named tools. No general database or shell access.
- Sensitive actions require human approval and are logged.
- Defences against prompt injection: instructions inside documents or messages are treated as data, not commands.
- Data minimisation: only the fields needed for the task are sent to the model.
Implementation
How we would approach it.
- Scoping workshopMap the process, the data and the risks. Decide what the agent will and will not do.
- Blueprint and proposalA written design, an evaluation plan and a fixed proposal.
- Build and evaluateTools, retrieval, approvals and logging, tested on real examples.
- Pilot and launchA limited rollout with supervision, then wider release.
Commercial considerations
What it costs, and what it saves.
- Build cost is scoped after the workshop and quoted as a proposal.
- Running costs are separate and visible: model usage, messaging fees, hosting, maintenance and supervision.
- Ownership, licensing and data-handling terms are set out in the proposal.
- The AI ROI calculator lets you test whether the numbers work before you commit.
Build this agent with Odontofusion.
Tell us the job. We will tell you honestly whether an agent is the right tool for it.