Scope is the whole design
The difference between a dependable assistant and an unpredictable one is almost entirely scope. A narrow agent that answers from your approved documentation, drafts a reply or retrieves a record can be made reliable; an open-ended agent with broad system access and vague goals cannot. We define precisely what an assistant is for, which tools it may call, and what it must refuse, and we treat that boundary as the core of the design rather than a setting added later.
Tools and permissions are granted on a least-privilege basis, so an assistant can only touch the systems and data its job genuinely requires. This keeps behaviour predictable and contains the blast radius if something goes wrong. We would rather ship an assistant that politely declines an out-of-scope request than one that improvises with access it should never have had.
Grounding, guardrails and approval
To keep answers trustworthy, we ground assistants in sources you control and design them to cite where information came from, so a person can check it. Just as importantly, we design them to say when they do not know rather than fill the gap with a confident invention. Input and output guardrails filter unsafe or out-of-scope requests, and every interaction is logged so behaviour can be reviewed and improved.
For anything consequential, the pattern is draft-and-approve: the assistant prepares the action, message or change, and a person reviews and approves it before it takes effect. This keeps a human accountable for outcomes while still removing most of the effort. You decide which actions are low-risk enough to run within tight automatic limits and which always require sign-off.
- A least-privilege, explicit set of tools the agent may use.
- Answers grounded in approved sources, with citations where possible.
- Designed to decline or escalate rather than guess when uncertain.
- Human approval gates for any consequential action.
- Full logging of inputs, outputs and tool calls for review.
When an agent is the wrong tool
Agents are not a universal upgrade. If a task is well defined and deterministic, a conventional workflow or a simple form is cheaper, faster and more reliable than a model deciding what to do. We reach for an agent only when the flexibility of natural language and reasoning genuinely earns its place, and we will recommend a plainer solution when it does not.
They are also not autonomous decision-makers. An assistant supports people; it does not replace professional judgement, and it does not make regulated determinations on your behalf. We are explicit in documentation about what each agent can and cannot do, so expectations are set honestly and no one mistakes a helpful drafting tool for an authority on the matter.
