An agent is a system, not a prompt.
Six components have to exist and agree with each other before an agent is safe in front of a customer. Here is each one, and the job it does.
Six parts. Each with a job.
Skip one and you get the behaviour everyone’s afraid of: confident answers from nowhere, actions without permission, customers stuck in a loop.
ChannelsWhere conversations arrive
Email, chat, messaging, contact forms. The system meets customers where they already are — no new portal to learn, no context lost between channels.
OrchestrationWhat happens, in what order
The plan behind every reply: identify the intent, gather what’s needed, act or decline, record the outcome. Deterministic where it can be, model-assisted where it helps.
KnowledgeWhat the agent may say
Your policies, product docs and history — curated, versioned and cited. An answer that can’t name its source doesn’t get sent.
ToolsWhat the agent may do
Allow-listed actions into your systems: look up an order, check a status, draft a note. Reads by default; writes only where you’ve agreed, with checks in front.
GuardrailsWhere the agent must stop
Scope checks on every step: topics it won’t touch, actions it won’t take, thresholds that trigger handover. Tested as failure modes, not hoped for.
Human consoleWhere people stay in charge
A queue your team actually enjoys: handovers with context, one-screen review, and the record of every automated step when they need to look back.
What we build agents to do. And not do.
This list is written down with you before a build starts, and it ships as configuration — not as a slide.
- Answer from your knowledge — with the source attached, so anyone can check.
- Collect and structure — order numbers, details, documents, before a person gets involved.
- Act within agreed scope — look up, draft, file, schedule; the list is explicit.
- Decline gracefully — out-of-scope gets a plain answer and a route to a person.
- Leave a record — every step logged: what it read, what it did, why.
- Invent policy — no source, no answer. Silence routes to a human, not to a guess.
- Make commitments — refunds, exceptions, contract terms: a person decides.
- Write without verification — account changes wait for identity checks you define.
- Pretend to be human — customers are told they’re talking to an agent.
- Hide its mistakes — bad answers surface in review, with the trail to fix them.
The handover is the product.
Most agent projects treat escalation as failure. We treat it as the moment that earns trust: the customer hears a straight answer (“this needs a person”), and the person who picks it up starts with everything already gathered.
A good handover means the customer never repeats themselves. Ever.
That’s why every system we build ships with the human console first. Automation earns its scope over time, from evidence — yours.
Let’s scope your first agent properly.
We’ll start with your knowledge base and one conversation type — and we’ll define what the agent may not do before what it may.
Email info@finlogiq-ai.comNorwest, NSW 2153, Australia