A RAG AI agent that admits what it does not know
Queries the company knowledge, answers with sources, admits when it lacks enough information and hands off to a person.
AI agents that are bounded, evaluated and hand off to a person
A useful agent is not the one that answers everything: it is the one that knows what it may do, recognises when it lacks enough information, and passes the conversation to a person before inventing something.
Questions to answer before building it
A pilot without these answers ends as a nice demo that never reaches production.
What is the task?
A specific, measurable task, not "handle customers". If it cannot be described in one sentence, it cannot be evaluated either.
Where does the information come from?
Which documents, which systems and who keeps them current. The agent is only as good as its sources.
What can it do besides answer?
Reading an order is different from changing it. Actions with side effects are limited in code, never only in the prompt.
When does it hand off?
What triggers escalation and with what context it arrives. Without this, the agent becomes a wall.
Scope
What we implement
Assistants over your own knowledge
They answer from your documentation citing the source, and admit when they find insufficient support.
Agents connected to your systems
They check the state of an order, an invoice or a ticket in the real system, with least privilege.
See integrations : Agents connected to your systemsChannels where the customer already is
Web, WhatsApp or the product itself. The channel is decided last, not first.
Evaluation before production
A set of real cases with expected answers, so you know whether yesterday’s change made things better or worse.
Handoff and human review
Explicit escalation rules and human approval on actions that carry consequences.
Observability and cost
Conversation logs, errors and model usage, so the bill is not a surprise at month end.
What a pilot includes and what it does not
The pilot exists to decide with data whether to continue, not to leave something half-finished in production.
What it includes
- Definition of the task, the sources and the permitted actions
- Agent implementation over a bounded process
- Evaluation set with real cases and documented results
- Handoff rules to a person and conversation logging
- A continuation recommendation with estimated running costs
What it excludes
- Model usage, billed by the provider and varying with volume
- Cleaning and rewriting the source knowledge, if needed
- New integrations with systems that do not expose an API yet
- A guarantee of accuracy: no serious provider can give one
The pilot is quoted by scope once the task and sources are defined. Build, hosting and model usage are presented as three separate lines so you can compare them.
Recurring third-party costs: Model usage per conversation, hosting and infrastructure, and channel licences (the WhatsApp API, for example).
Evidence
The demo shows the three cases that matter
A valid query with sources, a query with insufficient information where the agent admits it, and a handoff to a person.
Can the agent be wrong?
Yes, and anyone claiming otherwise is not being honest. That is why the work concentrates on bounding what it may do, measuring with an evaluation set, and handing off when support is insufficient. We publish no accuracy percentages or savings: without a comparable baseline they mean nothing.
What does it cost to run each month?
It depends on conversation volume and the model chosen. We estimate it before the pilot using your real volume and present it separately from development: build, hosting and usage are three distinct lines.
Do I need my documentation in order?
It helps a great deal, but it need not be perfect. Part of the pilot is finding out what is missing. What is essential is that it exists and that someone owns keeping it current: an agent over stale documentation repeats errors very confidently.
Can it execute actions, not only answer?
Yes, with limits defined in code rather than in the prompt. Actions with side effects are explicitly restricted, and consequential ones go through human approval.
Related services
AI chatbots that know when to stop talking
Conversational assistants for support and sales: knowledge, tone, evaluation and escalation.
API, ERP and CRM integrations that hold up when something fails
Synchronisation between systems with error handling, retries, duplicate control and monitoring.
Custom software development that connects your operation
Your own systems and platforms, when no off-the-shelf tool solves the process.
Have a specific task to evaluate?
Tell us what the agent should resolve, with which information and what it should be able to do. We will say whether a pilot makes sense.