Tool surface
Typed functions with validated arguments for each action the agent may take, running behind the same authorisation as the rest of your system.
Solution
AI Agents and Assistants
Assistants embedded in your product or operation that answer from your data, take scoped actions through typed tools, and escalate to a person when they should.
Overview
What the system does
An agent is an application that uses a language model to decide which of your functions to call. The model supplies judgement about intent. Your code supplies the actions, the permissions and the limits.
That framing decides everything about the build. The interesting work is the tool surface, the retrieval quality, the guardrails and the evaluation, not the prompt.
Used well, an assistant handles the volume of routine requests that currently consume a team, and hands over the ones that need a human with the context already gathered.
Who needs this
Teams with document-heavy questions or repeatable tasks that need language interpretation and controlled access to business tools.
What we build
Build an LLM integration around retrieval, typed tools and measurable tasks. RAG retrieves authorized passages; vector search alone does not establish permissions or answer quality. Tool calls propose validated actions, with human approval for consequential changes.
Modules
Core modules
Retrieval
An index over your documents and records, filtered by the requesting user permissions, returning passages with their source so answers can cite them.
Conversation state
Session memory, and durable memory where it earns its place, with a clear policy on what is retained and for how long.
Escalation and handoff
A defined path to a person, carrying the transcript and the gathered context, triggered by topic, sentiment, repeated failure or explicit request.
Approval steps
Actions with a cost prepared as proposals that a person releases, so the agent can do the work without holding the authority.
Guardrails
Scope boundaries, spend caps, per user rate limits and refusal behaviour enforced in code rather than requested in a prompt.
Observability
Redacted tool-call traces, model versions, error states and review outcomes. Transcript retention is limited to the approved purpose, with access control and deletion rules for sensitive content.
Evaluation
A fixed set of real cases with expected outcomes, run before every prompt or model change.
Architecture
How it fits together
A typical implementation flow, with explicit responsibilities at each boundary.
Integrations
- Language model providers, hosted or private
- Your CRM, ticketing or operational system
- Document stores for retrieval
- Messaging channels such as web chat and email
- Workflow orchestration for multi system actions
- Analytics for containment and escalation reporting
- 01
Ground
Resolve the user, retrieve permitted sources and attach citations.
- 02
Propose
Validate tool arguments and prepare an action within a narrow permission scope.
- 03
Evaluate
Check the result, require approval where needed and record redacted operational traces.
Security and controls
Design for the exceptions, not just the happy path.
Treat source documents as untrusted input. Enforce authorization outside the model, cap spend and iterations, redact logs, and test prompt injection. Evaluation sets, monitoring and a human handoff are delivery requirements.
Technology options
We select the final stack after reviewing your existing systems and operating requirements.
Security and IP practicesFit
Industries that use this
FAQ
Questions people actually ask
What can an agent do that a chatbot cannot?
Act. A chatbot answers from a script or a page. An agent reads your records, calls your functions, and changes something, which is why the permission model matters more than the wording.
How do we stop it inventing an answer?
By grounding answers in retrieval with citations, and by designing the behaviour for low confidence: a short deterministic response and a handoff rather than a guess.
Can it work in more than one language?
Yes. Model based assistants handle multilingual conversation well. Where accuracy matters, the evaluation set covers each supported language rather than assuming parity.
What is a reasonable containment rate?
It depends entirely on the request mix, and we will not quote a figure before measuring yours. The evaluation set and the first weeks of transcripts give you a real number.
Where does the conversation data live?
In your database, with a retention policy you set. Where data cannot leave your environment we design for a private or on device model.
How do we improve it over time?
Transcripts feed the evaluation set, failures become cases, and changes are scored against that set before they ship.
Let’s talk about the system you need.
Tell us where the current process breaks down. We can map the workflow, examine the integration requirements and define a practical scope for the build.
- A mutual NDA is signed before project details are shared.
- Scope is fixed in writing before any build starts.
- You own the source code, the IP and the documentation at delivery.
- Support plans are available and tailored on a discovery or technical call.