What we doAIWorkWhereHowStart a project
Home/What we do/AI & Data
AI agents & workflow automation

An agent that checks real stock,books real slots and knows whento hand over.

Agents that read your live data, act behind server-side guards and escalate to a person with the full context — on WhatsApp, on the phone, on your site.

Get in touchRelated work
WhatsApp · Voice · WebTool-callingBooking & refund guardsTest labCost per conversation
guards: server-side ✓cost / conversation $0.003
Why it works

Guards on the server. Model as a setting. Tested like software. The model is a configuration value.

What's included

Everything the engagement delivers.

Each piece already runs for a client — nothing here is a slide.

01

Discovery

Five days mapping real conversations, the systems the agent may touch, and its limits. Written plan + estimate.

02

Agent core

Tool-calling against your database and APIs; every write passes a server-side guard. In production · WhatsApp sales agent.

03

Channels

WhatsApp Business API, web widget, inbound voice, Telegram, email — one agent, one memory. In production · voice + WhatsApp + web.

04

Automations

Event-driven workflows: enquiry → CRM, booking → reminder, order → courier. Retried, deduplicated. In production · order workflows.

05

Test lab

Scripted conversations that must pass before any prompt, tool or model change goes live. Regression suite blocks deploys.

06

Operations console

Transcripts, tool-call traces, latency and cost per conversation, escalation queue, alerts. Handover + staff training.

How it runs

Guards on the server. Model as a setting. Tested like software.

A

The model is a configuration value

OpenAI, Anthropic and Groq-hosted open models in use; switching is a test run and a deploy, not a rewrite.

B

Retrieval is a last resort

Direct database and API tools answer exactly; vector search only when data outgrows direct queries.

C

The agent proposes, your code decides

Refund caps, booking rules, identity checks and rate limits are enforced in functions you own.

D

Every call is logged and priced

Transcript, tool calls, latency and cost per conversation; per-tenant cost caps with alerts before the cap.

Questions buyers ask

Before you write to us.

01

Which model do you use, and can we switch later?

+
Whichever fits the job on cost, latency and accuracy; today a mix of OpenAI, Anthropic and Groq-hosted Llama. The model is a config setting, so switching later is a test run and a deploy.
02

How do you stop it inventing answers?

+
It answers from tools that return your data, not from memory. Where no tool has the answer, it says so and offers a person. The test lab replays real conversations before every change.
03

Can it take actions, such as booking or issuing a refund?

+
Yes, within limits you set. Each action is a server-side function with its own rules; the agent can request it, your code decides.
04

What does it cost to run each month?

+
Model usage is billed at cost and shown per conversation; most deployments run at a few cents per conversation. The run retainer covers monitoring, tuning and evaluation.
05

What happens when a customer asks for a human?

+
The agent hands over immediately with the transcript, the customer's details and what it tried. Handoff rules, hours and destinations are set by you and reported monthly.

Find out where an agent
pays back first.

Five questions, a written report within 48 hours, no call required.

Take the free AI audit