Skip to content

AI agents that read, sort and reply, and know when to hand over

An AI agent is useful when your team spends hours reading messages, deciding what each one is about and typing a reply that is mostly the same every time. We build agents that do that first pass inside your inbox, WhatsApp or CRM, using your own documents as the source, and pass anything uncertain to a person with the context attached.

Typical timeline
3–6 weeks
You receive
Agent, test set, review dashboard
Models
OpenAI, Anthropic, others
A support agent wearing a headset works at his desk

Who it suits

A good fit if

  • Support and admin teams answering the same twenty questions all day
  • Clinics and firms where one person sorts a shared inbox every day
  • Sales teams that reply to new leads hours after they come in

Probably not a fit if

  • Decisions that need a licensed professional, such as a diagnosis. The agent can prepare the case, not make the call
  • Tasks where an occasional wrong answer is unacceptable and no human can review it

Agents we build

  • Inbox triage

    Reads every email or message, works out what it is about and how urgent it is, pulls out the key details, and routes it to the right owner.

  • Support and knowledge assistants

    Answers questions from your policies, price lists and past replies, cites where the answer came from, and asks a person when it isn't sure.

  • Lead qualification

    Answers new enquiries on WhatsApp or email as soon as they arrive, asks your qualifying questions, and books a call for the ones that fit.

  • Document reading

    Pulls fields out of forms, referral letters, invoices and contracts into structured data your systems can use.

Built with

  • OpenAI
  • Anthropic
  • n8n
  • Gmail
  • Outlook
  • WhatsApp
  • HubSpot
  • Salesforce
  • Notion
  • Slack

From first call to live

  1. 01

    Collect real examples

    We start with 100 to 300 real messages or documents from your team, with the replies they sent. That becomes the test set.

  2. 02

    Build the first version

    A working agent connected to a copy of your inbox or CRM, scored against the test set so we can see where it is wrong.

  3. 03

    Draft mode

    For two weeks the agent drafts and your team approves. Every correction goes back into the instructions and the test set.

  4. 04

    Go live with limits

    The agent sends on its own only for the categories it handles reliably. Everything else stays in draft for a person.

How we keep agents accurate

Every agent has a written test set built from your real messages. We measure it before launch and after every change to the instructions or the model, so you can see accuracy as a number rather than a feeling.

Agents answer from your documents, not from the model's memory, and they show the source. If the source doesn't cover a question, the agent says so and hands over.

Your data and the models

We use the business APIs of OpenAI, Anthropic and others, which do not train on API data by default, and we send the model only what the task needs. Where a provider offers zero data retention, we turn it on for sensitive work.

Where we've done this before

Common questions about ai agents

It can, which is why we design against it. Agents answer from your documents, show the source, and hand over when the documents don't cover the question. We measure accuracy on a test set built from your real messages before launch.