Skip to content

AI Agents & Assistants

AI agents that know when not to answer.

We build AI agents that work from a knowledge base you control. They answer in your voice, show where an answer came from, and hand over to a person when a question falls outside what they know.

Book a call See the work

What we build

The work, in plain terms.

  • 01

    Customer-facing assistants

    Website and storefront chat that answers from live data, not from the model’s memory.

  • 02

    Internal knowledge assistants

    Assistants in Slack over documentation, tickets and unwritten expertise.

  • 03

    AI sales agents

    Reply handling across email and SMS, with qualification and booking.

  • 04

    Guardrails and testing

    Input checks, output filtering and a question set that runs after every change.

How we approach it

Three rules we do not bend.

  • Retrieve first, then generate

    The model only sees content that was looked up and verified. It cannot describe something it was never given.

  • Unsure means hand over

    Below an agreed confidence level, the agent stops and passes the question to a named person.

  • Accuracy is a number

    We agree a benchmark with you before the build and test against it.

Our work here

The proof, not the promise.

See all work

Home & Relocation Services

AI phone and chat agent that captures leads after hours

We built a voice and chat AI agent for a senior move management company, qualifying callers around the clock and writing booked consultations straight into their industry CRM.

24/7Inbound call coverage

Manufacturing

Internal AI assistant built for accuracy, not confidence

We designed an enterprise knowledge assistant that answers technical support questions in Slack from scattered documentation and undocumented expertise, with citations and escalation instead of guesses.

5Layers of accuracy safeguards

Questions

What people ask before they start.

Will the agent make things up?

We design it so that it cannot answer from its own memory. It retrieves from your approved content first and must point to a source. No system is fully immune to error, so we add layered checks and test against real questions before launch.

How do you keep the agent on brand?

We write it against your own material and existing messages, then test it on real conversations and correct anything that reads wrong before it speaks to a customer.

What happens to conversations the agent cannot handle?

They go to a person on your team, with the conversation so far attached, and your team is alerted straight away.

Insights

Related reading.

Industries where we have done this

Need this in your business?

Book a call See the work