AI agent development

3–6 weeks typical Fixed quote after scoping

Design and build of a production AI agent — customer- or staff-facing — grounded in the business's own data, with server-enforced guardrails and evaluation.

Scope your project →

What we build

A production AI assistant for your customers or staff: it converses naturally, answers only from your business's real information, escalates to a human when it should, and cannot invent facts — every consequential rule is enforced in server code, not left to the model's judgment. An evaluation suite proves how it behaves before and after launch. The content it answers from — your catalogue, policies, documents — is yours to author and maintain: we build the machine that uses your information, not the information itself. Choose this when you need a new assistant built; if you mainly need AI connected to systems you already have, see MCP server development.

What you get

  • The agent's source code in a Git repository
  • A server-side validation layer that rejects out-of-policy tool calls, with tests in the repo
  • A scenario evaluation suite scoring the agent's conversations, runnable with one command

How the work unfolds

  1. Design the agent loop and trust boundaries
  2. Engineer the system prompt and playbook
  3. Wire the agent's tool access
  4. Add structured confirmations for irreversible steps
  5. Meter token spend per conversation
  6. Test the agent against real scenarios

What shapes the price

Before you see a number, our scoping conversation asks:

  • How many of your systems should the assistant be able to act in? Count everything it touches — your CRM, calendar, ticketing, chat tools, payment provider. Each system the agent can touch is its own tool surface, permissions and failure modes. Count the systems the brief already names before falling back to the default — a brief that lists a chat tool, a project board and a payment provider has named three.
  • Will the assistant take actions that are hard to undo — sending money, deleting records, emailing customers? Structured confirmations before irreversible steps are up to four days of work.
  • Do you need per-conversation cost tracking from day one? Metering token spend per conversation is a day or two.

How an engagement starts

This work builds on Discovery workshop, so scope, boundaries and constraints are agreed before anything is built.

Teams often combine it with:

Get a priced plan →