AI development, designed around your business

The hard part of an AI project is rarely the AI. It is understanding how an organisation runs, in enough detail to know where a model earns its place.

So we treat it as a design problem before a technical one. Before we write a line of code, we map how the work happens, the tasks people repeat, the handoffs between systems, and the places where hours and money leak away. It is the same service-design practice we wrote about here, and the one we run on ourselves. A project with us usually goes like this.

1. Start with the day-to-day. We trace the real work, from reporting and approvals to document review and the requests that never stop, and find where the time goes. It is rarely where people think.
2. Separate the routine from the judgement. Repeatable, rule-based work is what an agent can take on. The calls that need a person stay with a person.
3. Get the data right first. An agent is only as good as the data it can reach and trust, so the first job is often bringing that data together. It is worth doing on its own.
4. Build where people already are. In Slack, Teams, or whatever your team already has open, over the systems you already run.
5. Keep a person on what matters. The routine runs itself. Anything with real consequences is signed off by a human.

Do it this way and the AI fits how people work, which is what carries a project past the pilot and into the business for good.

Any team can wire a model into an app. What turns that into a system your business can run on, safely and affordably, day after day, is everything around the model. That is the part we design and build.

The model is a component, and a swappable one. We build so that changing it, OpenAI for Anthropic, or whatever comes next, is a small, contained change. That keeps you off any single vendor’s roadmap and lets us fit the model to the job. Often the right choice is a small, cheap model, or no model at all, just a script, because not every job needs a large language model. Right-sizing like that keeps the running cost sensible, and the running cost is forever while the build is a one-off.

Then there is the foundation the model stands on, the data layer it reads from and writes to, the integrations into the systems you already run, the security built in from the start, and a person kept on the decisions that carry weight. This is the difference between AI that holds up in production and AI that stalls after the pilot.

We did not just design this approach. Conduct has fully embraced AI, using it throughout our everyday work.

Our own platform, Waypoint, does the work we used to pay seven subscriptions for, Confluence, Smartsheets, Jira, Visio, Zendesk, Harvest and Harvest Forecast. That is one system we own instead of seven we rented.

Behind it, around twenty-seven agents run the back office over Xero and Slack, some driven by a model and many of them plain scripts. They reconcile payments, age invoices, prepare our government-panel reporting, audit compliance, process leave, and handle HR through Slack apps we built. A person still signs off anything that carries weight.

Underneath sits a data layer, with deep API and agent (MCP) access that lets those agents work the platform directly. It is what lets us forecast cost and delivery from our own history, and it is why our reporting writes itself.

This is how a modern team works now, and it lets the same people do far more than they used to, a productivity shift we wrote about here. The bigger change, though, is what it frees us to do. With the admin and the grunt work handled, our people are back on the craft.