Four stages. One loop.

We work with the people who do the work to write down how it runs: the things your business deals with, the steps they take, the judgement calls and the rules. The interviews are structured, not open-ended. Getting people to articulate how they really make decisions is a skill in its own right, and Foundations is built around it.

What they tell us is encoded into the product. Most of this is generated with AI, supervised by people who know your industry — which is why it's fast — and reviewed with your team, who correct it in their own words.

You get: your operation written down in your own terms, and connected to the systems it runs on, ready to be checked against what actually happened.

Changes arrive in two ways. Groundwork proposes them from what the last Run recorded — a concept it keeps meeting, a workflow branch for a request type it keeps escalating, a rule worth tightening — and you accept, edit or reject each one. Or your team makes them directly, with Operator, an agent for your own people that knows your Context Layer and helps draft the change.

Either way a change is a new version of the workflow. The old one keeps running until the new one has been proved.

We take a representative sample of your real historical cases — conversations, tickets, requests — and work through them against the map with the people who handled them. This shows how your business actually behaves (as opposed to how the process doc says it does), where the judgement calls are, and what "good" looks like in your terms. Every gap becomes a correction to the map, and the cases themselves become the evaluation set.

You get: a working Context Layer, corrected against your own history; an evaluation set built from your own cases; agents configured and ready to prove; and an honest view of what Groundwork will automate now, what stays human-in-the-loop, and what shouldn't be touched. This is the Foundations deliverable — it's yours, in your Groundwork account, whatever you do next.

The product replays your own live traffic. Every conversation the agents handled, and every one they escalated, is on the record with its attributes and the decisions behind it. Escalations show where the map is incomplete, corrections your team made show where it is wrong, and new request types surface as they appear.

A conversation that went well can be pinned as a golden case, its outcome becoming the expected one. A conversation that went badly can be corrected by hand and kept as a guard against the same failure. Either way the evaluation set grows from your own history, not ours.

Foundations ends here. Agents run against the evaluation set built in Replay, and you see, case by case, what they would have done: the same as your team, or different, and why.

The bar is yours to set — which case types the agents must handle, what accuracy is acceptable, and what stays human. Go-live is a decision you make on the evidence, not a date on a plan.

Every change goes through Prove before it is live. A new workflow version, an edited policy, a renamed concept — each is run against the evaluation set, and the diff shows exactly which past cases would now be handled differently. If the only differences are the ones you intended, it goes live. If not, it doesn't.

Go-live is not a switch-over. Workflows are enabled one at a time, so you can start with the request types you are most confident about and add the rest as the evidence comes in. Actions that write to your systems are staged for a person to approve until you decide otherwise.

Agents work from the current proved version of the map. Escalations arrive with the conversation summarised and its attributes attached, and every reply and decision is written to the record — including the ones your team makes.

That record is what the next Map works from.