One project state for the whole team — whichever AI each consultant uses.

Business situation

A London consultancy is running a review for ABC Limited. One consultant does research with Claude. Another drafts deliverables with ChatGPT. The engagement lead uses Microsoft Copilot. Decisions are agreed with the client in weekly meetings, and interview notes from ABC Limited’s staff are sensitive.

Why existing tools alone are insufficient

Each AI keeps its own memory. A finding one consultant establishes with Claude is invisible to the colleague using ChatGPT, so it gets re-explained, re-pasted or lost. Nothing stops an AI from quietly rewriting a decision the client has already agreed. And interview notes shared into one chat are then outside anyone’s control.

Systems involved

• Kroy State — the engagement’s decisions, findings, open questions and deliverables.
• SharePoint — working documents.
• Microsoft Teams — the project channel.

Kroy architecture

All three AIs connect to Kroy. The engagement is an entity in KroyDB with its State. The Update Project State Skill may add findings and open questions; the Summarise Findings Skill may read State and draft deliverables. Decisions marked Client-agreed are locked: changing one requires the engagement lead’s approval. Interview notes carry a restricted classification, readable only by people cleared for them.

Workflow

1. The research consultant asks Claude: “Record what we learned from this week’s process interviews.” Claude writes three findings and an open question with state.update, each linked to its source.
2. The drafting consultant asks ChatGPT: “Draft the current-state section using the latest findings.” Kroy returns the same State with state.read, and ChatGPT drafts the section with artifact.create.
3. ChatGPT requests the underlying interview notes. Kroy refuses: the drafting consultant is not cleared for restricted notes. ChatGPT works from the findings instead.
4. ChatGPT proposes changing the agreed scope decision to include procurement. Kroy refuses the change and routes it to the engagement lead as an approval request.
5. The engagement lead asks Copilot: “What has changed on ABC Limited this week, and what needs my decision?” Copilot summarises the new findings and the pending scope change.
6. The engagement lead discusses it with the client and approves the change. The decision is updated, with the approval recorded.

Agent permissions

Each AI acts under its own person’s delegation. All can read the engagement’s State. Only people cleared for restricted notes can have their agents read them. No agent can change a locked decision directly.

Human permissions

Consultants add findings and draft. The engagement lead approves changes to client-agreed decisions and manages clearance.

State changes

• Three findings and an open question are added, each with its source.
• A draft section is linked to the findings it used.
• The scope decision moves to a new version after approval.

Audit outcome

Kroy records which agent added or read each item, the refused request for restricted notes, the refused direct change, the approval and the new decision version.

Security considerations

• Context moves between AIs through Kroy State, not through pasted chat transcripts.
• Classification and locked decisions are enforced by Kroy, not by instruction.
• Each finding carries provenance, so the team can see where it came from.

Setup requirements

• The engagement modelled as an entity with State in KroyDB.
• Microsoft 365 connected to Kroy.
• Claude, ChatGPT and Microsoft Copilot connected to Kroy.
• Classification and decision-lock policies configured.