Recommendation is the easy part. The work that earns WorkReef its place is what happens after approval. Provisioning the agent. Wiring the integration. Running it shadow alongside the human baseline. Measuring agreement. Proposing promotion when the bar is met. Refusing to promote when it isn't.
You hand WorkReef admin consent on Microsoft 365 and three connector tokens. The Cartographer agent goes to work. By the end of the hour you have the org as it actually operates. Not the Visio diagram your COO drew in 2023.
147 stakeholders pulled from Entra. 23 free-text department strings collapsed into 9 canonical departments. 11 workstreams nobody had written down. A list of every Claude, ChatGPT, Copilot, RPA bot, and "AI Power Automate flow Greg built last quarter" inferred from the data. You haven't typed a name. You haven't filled a form. You will not be asked to write a spreadsheet at any point in this product.
The Concierge agent walks you through the rest of the connections in one chat, paste at a time. It opens the right vendor admin page for each, probes the credential the moment you paste it, narrates what it found ("87 tickets, 12 tags, 6 active agents in last 30 days — looks good"), and surfaces any scope shortfall inline with a fix link.
A position has a name, capacity, cost, and a task portfolio. The holder is either a human or an AI agent. The rest of the platform does not care which. When AI takes a task, the assignment moves from one position to another, and capacity rebalances cleanly across the department. There is no "AI department" running parallel to the real org.
The capacity surface shows you who is at 87%, who is at 110%, and who has slack. Before you consider AI takeover, the rebalance suggestions surface first. Sometimes the answer is "Sarah is buried, give three of her tasks to Maria, who is at 60%." The platform proposes it, you approve, it lands atomic.
The Steward agent watches your AI spend the way a comptroller watches yours. Daily collection from Anthropic, OpenAI, GitHub Copilot, Power Automate, anything else you have wired up. Weekly digest into Teams on Monday morning. Spend anomaly above 50% week-over-week? It tells you. A tool that used to be heavy and is now dormant? It tells you. A dollar of measured value ranked above a dollar of self-reported value when it scores cost-per-value.
For every task that might become AI, the Architect agent drafts the analysis: scores, cost-benefit, risk register, governance plan, rollout phases. Then Claude, GPT-5, and Gemini each get the same prompt independently. They vote on four questions. Does the math work. Is this human-sensitive. Is the customer at risk. Does compliance allow it.
Sometimes they agree. The recommendation gate writes do_now with high confidence, and the candidate clicks through to shadow phase. Sometimes they don't. If Gemini flags 11% of the tickets as PHI-touching and the other two missed it, the gate downgrades the recommendation to pilot silently and surfaces Gemini's specific objection in the disagreement notes. You see the math, not the headline.
The platform also rates each candidate on three impact dimensions (cost, service, risk) at none / low / medium / high, with a one-sentence rationale per dimension. A pure-risk candidate (the AI that PHI-redacts every outbound support ticket) wins on its own. A service-only candidate (faster MTTR, no money saved) wins on its own. A candidate that fakes a high score it cannot defend fails the audit invariants and is rejected at the test layer before it ships.
Approve a candidate. The platform provisions the agent itself. A placeholder lands in the org-map with an intentionally narrow tool scope: no Teams, no calendar, no destructive surface, until you wire real credentials. It runs shadow against the human baseline. Each shadow output is persisted with the prompt, the tool calls, the output, and the human decision it was scored against. Agreement rate climbs. Or it doesn't. The AI Supervisor (the new role the platform creates and names) works the disagreement queue. The promotion gate refuses to advance the candidate forward until thirty runs at eighty-five-percent agreement. Backward moves are always allowed: autonomous, assist, off. Restricted actions block until a named approver releases them. The monthly AI spend cap gates every cost-incurring path. Suspended runs surface as a signal on the on-call dashboard, not a silent overspend. Once the agent is live, it works in Teams. With its own name. With its own avatar. As a participant your team recognizes.
Agents are real first-class participants in your tenant. They do not live in a tab nobody opens.
The Steward posts your Monday digest into the team channel you authorize. As a named participant with its own avatar, not a generic notification bot. One-click follow-ups go back into the canvas as a conversation.
The Concierge needs ten minutes to walk you through Pylon. It sends a calendar invite. From itself. On the right person's calendar. With the right context already loaded into the agenda.
People come from Entra. Departments are inferred from groups. The reporting hierarchy is what Active Directory already says. You bootstrap nothing by hand.
Every agent has a stable identifier, a system brief that reflects your industry, a schedule, and a tool scope you authorize. Every run lands in the audit log with prompt, tool calls, and output preserved.
Drafts your morning briefing at 7am. Synthesizes capacity, transformation candidates, and overnight anomalies into a read you can finish in two minutes on your phone. Holds the persona-tuned conversational canvas. Has its own name, its own voice, and a memory scoped to your tenant.
Maps your organization across twelve dimensions, in passes. Pulls people, departments, applications, workstreams. Surfaces the AI already running before proposing any new AI. Never writes directly to the org-map. It proposes, the operator approves, the platform commits. The form your team didn't have to fill out.
Daily spend collection. Weekly digest to Teams. Flags anomalies above 50% week-over-week. Rates measured value above self-reported value when it scores cost-per-value. Posts the worst ratio first. Does not moralize about spend. Spend can be high and right, low and wrong. The Steward shows the math and lets you decide.
Takes one task. Produces a defensible analysis. Cost-benefit, risk register, governance plan, rollout phases. Hands the analysis to the quorum panel for vote. The recommendation gate refuses to surface do_now unless the panel agreed. PHI tasks stay human-only even at perfect recurrence numbers; the Architect does not argue with the compliance veto.
Walks you through every connection in one conversation. Generates the right vendor admin link with the right scopes pre-selected. Probes the credential the moment you paste it. Tells you what it sees. Diagnoses scope shortfalls inline. Skips integrations you say don't apply and stops pestering. The reason your stack is fully connected by lunch on day one.
Plus your own specialist agents, one per department. The agents you build, not the ones the platform ships with. Each gets the same audit treatment, the same approval gates, the same spend cap.
Connectors that claim a capability have to actually implement it. The catalog cannot lie about what data is flowing. "Coming soon" is a chip on the card.
Three connectors. One hour. You'll see the AI you already run before you make any decision about new AI.