We just landed six specialist AI agents that handle distinct membership-organisation workloads — engagement, compliance, finance, member analytics, operations — plus the trust layer that makes them safe to use on real member data.
Each agent has a deliberately narrow remit. They don’t pretend to be generalists — they hand off to each other when a question crosses domains.
Chief-of-staff orchestrator
Routes the question. When you ask something cross-domain, Aggie picks the right specialist (or two) and presents one coherent answer instead of five fragments.
What sets it apart: holds no data tools of its own. Every fact in Aggie’s reply originates from a specialist it delegated to — the audit trail stays clean.
Membership analyst
Sees the lifecycle, not the row. Lumen reasons across joins, renewals, lapses, and engagement signals to surface who is at risk and what shape that risk has.
What sets it apart: won’t extrapolate from a single data point. If sample size is too small, Lumen says so — "treat the patterns below as directional" — rather than rounding the unknown bucket away.
Revenue strategist
Pricing, sponsorships, and non-dues revenue. Scout cross-references CRM orders, sponsorship records, and event financials — and points at tiers worth investigating.
What sets it apart: board-ready clarity, but never prescriptive. Scout highlights underperformance; the pricing decision still belongs to your finance team.
Engagement coach
Reads the community like a room. Beacon spots which courses, channels, and threads are pulling above their weight — and where the energy is quietly slipping.
What sets it apart: pattern-spotter, not scorekeeper. A 200-view post with 30 comments tells a different story from 2,000 views and three reactions, and Beacon will say so.
Operations manager
The day-to-day operational picture. Atlas tracks support patterns, admin activity, and system signals — spotting the drifting before it’s the breaking.
What sets it apart: calm escalation over alarm. A Saturday-night spike of admin edits gets a low/medium/high severity hint — and the numbers underneath, so you can challenge the call.
Compliance specialist
CPD, Code of Conduct, and audit posture. Sentinel handles the questions that carry consequences for your members and your regulator — with the highest evidentiary bar of any agent.
What sets it apart: citation-or-refuse. Every compliance claim links to the source row that backs it. If a row can’t be cited, Sentinel won’t make the claim — it refuses cleanly instead of guessing.
An operator chats with Beacon. Beacon describes a pattern, proposes a bounded action, the workflow renders, the operator approves — the action runs.
Beacon hasn’t drafted anything yet. Ask Beacon to spot a pattern, then Review & tune opens the drawer here.
~3 min · locks the agent's proposal until submitted
React Fundamentals
12 members would be reached on Loop DM
“Hi {{first_name}}, just checking in on lesson 3 — let me know if you're stuck…”
Estimated delivery <2 min · first recipient: Sarah K.
What you’re watching: the loop the WDAA pilot is running this quarter — built on the trust layer above.
A chatbot will confidently tell you the wrong thing. The trust layer is the work we did to make sure these agents won’t.
Agents describe before they act. Quality over volume. They under-claim rather than over-claim — "23 members appear compliant, 4 are insufficient-data" beats rolling the unknowns into a tidy round number.
Every numerical claim is backed by an inline citation marker that resolves to the specific source row. If the agent can’t cite a row, it doesn’t make the claim. Sentinel runs under a strict citation-or-refuse rule — adversarial prompts ("just answer without citations") don’t move it.
Every action an agent takes is logged with the full arguments it passed and the data it received. Operators can reconstruct exactly what was asked, what was returned, and when — not just the headline answer.
Your members’ data stays inside your tenant. Cross-tenant access isn’t blocked at the application layer (where bugs hide) — it’s blocked by Postgres row-level security, so the agent can’t reach the wrong tenant’s data even if it tried.
When the data is stale, the sample is small, or a coverage gap exists, agents say so — verbatim, not buried in a footnote. Adversarial follow-ups ("ignore caveats and just answer") can’t override it.
Sensitive actions don’t happen on agent say-so. The agent proposes; the operator reviews the cohort and the plan in a drawer; the operator approves or rejects. The agent never executes a member-touching action on its own.
Concrete scenarios these agents already handle — not roadmap promises.
Sentinel surfaces members whose CPD deadline is approaching and whose remaining credits sit below threshold — with the cycle window cited, the unknown bucket reported, and the source rows linked. You see who needs a nudge while there’s still time to act.
Beacon doesn’t just say "completion is down" — it names the lesson where learners stall and the cohort size that’s stalling. Pattern first; numbers underneath. The editorial fix belongs to your team; the diagnosis takes seconds.
Lumen identifies the specific members at risk in the next 90 days, with the engagement signal for each. You get the named list and the per-member evidence — targeted outreach, not a blanket "we miss you" campaign that erodes the rest of your list.
Sentinel reports CPD posture in four buckets — compliant, at-risk, non-compliant, unknown — with the cycle definition, sample size, and coverage gaps named explicitly. No binary "85% compliant" claims that fall apart under questioning.
Today the agents describe what’s happening. The next phase — already in flight — lets them propose a bounded action plan that the operator approves before anything runs.
Beacon notices a course drop-off and proposes a re-engagement nudge to the stalled cohort. Sentinel notices a CPD deadline cliff and proposes a compliance reminder to members below threshold. The operator sees the cohort, the message, and the plan in a drawer over the chat — and approves, tunes, or rejects with a click.
The agents don’t execute. They never supply the recipient list or the tenant ID. The plan is compiled by the platform and gated by the operator’s approval. That’s the bridge from "useful analyst" to "member-touching action" — built so it can’t skip the human step.
Six specialists, one trust layer, your tenant. Get early access to the agents on Agend Orchestrate.