AI agents

Find it: Left nav → Agents (what agents did + your review queue) · Settings → AI agents (the controls) · Who can use it: Anyone with donor-write access sees the Agents page and review queue. Only Org Admins / CDOs set the autonomy dial and the kill switch.

Read this first — the guarantees

The AI agent layer is designed so you can turn it on slowly and never be surprised. Before anything else, know these five hard rules. They are not settings you can weaken — they are built into the platform.

1. Nothing sends on its own until you opt in, per task type. Every task type starts at Prepare — agents draft, you approve. An agent only completes something end-to-end after you deliberately switch that one task type to Auto, and even then only where the task type allows it (see below).
2. Agents can never override ask-timing. An agent cannot complete or schedule an ask that the OK-to-ask engine blocks. There is no override path for an agent — the human-with-a-justification override exists for people, never for agents. Asks are always drafted at most, never auto-sent.
3. Consent and deceased status are always honored. Do-not-contact, do-not-solicit, unsubscribes, and deceased suppression apply to agents exactly as they do to staff. An agent will not touch a donor it isn't allowed to touch.
4. Every agent action is logged with its reasoning in plain English. Not just "did X" — why: e.g. "Sent monthly impact note to 214 Tier E donors — automated stewardship touches are permitted for low-tier donors." You can read exactly what your agents did and why, any time.
5. One switch pauses everything, instantly. The kill switch on the Settings → AI agents page stops all agent activity for your organization immediately. (ThankFirst also maintains a platform-wide kill switch as a backstop.)

The design goal: an admin can go on vacation for two weeks and come back to find low-tier stewardship current, receipts issued, tasks triaged, and a prioritized human to-do list waiting — with zero methodology violations in the log.

The autonomy dial (Settings → AI agents)

You control agents with one dial per task type, with four levels:

Everything starts at Prepare. That means the very first thing agents do is make your review queue useful — never act unattended — until you decide otherwise.

What can and can't be automated

The dial reflects each task type's execution class — how much a machine is ever allowed to do with it:

• Can reach Auto (routine, low-risk): low-tier (Tier D/E) thank-yous, receipt delivery (transactional, not a stewardship touch), creating reminder / follow-up tasks, next-action-date maintenance, stale-task reassignment, record hygiene (duplicate flags, missing-field nudges), and applying donor self-service changes you've already approved.
• Prepare-only (drafted for you, never auto): cultivation and follow-up to mid and major donors — any meaningful donor-facing message to someone who matters gets a human's eyes.
• Human-only (agents may add prep notes, never complete): calls, visits, and every ask. Agents can attach donor context and suggested talking points to these, but a person always does them.

This mirrors the methodology's own automation tolerance: mass, low-tier touches are fine to automate; the relationships that carry your major gifts are done by people.

The Agents page — what agents did, what's waiting

Two sections:

• Waiting for you — donor changes to review: when a donor updates their own details in the Donor portal, the change lands here instead of silently overwriting your record. Approve or reject each one; approving is what actually updates the donor.
• Recent agent activity: a running log of what agents have done, each with an outcome — done, escalated, refused (it declined because a rule said no), or failed — and the plain-English reasoning behind it.

What lands in your inbox

Agents also produce three recurring summaries, emailed to the recipients you choose in Settings → AI agents (org admins by default):

• Daily digest — "what your agents did / what's waiting for you." A plain, factual recap so you always know the state of play.
• Weekly briefing — an AI-written, methodology-driven strategy note ("thank first, steward second, ask last") highlighting who needs attention this week.
• Monthly board packet — a board-ready summary (the full printable Board Packet lives in the app).

Failure discipline

A task an agent can't complete is never silently dropped and never silently retried forever. It stays open with an error note and escalates to a human. Agent runs are idempotent — if a run crashes and restarts, it won't send anything twice. Agent activity also counts against your plan's AI generation limits, just like manual AI drafting.

How to turn agents on gradually

1. Go to Settings → AI agents. Everything is at Prepare — leave it there at first.
2. Work your review queue for a week. See what agents draft; approve what's good.
3. Once you trust a routine, low-risk task type (say, low-tier thank-yous or receipt delivery), switch just that one to Auto and confirm.
4. Keep calls, visits, and asks human. Let agents prepare them, but do them yourself.
5. If anything ever feels off, hit the kill switch — it stops everything at once, and you can switch any task type back to Prepare whenever you like.

Good to know

• Agents are premium. The agent layer is part of the AI Development Associate, available on Growth and Enterprise plans. See Billing.
• Prepare is safe by default. A brand-new org has agents drafting-for-approval only — you have to choose Auto, per task type, for anything to send unattended.
• The log is the proof. Because every action carries its reasoning, the audit trail answers "why did my donor get this?" for you and your board. See the audit log.
• Related: the on-demand AI drafting inside a donor's profile and Communications is the same assistant working one message at a time; this page is about it working your whole queue.