CRM Automation & Donor Management
How shluchim use AI to enter donations, log contacts from voice notes, and build follow-up lists in their CRM, with the guardrails that kept it from going wrong.
On this page
When the group was asked in March 2026 what an AI agent should do for them, the top answers were about the CRM: a short daily list of lapsed donors to call, contacts updated from forwarded emails and voice notes, and follow-ups put on the calendar. By October 2026 several shluchim had working versions. This page covers how they did it. For the systems themselves, see CRM Tools.
Four ways AI reaches a CRM
1. A connector (MCP) on a real CRM. The cleanest route. Salesforce, Attio and Tally have official MCP servers, and newer shluchim-built CRMs (DAAS, MYS) come with their own AI. Claude or ChatGPT then reads and updates records through the vendor's own rules. See MCP Connectors.
2. An agent with written boundaries. A shliach on Salesforce NPSP (August 2026) first had Claude review his setup and data, then wrote a skill file for his GabAI agent spelling out which objects it may read and write and what is off limits. He reports it asks before updates and bulk changes. It runs reports, updates contacts, creates pledges, gifts and payment plans, and lists birthdays for the week. One lesson from him: it got birthdays wrong around shkiah because Hebrew birthdays were not stored.
3. A browser agent. For CRMs with no usable API, Claude Cowork, Claude in Chrome, ChatGPT's agent mode or a personal agent clicks through the web screens like an assistant would. It works but is slow, and every site redesign breaks it. Give it its own login with limited rights where the CRM allows it. One shliach used Cowork to enter Zelle donations and had it keep notes so it would not repeat mistakes (August 2026).
4. Import, not direct entry. The option with the least that can go wrong: the AI prepares a spreadsheet and you import it after checking. In October 2026, worried about "AI messing with the real database," members on CMS Cloud used its Upload Donations feature for exactly this.
Workflows that are running
Payments from email into the CRM
Zelle, Venmo, PayPal and CashApp send notification emails. Several shluchim had AI write a Google Apps Script that reads those emails into a Google Sheet, then import the sheet or have an agent enter the gifts. The full recipe is in Bookkeeping Agents. In October 2026 another member had Muse enter all donations into ChabadSuite (Zelle, Venmo, CashApp, event registrations, photos of checks) from a dedicated forwarding address, and said it saved "hours and hours" over Tishrei.
Voice note to CRM
A campus shliach records a voice note after each conversation; his agent transcribes it and logs the interaction in the COCi system through its API. Changes wait 24 hours on a review list before they go in. He reports saving several hours a week (August 2026). Hecher's WhatsApp helper does the same for its own CRM.
Weekly report
A shliach on Salesforce NPSP shared standing instructions (October 2026) for a weekly report: a trigger phrase and date window, data from NPSP plus Google Calendar to fill gaps, sections for personal touches, events, milestones, learning, gifts and new contacts, and a check-in list of people who are drifting ("3+ touches in prior 8 weeks, none in last 3"). It outputs a Word document. Instructions like this, saved in a Claude Project or a skill, are reusable by anyone with a similar CRM.
Lapsed donors
A CMS Cloud user asked in January 2026 for AI to surface two lapsed donors a day. The replies: most CRMs can run that report without AI, and if yours cannot, export names and yearly totals and ask an AI to analyze them. CMS Cloud's AI Report Assistant (August 2026) now builds such transaction reports from a plain-language request.
Guardrails that came out of mistakes
- Draft and review first. A member using the LGL API in March 2026 found it pushed records straight in, sometimes wrong, with no preview. Run new automations in draft mode, or with a review list, before letting them write.
- Start read-only. Members setting up Salesforce agents were advised to begin with a read-only user.
- Expect lockouts. In September 2026 a member warned that Salesforce locked him out after agent use. Agents also fail on passkey and security-key prompts.
- Unknown donors. Zelle notices often lack a full name. Have the AI ask you rather than guess.
- Auto-filled fields. An agent-made event invoice put the wrong spouse's name on it (April 2026). Review personal fields.
- Personalized cold emails. A vendor's AI-written emails to shluchim, with a paragraph about the recipient's family, drew "it's obviously AI" and "turns me off" (August 2026).
Building your own CRM
In July 2026 a group admin said he was building his own. The strongest counter-view came from a technical member in June: AI-built CRMs and accounting systems put data integrity, audit trails and irreversible changes at risk, and "integrating into and out of a real system" is the safer goal. A shliach who tried it said it was "very difficult" even after hiring a developer. See Vibe Coding.
Still open
- A safe way to push donations from DonorBox or a payment processor into ChabadSuite (asked August 2026, no answer).
- A communication-preference page that pauses outreach for people who are traveling (an October 2026 idea).
- Whether the reported ChabadOne AI integration will allow data entry.