Watch
It notices when work arrives
A form submission, a Sunday of check-ins, a deadline, a batch of donations. The agent is already listening; nobody has to open an app to start it.
Hybrid agentic automation for nonprofits
Agents run the operations.
Agents run the operations.
People run the mission.
Reformed AI deploys AI agents inside nonprofit organizations, built on our own hybrid engine: the exact work is done by software, the judgment work by AI, and the final call by a person. Each agent watches a stream of work — check-ins, schedules, giving, reports — and hands your team a decision instead of a task.
20h → 10mweekly admin, first deployment
4agents on the platform
1 approvalper outcome, always human
Trigger
run #0 · sample data
Every agent runs the same loop
Not a chatbot. Not a dashboard. A worker with a narrow job, a clear stopping point, and a person at the end.
Do
It does the routine part completely
Cleans up records, matches names across languages, reads free-text notes, spots trends, drafts the message or the report. Software where the answer is exact, AI where it takes judgment.
Hand off
It stops at the decision
A short list, a draft, a filled schedule with two gaps. A person approves, edits, or declines. Nothing goes out to a member or a donor without that step.
Under every agent: our hybrid engine
Every piece of work passes through four stages. At each one, our engine decides: is this a job for software, which is exact and free, or for AI, which can read and judge but costs more and needs checking? That split is why the agents are fast, affordable, and right.
Software (exact, instant)AI (reads and judges)Highlighted: the path for the selected input
01 · COLLECT
Intake
Spreadsheets, forms, photos, and notes arrive. Each one is sorted by what it is.
Receive and sortRecognize the format
02 · EXACT WORK
Software
Anything with a right answer is done by software. No AI involved, nothing to get wrong.
Clean up recordsCount and comparePut it in tables
03 · JUDGMENT WORK
AI
Only what needs reading, matching, or writing goes to AI, and its answer is checked before it’s used.
Understand names and notesMatch against what we knowTurn text into fields
04 · ANSWER
Hand off
Clean records saved. A short list, a draft, or an answer to a question, ready for a person.
Save the recordsAnswer in plain language
Three questions the router asks
1 · WHAT IS THIS?
2 · EXACT, OR NEEDS JUDGMENT?
3 · IS AI WORTH IT HERE?
Route:
Four agents. One platform.
Pick one to see its run. Attendance is live with a pilot organization; the rest are in development with pilot partners.
Trigger
sample data
Built to be trusted, not just demoed
What every agent shares underneath:
Connections
Forms, spreadsheets, giving platforms, text and KakaoTalk. Everything lands in one place, in one shape.
Hybrid engine
Software for anything with a right answer. AI only for reading, matching, and drafting, and its work is checked before it’s saved.
A person approves
Every message, schedule, or report is a proposal until a named person approves it. Who approved what is recorded.
Everything is traceable
Each run records what it saw, what it did, and what it changed. Any number in a report links back to where it came from.
First deployment: attendance, a church, 20 hours a week
Volunteers took attendance on paper and re-typed it into a spreadsheet during the week. The agent now collects the check-ins, cleans the records, and produces Monday's follow-up list. The same week takes one person under ten minutes.
Before: paper, spreadsheet, volunteers~20 hours / week
After: agent run + one approvalUnder 10 minutes
Your data is the organization's, not ours
Private by default
Each organization gets its own private database, hosted in the US and encrypted. Your data is never used to train anything for anyone else.
No passwords lying around
Our systems talk to your database with keys that expire in minutes. There is no password anywhere to steal.
Leave with everything
Export every record and every run log as a spreadsheet at any time. No lock-in.
Questions we get
Why "hybrid"? Why not just use a model for everything?
Because most of the work has a right answer. Counting absences, building a table, matching two records that share an ID: software does that exactly, instantly, and for free. AI is slower, costs money every time it runs, and can be wrong. So our engine sends each step to software where it can and to AI only where it must, then checks the AI’s work before saving it. That’s what keeps the agents affordable on a nonprofit budget.
What makes this "agentic" rather than just automation?
Automation runs a fixed script. An agent is given a goal and a boundary — "produce Monday's follow-up list from this weekend's check-ins" — and decides the steps: which records need matching, which notes need structuring, what to flag. It also knows when to stop and ask. The boundary is what keeps it safe. The goal is what makes it useful.
Can an agent send a message to a member or donor on its own?
No. Anything outbound is drafted and queued for a named person to approve. The approval and the sent message are both logged.
Why start with churches?
Heavy weekly admin, tiny staff, and volunteers who already know what happened but hate writing it down. If the loop works there, it works for youth programs, community centers, and mutual-aid groups — the same agents with different data.
Do our volunteers need training?
No. They keep doing what they do — tap a roster, fill a form, upload a file. The agent meets the data where it already is.
What does it cost?
Pilot organizations use it free while we build. We set pricing with them, not for them.
Hand the admin to an agent.
We're onboarding a small number of pilot organizations this fall.