Synthetic nonprofit/foundation case — A2 agent for grant dossiers

Fictional organization. Fondation Lien Local, its programs, applicants, volumes, costs, and results are synthetic. The case illustrates a decision method; it does not prove the performance of any tool, supplier, sector, or real organization.

Measured level: administrative A2 business agent. The system carries a standard case to a reviewer-ready packet and performs external effects after approval. It does not score merit, rank needs, recommend an amount, or decide an award or refusal.

1. Mission and starting point

Fondation Lien Local is a fictional 14-person Swiss foundation. It supports community projects through CHF 5,000–25,000 micro-grants. Twenty-four volunteer reviewers and a seven-person committee examine about 720 applications per year in French, German, and Italian. Telephone and paper channels remain available.

A baseline across two calls covers 241 applications. For a standard case that is ultimately sent to reviewers, the secretariat spends a median 96 minutes of active human work: check consent and program, inventory documents, extract budget and schedule, verify administrative completeness, pseudonymize the review copy, request missing information, record the case, and assign two reviewers.

The pilot does not try to accelerate funding judgment. It addresses a narrower problem:

carry an administratively standard case from receipt to two authorized reviewers without turning AI into a judge of the mission, people, or funding.

2. Mandate and population

“Admitted to the workflow” only means that a case can follow the standard administrative path. It does not mean eligible for funding, recommended, or prioritized.

The A2 workflow accepts only a known program, recorded consent, identified applicant, readable documents, and defined document categories. It stops for an offline application, assistance need, new program, ambiguous funder restriction, unexpected highly sensitive data, conflicting identity, inaccessible document, conflict of interest, or any situation requiring mission interpretation.

3. Boundary between administration and judgment

The agent produces no risk, vulnerability, “quality,” or funding-probability score. Reviewers see source documents and work without a prefilled recommendation. Pseudonymization reduces some visible information; it does not claim to remove every social or cultural signal.

4. Workflow and permissions

The agent performs seven steps:

1. check consent, program, channel, and workflow-entry rules;
2. inventory documents without copying unnecessary data;
3. extract administrative facts with page-level source citations;
4. apply the deterministic completeness checklist;
5. prepare a document request or pseudonymized reviewer packet;
6. show sources, transformations, recipients, and effects to the approver;
7. after approval, send, write to the grant system, assign two reviewers with no known conflict, and read back effect identifiers.

Raw documents remain in the authorized system. The model receives only needed extracts. Every connector uses a separate identity. Writes use an idempotency key; a timeout triggers a read-back before retry. The kill switch blocks messages and writes without removing the manual path.

5. Planning range

The range applies to administrative time only for cases that can follow the standard workflow:

hours released / month = 60 applications × workflow share × 96 min × reduction / 60

With CHF 12,000 of setup and CHF 750 in recurring monthly cost, simple theoretical payback ranges from 4.6 to 16.6 months after go-live; the central case is 6.9 months. The high bound is not a mission target.

These figures are theoretical administrative capacity—not another donation or social impact. Released time creates value only when actually reassigned to applicant support, decision quality, or shorter applicant delays.

6. Gates before live use

The 60 frozen cases cover three languages, small and large organizations, portal and scans, multi-sheet budgets, missing files, unreadable documents, conflicting identity, minors, unexpected health data, conflicts of interest, an injected instruction in an attachment, unavailable services, post-write timeout, and replay of the same case.

One suggested funding decision, data leak, channel exclusion, unauthorized action, or untraceable transformation stops the pilot regardless of time gain.

7. Sixty-day pilot

• Days 1–10 — Map and minimize. Test rules, documents, consent, languages, channels, roles, retention, pseudonymization, and fallback.
• Days 11–20 — Replay and contradict. Run all 60 frozen cases after any model, instruction, rule, language, source, or connector change.
• Days 21–30 — Shadow mode. The secretariat prepares the real packet before seeing the proposal. No applicant or reviewer receives an effect.
• Days 31–60 — Live A2. A manager approves the message, packet, and two recipients; reviewers and committee then work without an agent score or recommendation.

8. Synthetic results

Across 120 inbound applications:

• 86 enter the standard administrative workflow;
• 34 remain outside the workflow and receive human service, without being treated as refused or “bad” applications;
• 58 packets are reviewer-ready without correction;
• 21 require a minor correction and are then accepted;
• 7 stop and escalate before effect;
• no decision, amount recommendation, or priority is generated;
• no offline channel is removed and no action occurs without approval.

The 79 accepted packets are 65.8% of all 120 applications. Assuming the same reference time across the portfolio, 65.8% × 59.4% gives a weighted ceiling of about 39% less administrative time across all applications—not 59% of the foundation's entire activity.

The pilot releases about 75 hours over 60 days, or 37.5 hours per month. At CHF 62/h and after CHF 750 in recurring cost, net monthly capacity is about CHF 1,575, for a simple payback near 7.6 months. No additional social impact or funding is automatically attributed to the agent.

9. Mission gate

Results are segmented by language, channel, and organization size. The pilot does not show a gap above the preregistered major-rework threshold. Two applicants use the telephone channel; their cases remain outside the agent and advance to the same human gate. No AI-treatment complaint is logged in this small sample; that does not prove the absence of a trust effect.

Decision: keep A2 for standard administration for 90 days; prohibit every automated funding decision or recommendation.

The system reduces real work without crossing the mission boundary. Any extension requires a new mandate, consultation with affected people, a larger sample by language and channel, a tested challenge mechanism, and a decision independent from the administrative success.

10. External anchors and limits

The note AI and grantmaking: external anchors and limits separates sector survey, functional analogues, provider story, Swiss law, and evaluation guidance. Sources make the workflow plausible; they validate neither the figures nor the boundary chosen for Fondation Lien Local.

11. Evidence packet to hand over

The handoff contains the mandate, mission guardian, consulted groups, baseline and denominators, workflow-entry rules, data and consent inventory, permission matrix, funder register, threat model, frozen set, results by language and channel, ledger for all 120 applications, pseudonymization transformations, approvals, corrections, escalations, recipients and read-back effects, complaints, costs, gate decision, fallback, and review date.