AI-assisted nonprofit decisions: connect each recommendation to purpose and control

Start with the charity's purpose and decision authority

Write the decision in terms of the organization's governing purpose, the available funds and the body authorized to decide. “Choose the strongest application” leaves the criteria undefined. A better exercise question is: “Which proposals, if any, should the fictional board advance to due diligence under the stated community-learning fund criteria?” Record conflicts, delegations, quorum or approval requirements supplied in the case materials before asking AI to compare proposals.

The Charity Commission's article on charities and artificial intelligence says trustees remain responsible for decisions and describes the need for independent checks and human oversight when AI is used. Its decision-making guidance for charity trustees sets out principles including acting within powers, being sufficiently informed, considering relevant factors and managing conflicts. These sources apply to charities in England and Wales. They inform this fictional review but do not determine the duties, governing rules or right outcome for another organization or jurisdiction.

Sources: [1] [2]

Build a purpose-and-control record

Imagine a synthetic nonprofit with £60,000 available for community learning projects. Three short applications, a fund note and trustee declarations are provided. Project Cedar requests £35,000 for workshops and £8,000 for unrelated building repairs. Project Harbor requests £25,000 and includes a delivery partner chaired by one trustee's sibling. Project Lantern requests £30,000 but its participant estimate appears in a marketing slide without a calculation. An AI comparison ranks Harbor first because its narrative uses the most confident outcome language.

Create one record with a row for each proposal and columns for stated activity, connection to the fund purpose, requested amount, restricted or excluded cost, evidence supplied, missing evidence, conflict or relationship, proposed control, decision status and owner. Link every entry to the application, fund note or declaration that supports it. Put inference in a separate field. The record should not convert persuasive wording into evidence of need or delivery capacity.

For Cedar, separate the workshop request from the building repair and check each against the fictional fund note. For Harbor, record the relationship without assuming wrongdoing, then route it through the case's conflict process. For Lantern, preserve the unsupported participant estimate as a question rather than filling in a plausible calculation. AI may draft the table, but the worker checks amounts, conditions and source locations and removes any invented scoring rationale.

Use a decision pause test

Apply a pause test to each proposal: what unresolved fact would make it premature for the authorized body to decide? Cedar may need a revised budget that removes or separately justifies the repair. Harbor may need a recorded conflict disposition and an independent comparison. Lantern may need the assumptions behind its participant estimate. The purpose of the pause is not to reject an application; it is to distinguish a decision-ready record from an attractive narrative.

Next, change one fact. Suppose the synthetic fund note says unused money must be returned and cannot support capital costs. Does the recommendation update Cedar's treatment and the available total, or does it preserve the earlier ranking? Suppose the related trustee leaves the meeting. Does the record show how the remaining decision was authorized under the supplied procedure, rather than merely deleting the conflict field? A useful final brief gives the board options, unresolved issues and accountable next steps without pretending the AI chose the charitable outcome.

Preserve governance and assessment boundaries

Real nonprofit work must follow the organization's governing document, fund restrictions, applicable law, safeguarding and financial controls, with qualified advice where needed. This synthetic method does not establish eligibility, trustee compliance, funding impact or fraud risk. Two Prune could observe whether a participant connects evidence to purpose, keeps conflicts and restrictions visible and communicates a calibrated recommendation. Its report would describe performance within a bounded assessment for human interpretation, not direct charitable funds or automate a consequential decision.