AI agents for nonprofits

Most nonprofit AI use today is conversational: you ask an assistant for a draft, it answers, you copy the result into your real work. That works, but it is no longer the whole picture. AI agents can now carry a multi-step task, or a whole small project, from your goal to a finished result, checking in with you along the way.

The sector context makes this worth your attention. The 2026 Nonprofit AI Adoption Report (346 organizations surveyed) found that 92% of nonprofits use AI in some form, but only 7% say it has genuinely expanded what their team can accomplish. The difference between those two groups is rarely the tools: it is moving from one-off prompts to repeatable, delegated workflows. That is exactly the gap agents are built to close.

This guide covers what agents are, the main platforms (they are more alike than their marketing suggests), what agents can realistically do for a nonprofit today, how to run a safe first task, and the guardrails that keep an agent from becoming a new risk.

What an AI agent is (the levels of autonomy)

The simplest way to place agents is as the fourth level of a ladder you are probably already climbing:

• A chat assistant (e.g. ChatGPT, Claude, Gemini) answers one request at a time. You do the steps; it helps with each one.
• A custom assistant (a custom GPT, Gem, or Claude Project) is the same thing preloaded with your context and rules, so every answer starts on-brand.
• An automation (e.g. built with Zapier, Make, or n8n) runs fixed steps when a trigger fires. Reliable, but it only does exactly what you wired.
• An agent works toward a goal you set. It decides its own steps, uses tools (your files, a browser, connected apps), keeps going until the task is done or it needs you, and asks approval for the important actions. Anthropic’s building effective agents guide is a clear technical explanation of the difference, including when an agent is not the right tool.

The practical difference: an automation repeats a workflow you designed; an agent figures out a workflow for a goal you described. Both are useful, for different jobs.

If you have used a deep-research mode (in ChatGPT, Gemini, Perplexity, or most other major assistants), you have already used an agent with one fixed goal type: it plans searches, reads sources, and assembles a report on its own for several minutes.

The main agent platforms

Every major AI vendor now ships agents, and they are converging on the same core idea. It helps to sort them into two families.

Agents for everyday knowledge work, built for documents, research, spreadsheets, files, and connected apps. The most visible options right now include:

• Claude Cowork (Anthropic): reads and writes real files, runs long multi-step tasks and scheduled tasks, and asks approval for sensitive actions. Included in the paid Claude plans; see the getting started guide.
• ChatGPT Work (OpenAI): runs in the cloud, gathers context from your connected tools, produces finished documents, spreadsheets, and slides, supports scheduled tasks, and can stay with a project for hours. Rolling out across paid ChatGPT plans; it builds on the earlier ChatGPT agent mode.
• Browser agents operate websites for you: filling forms, pulling data from pages with no export button, doing web chores. Most major vendors offer one (e.g. Claude for Chrome, or ChatGPT Work’s built-in browser and Chrome extension).

Coding agents, built first for software work but increasingly used for much more (next section). The most used options right now include:

• Claude Code (Anthropic)
• OpenAI Codex (OpenAI)
• Google Antigravity (Google): an agent platform in free public preview for individuals at the time of writing, notable for letting you choose among models from several vendors.
• OpenCode: a widely used open-source coding agent that also works with models from different vendors.

These are the main examples as we write, not the full field: many other agents exist in both families, and new ones launch constantly.

Two more categories to know:

• Agents arriving inside tools you already pay for. CRMs, email platforms, and productivity suites (including Microsoft 365 Copilot and Google Workspace with Gemini) are adding agent and scheduled-task features under their own names, and the pace is fast. Before adding a subscription, check what your existing stack has gained recently.
• Open-source personal agents (e.g. OpenClaw) run on your own devices, for technically confident teams that want full control. Treat these as an advanced path, not a starting point.

*️⃣ Note

Tool mentions here are examples to evaluate for your own needs and budget, not endorsements. This is one of the fastest-moving corners of AI: vendors that look behind today will ship strong agents soon, features move between products, and the best option for your team can change within months. Compare current options against your real tasks before committing, and re-check quarterly.

Coding agents are not just for coders

The “coding” label undersells what coding agents do. Under the hood, a coding agent and a knowledge-work agent are usually the same core system with a different harness: different default permissions, tools, and instructions. Anthropic’s Cowork runs on the same agentic architecture as Claude Code, and OpenAI reports that non-developers (analysts, marketers, operators, researchers) are about 20% of Codex users and its fastest-growing group.

For a nonprofit without developers on staff, a coding agent can still be the right tool when the task is technical-adjacent:

• Website changes: fix a broken page, update text across many pages, or build a new landing page for a campaign.
• Files and data in bulk: clean a messy spreadsheet export, merge donor lists, rename and reorganize hundreds of files, convert formats.
• Documents and assets programmatically: generate a batch of certificates, personalized letters, or image variants from a list, instead of one at a time.
• Troubleshooting: diagnose why a form, embed, or integration stopped working, with the agent reading the actual error rather than you describing it.
• Security and quality reviews: scan your website or a vendor’s code snippet for obvious problems before you ship it.

Coding agents can also answer ordinary questions and do web research; they are not limited to code. The practical rule: start with a knowledge-work agent for document-and-research work, and reach for a coding agent when the task touches a website, a data file, or anything a developer would normally handle. Expect coding agents to assume more technical comfort in how they communicate.

What an agent can do for a nonprofit today

Agents still make mistakes, so every use below assumes a person reviews the output. With that standing rule, realistic current uses include:

Communications and marketing:

• Turn a webinar recording and its slides into a blog post, a clip list, and a social pack in one delegated task.
• Mine a folder of your best past newsletters and appeals into a written voice-and-style guide.
• Monitor your mentions and sector news into a Monday-morning brief, on a schedule.
• Audit your website for broken links, outdated pages, and missing alt text.

Fundraising and grants:

• Research the funders most likely to support a program and deliver a sourced brief on each.
• Assemble a first-draft grant application package from your existing boilerplate, budget, and program documents.
• Track funder deadlines and requirement changes into a live calendar with alerts.

Operations and finance:

• Reconcile and clean spreadsheet exports, flagging anomalies for a person to check.
• Summarize a stack of vendor contracts into a comparison table with renewal dates.
• Draft board-pack sections (finance summary, program dashboard) from your source files on a monthly schedule.

Programs and people:

• Synthesize a folder of session notes or survey exports into themes, quotes, and a report draft (strip identifying details first).
• Compile evidence and best practices for a program design question into a sourced memo.
• Draft volunteer onboarding materials from your existing policies and a role description.

Cross-team:

• Organize a chaotic shared drive: consistent names, folders, an index.
• Prepare meeting briefs from the agenda plus the relevant documents, before every recurring meeting.
• Build simple internal tools (a lookup page, a form, a tracker) that you would never commission from a developer.

This list is a sampler, not a ceiling: there are far more documented uses than fit here. For much bigger idea lists, see Google Cloud’s library of real-world AI use cases (a regularly expanded collection, over 1,300 examples at the time of writing, organized by industry, with a public sector and nonprofits group, and by six agent types), Microsoft’s collection of customer AI stories (1,000+ examples, many of them agents), and our own database of AI use cases for nonprofits (50+ use cases with 300+ examples; not agent-specific, but most multi-step entries in it can be delegated to an agent). The two vendor lists skew to large organizations, but most of the patterns translate directly to a small team.

How to run your first agent task

1. Pick a task that is multi-step, non-sensitive, and easy to check. A sourced research brief, a synthesis of your own published material, or a folder cleanup all qualify. Nothing involving personal data or live accounts.
2. Brief it like a capable new volunteer: the goal, the context, the format you want back, and what to ask you about rather than guess.
3. Limit its access: one folder and the specific tools the task needs. Keep approval required for anything that sends, posts, or deletes.
4. Let it work, then review the result like a contractor’s first deliverable. Check facts, sources, and anything it decided on its own.
5. Improve the brief and run it again. Agents reward a better brief even more than chat assistants do.
6. If it earns a repeat, make it standing: save the brief, schedule it if it recurs, and record the use wherever you track your AI work, with an owner and a review date.

✅ Example

A copyable first brief (adapt the specifics):

You have read access to the folder "board-reports" (our last 8 published
board reports; nothing sensitive).

GOAL: a 2-page "state of our programs" summary a new board member could
read in 10 minutes.
1. Read everything in the folder.
2. Identify what changed over time: programs, reach, funding mix.
3. Draft the summary with every number traced to its source report.
4. List anything unclear or contradictory as open questions for me.
Ask me before reading anything outside that folder. Show me an outline
before writing the full draft.

Extend an agent with skills, plugins, and connectors

Out of the box, an agent is a generalist. Three kinds of add-ons make it a specialist:

• Skills: reusable instruction packs that teach an agent a specific job.
• Connectors: links that let the agent reach your other tools and data, increasingly built on MCP, an open standard most major AI platforms now support.
• Plugins: bundles of skills and connectors installed in one click. Most major agent platforms now have a plugin system and directory of their own.

Our companion guide, Agent skills and plugins, explains all three in depth, including how MCP works, how it compares to the alternatives, where to find vetted add-ons for each platform, and how to install them safely.

The guardrails (read before your first real task)

⚠️ Warning

An agent is more capable than a chatbot precisely because it can act: open files, browse, and use your connected tools. That also means new risks. The two to know by name: prompt injection (a malicious web page or document contains hidden instructions the agent might follow) and unvetted add-ons (a skill or plugin written by a stranger is instructions you are choosing to run). Neither is a reason to avoid agents; both are reasons to follow the rules below.

Five rules cover most of the risk:

1. Give the agent only what it needs. Point it at one folder and the specific tools for the task, never your whole drive by default.
2. Approve anything that sends, posts, pays, or deletes. Keep the agent in “ask first” mode for outward-facing and destructive actions.
3. Start read-only. For the first weeks, use the agent to research, summarize, and draft, not to act on live systems.
4. Install skills and plugins only from vetted sources, and read what they do first. See Agent skills and plugins for how.
5. Keep the activity log on and skim it. Most agent products log every action; that log is your audit trail and your early-warning system.

Your existing rules still apply on top: the same data rule (no donor or beneficiary personal data in tools without proper terms) and human review before anything publishes. Treat any pilot that gives an agent access to real accounts, files, or a browser as higher-stakes by default, and run it alongside your current process before switching over. Our AI risk database for nonprofits and new AI pilots checklist both apply to agents unchanged.

If leadership or the board asks about oversight, these five questions (adapted from a good governance-first explainer for nonprofit leaders) are the ones to have answers for:

• Who defines the goals and boundaries each agent operates within?
• Which actions require human approval, and which run automatically?
• How are the agent’s actions logged and reviewed, and by whom?
• What data can each agent reach, and how is sensitive information kept out?
• Who monitors for errors and bias, and who can switch an agent off?

What it costs

Prices change too fast to quote, but the pattern is stable enough to plan around:

• Included in paid assistant plans: the main knowledge-work agents ship inside the paid tiers of the assistants many teams already have, so trying one often costs nothing extra.
• Nonprofit discounts exist: check Claude for Nonprofits, OpenAI for Nonprofits, and the nonprofit programs of Google and Microsoft before paying full price.
• Free previews and open source: some agent platforms run free previews (e.g. Google Antigravity for individuals at the time of writing), and open-source agents cost nothing beyond the AI model usage they consume.
• The real cost is review time. Budget staff time to check agent output, especially early. An agent that saves three hours but needs two hours of fixing is barely worth it; measure both sides before scaling.

Useful resources

• Claude Cowork: getting started: setup, permissions, and scheduled tasks; the product guide adds seven worked example workflows.
• ChatGPT Work: OpenAI’s agent for everyday work, with connected tools and scheduled tasks.
• Google Antigravity and OpenCode: an agent platform in free preview and a widely used open-source agent, both of which let you choose among AI models from several vendors.
• Agent skills and plugins: our companion guide to extending an agent safely, whichever platform you use.
• Google Cloud: real-world AI use cases: the biggest well-curated idea list we know (1,300+ examples at the time of writing), browsable by industry and agent type; skim the public sector and nonprofits group first.
• Agentic AI for nonprofits: a practical explainer for leaders: the governance-first read for your director or board.
• Anthropic: building effective agents: a clear explanation of what agents are and when not to use one.

📋 Summary

• An AI agent works toward a goal you set, deciding its own steps across tools, while an automation only repeats steps you wired in advance.
• If you have used a deep-research mode, you have already used an agent with one fixed goal type.
• Every major AI vendor now ships agents, in two families: knowledge-work agents (e.g. Claude Cowork, ChatGPT Work) and coding agents (e.g. Claude Code, OpenAI Codex, Google Antigravity, OpenCode, among many others), and the landscape shifts fast enough that no vendor or shortlist is a safe default.
• Coding agents are useful to non-coders too: website fixes, bulk file and data work, batch document generation, and troubleshooting are all agent tasks, no programming required.
• The best first agent tasks are multi-step, non-sensitive, and easy to check: research briefs, syntheses of your own material, file organization.
• An agent’s power comes from acting, so limit what it can touch, approve outward-facing actions, and start read-only.
• Prompt injection and unvetted add-ons are the two agent-specific risks to know by name; permissions, vetted sources, and activity logs are the countermeasures.
• Review an agent’s work like a contractor’s first deliverable, then improve the brief: agents reward a better brief even more than chatbots do.
• Trying an agent often costs nothing extra (they ship inside paid assistant plans), but budget real staff time for reviewing output before you scale anything.
• Agent features are arriving inside tools you already pay for, so check your existing stack before buying anything new.