Quick Answers

Choose based on the workflow and who will maintain it.

n8n fits when the workflow has several steps and AI handles one or more of them, such as classifying, drafting, or transforming data between systems. It can be self-hosted, which helps with data-sensitive work, and is the most flexible of the visual tools for AI-heavy orchestration. The tradeoff is a steeper learning curve.

Make fits when you need to connect familiar SaaS tools quickly — forms to spreadsheets to email to Slack — with modest logic in between. Its visual builder and connector library make common integrations accessible to non-developers.

Apps Script fits when everything already lives in Google Workspace. It is free, runs inside your Google account, and is useful lightweight glue: processing Gmail, updating Sheets, generating Docs. It requires writing some JavaScript, though AI assistants make that more approachable than it used to be.

Custom code fits when logic or integrations outgrow no-code — complex business rules, unusual APIs, performance needs, or workflows you will maintain for years. You take on hosting and maintenance, so a developer, staff member, or reliable volunteer needs to own it.

A practical tiebreaker: who will maintain this in two years? Pick the option that person can troubleshoot. Prototype before committing; an afternoon in each tool tells you more than any comparison page.

Permalink

All three run the same Claude models. What changes is how much of your computer each one can reach.

Claude Chat is the familiar chatbot. It answers questions and holds up well for back-and-forth strategic thinking. It is also the least capable of the three.

Cowork sits in the middle, and it is where we suggest starting. It is the same chatbot, except you give it a folder on your computer to work out of. It reads the files in that folder to find what it needs, and it can create new files there or change existing ones.

Claude Code does much more than its name suggests. It has far fewer restrictions, so it can take almost any action on your computer, usually asking permission first. It can open any folder, read or change any file, write small pieces of software, and put a working website online. That freedom carries two costs. It can delete a file or break a running site by accident, though this has become rare. It also has the least intuitive interface of the three, which some people find intimidating to look at.

There is one reason to prefer Chat. It uses up your weekly Claude allowance more slowly for the same task. Cowork and Code can both see folders on your computer, and they search through them on their own initiative, sometimes spending your allowance on irrelevant files. Chat does not sit in a folder, so it never does this.

If you regularly run out of Claude before your usage resets, send simple questions to Chat. Otherwise use Cowork or Code for everything.

Go deeper: Claude · What is an AI agent?

Permalink

Deep research is a mode in tools like Claude, ChatGPT, and Gemini. You give it a question, and the AI searches the web and any connected sources you allow, opens relevant pages, compares evidence, and writes a report with citations and a source list. Depending on the tool and question, the report often runs several pages and may take a few minutes or longer.

A high reasoning setting gives the model more effort to analyze the context already available. Deep research adds a multi-step workflow for finding and comparing sources before it answers. Some tools let you choose a stronger reasoning model inside deep research. The research workflow and reasoning level are separate choices.

For advocacy nonprofits, it is well suited to landscape work: scanning corporate welfare commitments in a region, mapping which organizations work on an issue, summarizing the state of research on a question, or gathering background before a campaign or grant application.

For a broad question or a decision your organization will act on, try multiple reports with different tools or prompts. Ask AI to combine them, carry forward the best-supported findings, and flag disagreements or missing areas that deserve another search. Check the underlying sources before accepting the combined answer; multiple reports can repeat the same weak claim.

Output quality depends on the available sources, so niche or paywalled topics may produce thinner reports. Verify critical citations before you rely on or republish a report. Feature names and availability change quickly; check the current documentation for whichever assistant your organization uses.

Go deeper: Picking the right AI model and effort level · Claude overview

Permalink

Start with policies the movement has already published. The four below show a useful range of choices, from values to operating rules. They also span a real range of risk appetite, from cautious to bold.

Four policies from the movement

Stray Dog Institute (full policy PDF) — the values-led model. Each tool category gets a plain discussion of the considerations, then an explicit “our commitment” statement. It has the movement’s clearest position on photorealistic AI imagery: where an image serves as documentary evidence of what animals endure, it should be real — AI-generated substitutes can erode the public trust that investigators took great personal risk to earn. As a funder, SDI also commits to never weighing a grantee’s openness to AI in funding decisions.

Faunalytics — the tool-by-tool model. Organized by category (language models, image generators, machine learning, crawlers), each with a short explainer of how the technology works before the rules. Notable calls: internal drafting is broadly fine while external AI use is narrow and always labeled, and they deliberately allow AI crawlers to index their research, using language models as another channel for reaching people.

Animal Charity Evaluators — the operational model. The most rule-like of the four: two confidentiality tiers (sensitive material requires training-disabled modes; strictly confidential material never enters AI tools), connectors prohibited unless approved, consent required before AI note-takers join calls with external participants, and a commitment that belongs in any staff conversation: AI gains go to increased output for the mission, not to cutting staff.

Bryant Research — the high-trust model, and the boldest of the four. This research consultancy for the animal protection movement names its own bet directly: an overly restrictive AI policy threatens the organization as much as an overly permissive one. Staff are expected to try AI in every part of their work, and no manager sign-off is needed for any tool except photorealistic AI images, video, or audio in client-facing work. Client communications and survey or interview data are the only hard no for consumer AI tools. Most movement organizations will want more guardrails than this. Read it as the far end of the range.

Read all four and notice which choices fit your organization and which you would adapt. Those reactions are the start of your own policy.

What they share

Across these and the broader nonprofit policies we have reviewed, the same building blocks recur:

• An approved-tools list — which tools, on which accounts (organizational, not personal), and who approves additions.
• Data rules — a simple sensitivity classification, with supporter records, donor information, personnel matters, and anything legally sensitive in the “never paste into consumer tools” tier.
• Human review — anything published or high-stakes gets human review; accountability stays with the person, not the tool.
• Disclosure norms — when to say AI was involved, especially for public content and imagery.
• A named owner and review date — every policy above states who maintains it and how often it is revisited, because tools and terms change fast.

General-sector templates

For a document you can adapt, these options are free and current:

• NTEN’s AI resource hub — includes two free, directly downloadable policy templates, one from NTEN and one from ANB Advisory Group.
• Charity Excellence — a free UK-flavored template for small organizations; one of the few that addresses environmental impact.
• AI Impact Hub — Kyle Behrend’s nonprofit AI policy template and maturity model, designed to be customized in an afternoon (free, requires an email address).

Use each example as source material. Another organization’s policy reflects its risk tolerance, tools, and data. Write the rules that fit your own work for animals.

Go deeper: Create your AI policy · Common concerns about AI

Permalink

An AI agent is an AI system that can take a series of actions to finish a task. A chatbot replies with text. An agent can use tools: read and edit files, search the web, query a spreadsheet, call other software, and check its own work along the way.

Agentic systems sit on a spectrum. Plain chat answers your question. Tool-using assistants may search the web before responding. More autonomous agents can take a goal (“review these 40 applications against our criteria and draft a summary”) and work through many steps, deciding what to do next at each stage.

As of mid-2026, many staff will encounter agents through general-purpose tools such as Claude Cowork, Claude Code for technical work, and ChatGPT agent mode. These tools can work across files, websites, and connected apps. Automation platforms such as n8n, Make, and Zapier also let teams build agents for repeatable workflows.

That autonomy makes agents useful for nonprofit work such as triaging an inbox, monitoring broken links, and drafting routine emails. It also makes review checkpoints essential. The more steps an agent takes without you, the more a small early mistake can compound. As autonomy increases, place deliberate human checks before anything is sent, published, or deleted.

Capabilities change quickly; what counts as “agentic” this year was experimental last year. Start with low-stakes, reversible tasks and give the agent more autonomy as you build trust.

Go deeper: Data security and privacy basics · Common concerns about AI

Permalink

Nonprofit discounts can put major AI tools within reach of a small organization. Prices change, so check current rates before you budget.

Start with nonprofit pricing. The major vendors run dedicated programs well below list price:

• Claude for Nonprofits offers Team seats for $8 per user per month, with separate Enterprise pricing.
• OpenAI for Nonprofits offers ChatGPT Business for $8 per user per month when billed annually or $10 monthly, plus Enterprise discounts for larger organizations.
• Google for Nonprofits offers Workspace for Nonprofits at no cost, including the Gemini app with enterprise-grade data protections and NotebookLM as a core service.

Each vendor has its own verification process. Claude and OpenAI currently use Goodstack, while Google verifies eligibility through Google for Nonprofits. Check the current requirements before paying standard rates.

At full list price, paid assistant seats generally start around $20 per person per month, with team and business tiers higher per seat.

Automation platforms (n8n, Make, Zapier, and similar) have low-cost entry tiers. A small organization’s workflows often fit within tens of dollars per month.

Most organizations do not need a separate API account. APIs appear when a third-party app asks for your key or switches among models for you. If you provide a key, the usage appears on your API bill; otherwise, the app may include it in the subscription. Light use may cost only a few dollars a month, and many providers let you cap spending. Everyday staff chat stays in the per-seat apps above.

Start with a nonprofit plan, give paid seats to staff who will use them daily, and keep a small experiments budget. Expand when actual use justifies it.

Go deeper: What is an AI agent? · Is my data used for training?

Permalink

It depends on the vendor, your plan, and your settings. Check all three because policies change.

A few patterns hold across many major AI vendors:

• Business, enterprise, and education tiers typically do not train on your data by default. That commitment usually lives in the data processing terms, and it is a key reason organizations choose paid workspace plans over personal accounts.
• Consumer (free and personal) plans vary. Some train on conversations by default with an opt-out toggle; others have made training opt-in or off by default. This is the tier where settings matter most, and where defaults have shifted over the past few years.
• API access is generally not used for training by default.

What to do:

• Find the training or data-use setting in each tool your staff use, and document what you chose.
• Look for the vendor’s data processing addendum or trust page; search it for terms like “train,” “model improvement,” and “retention.”
• Treat personal accounts used for work as the common gap. Staff may paste supporter data into a free personal account without checking its settings or terms.
• Put your expectations in writing in your AI policy.

Use the source links below to read each vendor’s current policy. The overall pattern has been stable: paid organizational tiers are safer by default, and consumer settings need checking.

Go deeper: Data security and privacy basics · Create your AI policy

Permalink