Can AI connect to our CRM? How do I make this run every weekday at 5pm? Can I turn what I just did into something I can reuse next time? We hear questions like these every week, and our answer is nearly always the same: ask the agent.
An AI agent is an AI tool that can take actions for you, such as reading files, searching the web, or working in your other apps. Today’s AI agents can usually tell you what they can do, and many can set themselves up when you ask in plain English (or any other language you prefer). For questions about the tool itself, that often makes the agent a faster and more current source than a colleague or a blog post, including this one.
This works well in ChatGPT, Claude (including Cowork and Claude Code), and OpenAI Codex. Microsoft Copilot and Gemini can do some of it, but on work accounts more of the setup still happens in menus or through your administrator.
Three ways to ask
Ask it to do the task. You do not need to ask whether it can summarize a PDF. Give it the PDF and say “summarize this.” For features the agent may already have, such as memory or file uploads, trying is more reliable than asking. OpenAI’s own advice is to ask ChatGPT to do the thing rather than ask whether it has access.
Ask whether it can. Save this for things the agent is not set up for yet. “Can you read our donor database?” “Can you open this e-book file?” “What would you need access to in order to do this?” The agent will say yes, no, or yes once something is set up, and in that last case it will usually offer to walk you through it. Keep this question handy:
Can you [DESCRIBE THE TASK]? If not, what is the closest thing you can do?Ask it to set itself up. People skip this one most often, and in our experience it saves the most time. Agents can now create and change the things that shape how they work:
- Skills are saved instructions for a task you do again and again, such as writing a donor update in your house style. Instead of writing one yourself, ask the agent to “create a skill that does X,” or finish a task with it and ask it to turn what you just did into a reusable skill. Our humanizer skill how-to walks through one example.
- Memory is what the agent keeps between conversations. Instead of hunting for the memory settings, say “remember that we write in American English” in the middle of a chat.
- Scheduled tasks run on their own at a set time. Say “every weekday at 5pm, check the shared inbox and send me a summary of anything urgent,” and let the agent create the task and tell you what it needs from you.
- Connectors let the agent read or act in another app, such as Google Drive or your CRM. Ask “can you connect to [APP]?” and it will tell you whether a connector exists and how to turn it on. In most tools you still click to sign in yourself.
Which agents this works in
What each agent can set up for you, as of September 2026. Plans and features change often, so check current help pages, or ask the agent.
| Agent | Asking works | What it can set up when you ask |
|---|---|---|
| ChatGPT | Well | Memories and scheduled tasks on all plans. Skills on business, enterprise, and education plans. It can suggest apps to connect, and you connect them in its settings. |
| Claude, including Cowork | Well | Memories and skills on all plans, and scheduled tasks on paid plans. It suggests connectors, and you click to sign in. On Team and Enterprise plans, an owner may need to switch on memory and approve connectors first. |
| Claude Code | Well | Everything Claude can, plus edits to its own instruction files. It can also download and install the tools it needs, with your permission. Paid plans only. |
| OpenAI Codex | Well | Skills, including one built from the current chat, scheduled tasks, and connections to other tools, which it can often install when you ask. |
| Microsoft Copilot | Partly | Memories, and simple agents built from a description. Scheduled prompts and connectors on work accounts mostly go through buttons or your administrator. |
| Gemini | Partly | On personal Google accounts: memories, skills, scheduled actions, and app connections. On work accounts, including Google Workspace for Nonprofits, memory and skills are not available yet. Google plans to replace Gems with skills on those accounts in March 2027. |
Why asking works now
A year or two ago, this was poor advice. An AI model is trained on text written before it was released, so it knew little about its own newer features. Ask one what it could do and it would often guess, sometimes confidently and wrongly.
Two things have changed. Agents can now see which tools are switched on in the current chat, and many can search their own help pages when they are unsure. Claude Code, for example, has a built-in helper it calls on when you ask about its own features. Agents also sit inside the tools you use, so they can create and edit their own skills, memory, and schedules instead of only describing them.
Asking also works when you do not know the right words. A search engine needs you to know what the thing is called. An agent is fine with a vague question: “Is there a way to stop me retyping the same supporter numbers into every grant report?” is a poor search query and a good prompt.
It helps to stop picturing the agent as software with a fixed menu of features. If you look for “connect to CRM” in a menu and cannot find it, you might assume the answer is no. Agents have no such menu. What they can do depends on what you connect them to and what you ask for.
When to check the answer
Three things can go wrong.
It says no when the answer is yes. In one case we heard about, Claude Cowork insisted that its browser tools were “quite limited” and that it could not do a task, the day after it had done something harder in the same browser. Pushing back (“yes, you can, you did this yesterday”) fixed it. Before you accept a no, challenge it once, ask it to try, or start a new chat.
It says yes when the answer is no. This is the riskier mistake. AI models sometimes report that they finished a step they never ran. If it says it sent the email, check your sent folder. If it says it connected to your CRM, ask it to read back one record.
It tells you how, not whether. An agent will happily build you a scheduled task, a skill, and a CRM connection. It will rarely stop to say you did not need any of them. If you want that answer, ask for it: “Before we build this, is there a simpler way to solve the underlying problem?”
Before it sets itself up
- Check your organization’s data rules. A connector lets the agent see what you can see in that app. Read Data security and privacy basics before connecting anything that holds supporter records or investigation material, and check where that data is processed if you work with people in the EU or UK.
- Some settings belong to your administrator. On business and enterprise plans, an administrator often controls which connectors and features staff can turn on. If the agent says a feature is switched off for you, ask whoever manages your AI account.
- Be careful when it offers to install software. Claude Code and Codex can download and install tools from the internet. They ask your permission first. Only say yes to software from a source you trust, because downloaded code can contain malware.
- Review before anything goes out. Let the agent draft and set things up, and check its work before it sends, publishes, or deletes anything.
Prompts to start with
- “Can you do this? If not, what is the closest thing you can do?”
- “What would you need access to in order to do this?”
- “Turn what we just did into a reusable skill.”
- “Set this up to run every weekday at 5pm, and tell me what you need from me.”
- “You did something like this yesterday. Please try again.”
- “Before you do that, is there a simpler way to solve the underlying problem?”
None of these need any technical knowledge. Next time you are about to ask a colleague whether AI can do something, ask your agent first.
