Claude Code, agentic AI, and what METR's data means for charities

Earlier this year, Anthropic's Claude Code went generally available alongside Claude 4 Sonnet. OpenAI released an Agents SDK. Google has Project Mariner. Microsoft is building agent capabilities into Copilot. The common thread: AI is moving from "answers your questions" to "does things on your behalf."

For charities, the relevant version isn't writing code. It's tasks like processing 200 beneficiary feedback forms into a themed report, or reading an 80-page evaluation and drafting the impact section of a grant application, or cleaning up a messy donor spreadsheet and flagging duplicates. Multi-step work that previously took days of a staff member's time, handled by an AI agent that works through it while you do something else.

This is early. The tools work well for well-defined tasks with clear success criteria. "Process these forms and categorise them" is a good agent task. "Develop our fundraising strategy" is not. Agents are still bad at ambiguity and bad at knowing when to stop and ask a human. We've been using Claude Code for our own work and the productivity gains are real, specifically for repetitive processing of large volumes of information. We've also watched it confidently get things wrong.

"Set and check, not set and forget" is the right way to think about this stage.

Why the pace matters

METR, an independent organisation that evaluates AI capabilities, has been tracking what AI systems can do autonomously. Their data suggests that AI autonomy is doubling roughly every seven months. That's a rate of change worth paying attention to, even if the current tools feel limited.

Here's an analogy that might help. In 2003, if you walked into a Blockbuster, you'd be entering one of the largest entertainment companies in the world. The idea that a DVD rental company had essentially solved content distribution was reasonable. By 2008, it was obvious that something fundamental had shifted and the model was collapsing. The disruption wasn't a single event. It was a compounding change in underlying capability (broadband speeds, streaming technology) that crossed a threshold.

Software engineering feels like it's approaching a similar threshold. Not quite solved by AI, but very close. The tools are not yet reliable enough to work without human oversight, but the gap is closing measurably, and the rate of closure is accelerating.

What this means practically

For charities, the practical implication is about cost and possibility.

The cost of building custom digital solutions is dropping because the tools for building are getting dramatically better. Work that would have required a team of developers for months can increasingly be done by a smaller team in weeks. That changes the economics of "should we build something for this?" The answer is yes more often than it was a year ago, and it'll be yes more often again a year from now.

If your charity has a problem that you've assumed is too expensive to solve with custom technology, it's worth reassessing. Not necessarily today, but with an awareness that what's on the edge of the possible is worth planning for. The organisations that start thinking about their highest-value build problems now will be better placed to act when the tools catch up, which METR's data suggests will happen faster than most people expect.

This isn't an argument for rushing into AI projects. It's an argument for doing the preparatory work: getting your data in order, understanding your processes, identifying where the most time gets wasted. So that when you are ready to build, you can move quickly.

The governance questions get harder as AI takes actions rather than just answering questions. When AI is suggesting text, you can ignore bad suggestions. When AI is processing your data and producing reports, the consequences of errors are more immediate. Who checks the output? What's acceptable accuracy? What decisions should never be delegated to AI, even partially?

The charity sector will adopt agentic AI unevenly, with larger organisations experimenting first. But the underlying shift affects every charity eventually. The question is whether you're planning for that or whether it catches you by surprise.