How AI agents do real work, and who checks it
Over the last few weeks I’ve written a lot about where AI is heading, from software going headless to the different levels of AI agency. If you’re an ED or a board member trying to cut through the noise, here’s the short version, along with the question that comes up every time I explain it.
Where the technology is going
- Dashboards are on the way out. Software is being rebuilt so AI agents can operate it, which means far less time spent clicking through 15 browser tabs. Salesforce, the company best known for its dashboards, recently announced its own headless approach.
- Chatbots are the starting line. Typing prompts into a blank window and pasting the results around is closer to micromanaging than delegating.
- Agents that work across your tools are where this is going. They connect to your systems directly and can carry a job with many steps from start to finish.
The goal I’d set for a nonprofit team is to get people out of operating software so they can get back to leading the work.
Why your team still feels underwater
Most organizations have given staff access to AI by now, and plenty of teams still feel buried. A big reason is the app shuffle. Your design tool, CRM and email system each have their own built-in AI, but none of them talk to each other. So your staff act as the bridge. They take data from one app, prompt the AI in a second and paste the result into a third.
That’s the low end of the agency spectrum. The time savings show up when an agent sits above your tools and coordinates the work across them.
How agents get real work done
Two pieces make this possible. The first is MCP, an open standard that works like a universal plug between an agent and your apps, so it can read and write in them directly. The second is code the agent writes for itself. When it hits a gap between two tools, it can write a small script in a secure, separate environment, check that it works and finish the task.
Together, those give an agent the hands to log in and get the project done. They also raise the question I hear from almost every leader.
Who checks the work?
“Who’s making sure it doesn’t send something broken to a major donor?”
A person on your team does. Agents that work this way need a human in the loop, with clear lines between what the agent does and what a person decides:
- The agent does the manual labour. It handles the sorting, the drafting and the matching of data across tools.
- A person acts as editor-in-chief. They review the work, add the nuance and give the final approval.
With firm limits in place, a small team gets the reach of a much bigger operations team and still keeps control of its brand and its compliance. Any agent that can act on its own also needs an off switch that a person controls.
How to set up the checks before you hand over work
- List what counts as going out. Emails, social posts, reports to funders and data exports all belong on it. If it leaves the organization, it’s on the list.
- Make approval a step the agent can’t skip. Build it into the system itself. A line in the instructions asking the AI to check with you isn’t enough.
- Ask for a summary with every draft. The agent should show what data it used, what it changed and what it’s proposing. A reviewer can check that in a few minutes.
- Name a reviewer for each kind of output. Donor emails, social posts and board reports each need a person who signs off, plus a backup for when they’re away.
- Know where the off switch is. Make sure someone can pause the agent quickly, and that everyone knows who that person is.
For leaders: your role becomes editor-in-chief
As agents take on more of the manual work, your job and your managers’ jobs move toward review and judgment. That only works if people trust the system, and trust comes from knowing who checks what.
Before your team hands real work to an agent, get clear answers to these:
- What can our AI tools do without a person approving it?
- Who reviews each kind of output, and how long does that review take?
- If something goes wrong, who can stop it, and how fast?
A person approving the work has been part of Cosly, the AI staff team I’m building for small nonprofits, since the first version.