The work nobody should be doing by hand
How AI agents empower small teams and make big ambitions possible, starting with the three hours a week I took back for a four person consultancy in Milan.
Last spring I was on a call with a consultancy in Milan. Four people, all of them good at what they do, which is writing grant applications for nonprofits and cultural organisations.
Somewhere in that call one of them described her Monday morning. She opens a public registry of newly registered associations. She scrolls. She copies names into a spreadsheet. She searches for an email address for each one, which sometimes exists and sometimes does not. She writes a first message. By the time she starts the work she was hired to do, three hours are gone.
She had been doing this for two years.
We built something that handles the first three hours. The writing stayed with her, and so did the decision about who is worth contacting. Everything underneath moved.
This article is about what that thing is, why I think small teams are the right size for it, and how you would find your own version of those three hours.
What an agent is, once you take the vocabulary away
The word “agent” is doing a lot of work in tech right now, most of it decorative. Underneath it there is a simple idea.
When you use ChatGPT or Claude in a chat window, you are the engine. You open it, you ask, you read, you decide what to do next. It knows nothing about your Monday and it does nothing while you sleep.
An agent is the same intelligence with three things added: a reason to wake up, written instructions it keeps, and permission to use a few tools on its own. The result behaves less like software and more like a new colleague who works entirely from a brief.
In practice every agent is a combination of five parts.
Something wakes it up. A schedule, an incoming email, a button. In Milan it was a timer that runs every morning.
A brief tells it what it is doing. This is a document written in ordinary language. Who it is, what it produces, what counts as a good result, what it must never do. It is the longest part of the work and the part that has nothing to do with technology.
Tools let it touch the world. A language model on its own can only think and write. Tools let it read a website, write a row in a database, draft an email, save a file.
A loop lets it try again. This is the part that separates an agent from a script. It looks at what it found, decides what to do, acts, checks the result, and repeats until the job is done or it runs out of options. A search that comes back empty gets tried differently.
A gate stops it. Somewhere in the process it pauses and waits for a person. In Milan the agent prepares the emails and puts them in a queue. It has never sent one. A human reads each draft, edits it, and presses send.
That last part is the one I want to spend a paragraph on, because it is where most of these projects quietly fail.
The gate is the design work
There is a version of this technology that promises to remove people from the process. I find it both dishonest and, in a team of four, pointless.
Get Giada Pacchioni’s stories in your inbox
Join Medium for free to get updates from this writer.
A small team has no margin for a system that does something embarrassing at scale. One badly addressed email to a foundation you have been courting for a year costs more than the twelve hours the agent saved you that month. The maths only works if the output is reviewable.
So the interesting question stops being “what can it do” and becomes “where does it stop, and what does the person see when it does”.
That is an interface problem. It is the same question as any other piece of design work: what information does someone need, at what moment, to make a decision they can stand behind. The difference is that the user is one of four people in a small office, and the decision is whether something goes out with their name on it.
Everything I know about designing screens turned out to apply here. The queue has to be readable in the time a person actually has. The draft has to be editable without opening three tools. The reason a record was selected has to be visible, otherwise the person approves on faith, which means they approve everything, which means the gate is decorative.
One process, not a platform
The most common way this goes wrong is ambition.
A team decides to “adopt AI”, which means nothing, and ends up with a six month project that touches every part of the business and delivers something nobody trusts. Small teams cannot absorb that. They also do not need to.
The thing I would recommend is boring on purpose: pick one process, the most repetitive one, the one with the clearest inputs and the clearest output, and move that. One. Finish it. Live with it for a month and find out where it annoys you.
A good first candidate usually looks like this. It happens on a schedule. It has a defined starting point, like a registry, an inbox, a folder. The output is a draft or a list, something a person reviews rather than something that acts on the world. And nobody in the team enjoys it.
Common ones in the organisations I work with: monitoring funding opportunities and receiving a filtered weekly digest, turning meeting notes and PDFs into first drafts in the organisation’s own voice, keeping a contact or partner list current with a suggested first message attached, assembling the recurring skeleton of a report.
The honest part
Three things I would want to know before starting.
It is not free and it is not finished. Something changes on the other end, a website reorganises, a format shifts, and it breaks. Budget for maintenance the same way you budget for a website, because it is the same kind of object.
The savings are modest and that is fine. Milan got back roughly three hours a week. Three hours in a team of four is not a revolution. It is one afternoon a month of actual thinking, which turns out to be enough to notice the next thing worth moving.
Handling other people’s data brings obligations that do not disappear because a machine is doing it. If your process touches contact information or sends anything outward, that belongs in a conversation with someone who knows the rules in your jurisdiction, before you build.
Where to start on Monday
The question I now ask in every first conversation is this one:
What does your team do every week that nobody should be doing by hand?
Most people answer it in under ten seconds, which tells you the knowledge was never the problem. Everyone knows where their hours go. What they usually lack is a reason to believe the alternative is small enough to try.
It is. It is one process, one brief, and one moment where a person still decides.
I design digital products and, increasingly, the systems behind them for mission driven organisations. If you are staring at your own three hours, I am always interested in hearing what they are.
