Building an AI Agent for Real Team Work

Getting your Trinity Audio player ready...

Hi, and welcome to Try AI for Growth, a podcast out of Make Space for Growth. Here, I share short and sometimes surprising stories of how I’ve used technology to tackle everyday challenges. I’m your host, Sara Vicente Barreto, and today I want to talk about something I have been building quietly in the background — an AI Grant Scout agent for Um Pequeno Gesto, the charity I run.

Not a chatbot, not a one-off experiment. A proper, structured tool with a workflow, a set of rules, and a briefing document for the team that uses it. The commercial applications of a tool like this are immense. Whether you are fundraising, searching for clients or partnerships. So I want to tell you how it came together, what it actually does, and — honestly — what is still not quite right.

The Background

Imagine your job in a team is to find money. Not to just ask for it — that comes later — but to scan the landscape, flag the opportunities, score them against your priorities, and tell you which ones are worth pursuing. Literally, finding money. Precious right? And challenging. That is what I set out to build. It has taken time and multiple iterations. But it exists now. And it is already being used.

A bit of context, because this one has a history. Um Pequeno Gesto is a charity I founded that works in Mozambique. Charities need funding, ideally a multi-annual one. Which business does not need money to live? Grants are a critical part of that picture. And not one where we have been particularly successful. The hit ratio is low, so you need to spread your net but also to be laser-focused on the applications, or they can just turn into a huge time waster. Identifying the right opportunities, tracking deadlines, understanding funder priorities, deciding which ones are worth a full application — it is a significant workload, and it is easy to either miss things or spend time on opportunities that were never really a fit.

Evolving Early Experiments

I had started experimenting with this in ChatGPT. I built an early version with a scoring system and a list of funders, and it was useful enough to show me the idea had legs. But we kept getting the same suggestions. The scoring was fairly basic. Part of that was probably because we had not kept iterating. With these tools, you genuinely get out what you put in.

Then we made a decision that changed the approach: we moved the full team to Claude Teams. And with that came a proper rethink. Not just which tool, but how we were building these things at all.

Inside the charity, we now have a dedicated project — a space we use specifically to develop agents and workflows, to iterate, to test, to figure out what actually works in practice before we deploy anything to the wider team. Think of it as our internal lab. It is one of the things I find genuinely useful about Claude Projects: you can create a focused space that holds everything — the instructions, the reference documents, the rules — and whatever you build there stays coherent, because it has context. And that context matters.

It was inside that project that Grant Scout was born.

What are these projects or agents?

Let me explain what Claude Projects actually are, because if you have not come across them yet, it is worth a moment.

A project in Claude is a dedicated space where you give the AI a persistent set of instructions, upload reference documents, and set the rules for how it should behave. It is not a one-off conversation. It is more like briefing a team member properly before they start, and then having them remember that briefing every time you come back. The focus is the point.

A well-set-up project behaves very differently from a blank chat window, because it knows what it is there to do.

One of these projects is our Grant Scout. And building it properly took a few hours. I want to be honest about that, because I think people underestimate this part. It was not an afternoon of typing prompts. It was an iterative process: developing the workflow, writing the rules, expanding and improving the list of funders, building out the scoring system so it actually reflected our priorities rather than generic grant-hunting logic. Think of it as onboarding a team member. If you had someone new in the team to spend time on this process, you would have to give them the background, share important documents, explain your logic, and help them build a workflow.

Now, you may think drafting all this is daunting. And it is, which is why I also used AI for that part. So I first took Claude through what I already had, what was not working, and asked me to query me so we could improve the instructions. Through this process, Claude helped me produce the project instructions and the reference documents — because instructions have a character limit, some of the detail had to live in separate documents that the agent draws on. And then, once that was done, Claude helped me write a briefing for the team member who would be using Grant Scout day to day, so she understood the workflow and could hit the ground running.

So what does my Grant Scout do?

Here is what Grant Scout can actually do. It has different modes, and that structure was deliberate.

The first is the Scout — a broad scan looking for grant opportunities that match the charity’s profile and priorities. The second is a Quick Scan: when we receive a grant newsletter or a funder update, the agent reviews it quickly and tells us whether anything is worth a closer look. The third is a Deep Dive — when we have identified an opportunity that feels promising, Grant Scout goes deeper: more detail on the funder’s priorities, their track record, the fit with our work, the questions we need to answer before committing to a full application. By the way, we build in a final step that allows us to go back and review funders that we left as “keep monitoring” in case there are relevant updates or open calls. We have not tested it much, but it was on Claude’s recommendation, and I decided to keep it.

That structure — scout, scan, dive, review — reflects how grant research actually works. Or at least how it should work when you have time to be strategic about it.

Does it work?

Now, is it perfect? No. Definitely not. We are still fine-tuning the live search capability — making sure the agent can access current information directly rather than drawing on what it already knows, which for grant funding matters, because opportunities close and priorities shift. We are also getting used to the different modes and when to use which. It is a new way of working, and that takes time.

And there is one genuinely frustrating thing I want to be honest about: right now, only I can edit the agent. And the conversations I create within it are not visible to the rest of the team. That is disappointing. The whole point of building something like this is that it becomes a shared tool — not something that lives only in my hands. We are working around it for now. But I wanted to say it clearly, because I think it is important not to make these things sound cleaner than they are. Features evolve all the time, so who knows if by the time you listened to this, this issue is already solved.

The next step, by the way, is to build a Grant Writer — a separate agent that takes the opportunities Grant Scout has identified and helps with the application itself.

The workflow is becoming a proper pipeline.

Lessons Learned

1. The build is the work. It is easy to think of an AI agent as something you set up quickly and then it just runs. That is not how it works — not if you want it to be genuinely useful. The hours spent developing the workflow, refining the scoring system, expanding the funder list, writing clear rules — that is not overhead. That is the product. The agent is only as good as the thinking you put into it. I was mind-boggled, in a good way, at how much more coherent the output became once the instructions were properly structured. But it did not happen automatically. It takes investment.
2. Instructions need to be a document, not a sentence. One of the things I did not fully appreciate until we were deep into this is that the quality of the instructions is everything. Not just the content — the structure, the specificity, the rules. For my first agent with ChatGPT, I had developed the instructions. On this one, I worked through it with Claude and it wrote my instructions, checked character limits and created reference documents for me. This made instructions much clearer and much more specific.
3. Deploying to a team requires a different kind of thinking. Building something for yourself is one thing. Building something that someone else is going to use, reliably, without you in the room, is another. The briefing for the grants team member was not an afterthought, it was part of the project. What is the workflow? What does each mode do? When do you use which one? How do you interpret the output? If you are building agents for your team, that translation step matters enormously. Do not skip it. I see my teammate that is dedicated to grants takes her “briefing” on paper everywhere and jotting down notes on it to make sure we get it right.

What you can try

Is this a charity only type of model? Not really. It works well beyond. I built this for grant funding, but the structure — a scoping agent, a triage layer, a deep-dive mode, a pipeline toward action — applies to a lot of commercial situations. If you are a business looking for partnership opportunities, or a sales team that needs to qualify leads at scale, or a founder tracking funding rounds in a particular sector, the logic is the same. Build the agent around your actual workflow, not around what the tool can theoretically do.

If you are experimenting with Claude Projects or thinking about how to build something like this for your own organisation, I would genuinely love to hear how it goes. Subscribe so you catch the next episode — and if you try it, send me a message. That is honestly how I find out what is worth talking about next.

Grant Scout is not finished. We are still iterating, still finding the edges of what works and what does not, still waiting for some of the collaboration features to catch up with the vision. But it exists. It is being used. And the person responsible for grants has a tool, a briefing, and a workflow — where before she had a spreadsheet and a lot of hope.

Sometimes that is what progress looks like. Not a finished product. A better starting point.

Until next time — keep experimenting and keep having fun.

Photo by Božo Gunjajević