Start from the tasks, not the technology
Most small teams do not have an AI problem. They have a time problem. Five people cover strategy, delivery, admin, communications and whatever came in overnight, and the repetitive tasks land on whoever is free.
An agent is useful exactly where tasks repeat and the rules are known. Below is the list we use when we map a team's work, with examples from an agent we built for a policy research team in Washington, DC. It runs in WhatsApp.
Six jobs an agent can do today
- Monitor. Check websites, publications, news or an inbox on a schedule and send a digest of what changed. The policy team's agent logs new publications and events in its field every morning and flags the notable ones.
- Track. Keep shared logs of decisions, dates, roles and document status, and send each person their open items.
- Draft. Produce first drafts in your house template. For the policy team that means event concept notes and agendas in its own format.
- Research. Build profiles of people, organizations and topics from public sources, with links, so a briefing starts from facts rather than a blank page.
- Translate. Turn documents and messages into the languages your team and partners use.
- Relay. Draft a message to a colleague or partner, confirm it with the person who asked, and send it once approved.
What a shared memory changes
The feature small teams underestimate is shared memory. A chat assistant remembers the conversation in front of it. A team agent keeps logs that everyone reads and writes to, so agreed decisions apply across the team.
That is what turns an assistant into a team member. When a document is marked final, nobody has to be told twice. When the team agrees a format, later drafts follow it.
What an agent should not do
- Send anything unchecked. People approve anything the agent sends on their behalf. Draft, confirm, send is the pattern, every time.
- Make judgment calls. It can tell you an event looks notable. It should not decide which events your organization attends.
- Own a relationship. Partners, donors and clients should hear from a person. The agent prepares the message, the person owns it.
- Work without a log. Everything it does should be recorded so anyone on the team can check what happened and why.
- Expand on its own. New capabilities are added deliberately, after a review, not because the model could do more.
Guardrails that make it safe
Guardrails are not a brake on usefulness. They are what lets a team trust the agent enough to hand it real work.
- A fixed scope. Write down the tasks it does and the tasks it does not. Keep the list short at the start.
- Approval for anything outbound. Messages, emails and posts go through a person.
- Access limits. It reads what it needs and nothing more. Credentials are scoped and logged.
- A weekly review. Someone looks at what it did, what it got wrong and what should change. We do this for every agent we run.
- A clear owner on your side. One person who can say yes or no to changes.
How to start
Pick one workflow, not five. The best first agent does a single daily job that everyone will notice, such as the morning digest or the open-items list. A quick win builds the trust that later, more complex workflows depend on.
Then follow four steps. Map the work by listing the repetitive tasks that cost the most time. Build and train the agent on your documents, templates and rules. Launch it with clear limits on what it can send, share and change. Improve it every week based on what it actually did.
If you want to see what that looks like in practice, read the policy team case study. If you want to map your own team's tasks, book a free call and we will do it with you.