How to Build an AI Teammate That Does Real Work

Ellie at Willa, Editorial lead, Willa · Published September 1, 2026 · Updated September 2, 2026

If you have ever paid for expert help that did not pay you back, this story will sting a little. Divya Gugnani, a five-time founder who has sold three companies, once paid an agency tens of thousands of dollars to build the email flows for her business. When she later had Claude audit the account, she found that a single flow was driving 60 percent of the revenue. The audit did not come from a consultant. It came from an AI she had trained to work like one.

Divya, co-founder of Wander Beauty and founder of the fragrance brand 5 SENS, joined a live Willa session in August 2026 to show members exactly how she builds what she calls AI teammates: specialized agents that read her inboxes, prep her meetings, track her to-dos, and run market research while she saves her attention for decisions only she can make. She runs an investment fund, a real estate portfolio, and a consumer brand while raising three kids, and her system rests on a handful of moves anyone can copy. Here is the playbook from the session.

Follow Divya on Instagram and TikTok at @dgugnani, grab her free guides at divyagugnani.com, and explore Concept to Co.

What is an AI teammate, and how is it different from a chatbot?

A chatbot answers questions. An AI teammate owns a job. As Divya put it during the session, question-and-answer AI was 2023; the shift now is letting AI carry whole roles, so a defined agent handles your inbox triage, your meeting prep, or your research, end to end, and brings you only the parts that need a human.

The difference is not the underlying model. It is setup. A teammate has three things a fresh chat window never has: deep context about you and your work, skills that encode how experts do the job, and connections to the real tools where the work happens. You can build one inside Claude or ChatGPT using features you may already pay for, with no code.

If you want a primer on agents in general before going further, start with our guide to building your first AI agent, then come back here for the teammate setup.

How do you set up an AI teammate step by step?

Divya's framework comes down to context, skills, connectors, and voice. In practice, that looks like five steps.

1. Pick one role and give it a dedicated project. In Claude or ChatGPT, create a project for a single job: vacation planner, inbox chief of staff, meeting prep. Projects hold files, instructions, and memory in one place, so the context you build compounds instead of evaporating when a chat ends.

2. Have the AI interview you. This was Divya's most repeated instruction: "Every time you do a project, you always have it interview you. Do the interview. Don't waste the time not doing the interview." Open the project and prompt something like: "Interview me and ask me everything you need to know to be the best possible [role] for me." Her vacation planner knows her airline status, her kids' ages, and that her 2-year-old naps from 12:30 to 3:30, which is why it returns exact flights timed around nap windows instead of five generic resorts.

3. Add skills. Skills are downloadable packets of expertise, many free on GitHub, that you add through your AI's settings. Doing marketing? Download the marketing skills. Doing finance, legal, or content? Same idea. This is how Divya's Claude became an email marketing expert capable of the audit that found her 60 percent flow.

4. Connect the tools where the work lives. An AI cannot manage an inbox it cannot see. Claude and ChatGPT both have built-in connectors for tools like Gmail, Slack, Notion, Google Drive, and Shopify. Divya routes her many inboxes through Composio so her AI sees the whole picture across accounts.

5. Set guardrails and keep an approval step. Have the teammate put replies in drafts so a human clicks yes or no, at least until it earns trust. More on quality control below.

One more habit ties it together: talk instead of typing. Divya dictates nearly everything through voice tools like Willow Voice or Wispr Flow, because speaking gets far more context into the AI in far less time.

Why should each AI teammate have one job?

A Willa member asked whether one mega-project loaded with marketing, finance, and legal skills would work as well as separate specialized projects. Divya's answer was clear: specialists win. "Do you want a CMO who's been a career CMO for 20 years, or do you want an intern who's doing 5 different jobs and has worn 5 different hats?"

When one project holds every skill, the AI does not know where to pull from, and quality drops. Divya even split her organic social project from her paid social project after noticing the two voices bleeding together: organic sounds like texting a best friend, paid is built to convert, and one generalist project kept mixing them. Scope each teammate the way you would scope a hire.

Which AI teammate roles should you build first?

Divya walked through the four roles she recommends starting with, each with a copy-paste prompt during the session.

• Inbox chief of staff. After she appeared on Good Morning America, her fragrance brand gained 2,000 new customers in 24 hours and the customer service inbox flooded. She had Claude read six months of past inquiries, build a master FAQ document, and draft a reply to every incoming message, so her team just approves drafts instead of triaging hundreds of emails each morning.

• Meeting prep planner. Her AI reviews each day's calendar, checks who she is meeting, and pulls the relevant email history, open negotiations, and pending decisions into a morning brief. Her rule for what a meeting is even for: "Meetings are for decisions."

• Status update writer. Instead of drowning in AI meeting summaries, she keeps one master to-do tracker that her AI updates from voice notes and meeting notes, then surfaces a short daily list of what actually needs her. "All the stuff that used to live rent-free in my head and drive me crazy no longer lives rent-free in my head and drives me crazy."

• Research analyst. Before product or investor conversations, her AI pulls Google Trends data and Reddit threads on the topic, like fragrance note trends, and turns them into data-backed talking points she can bring to the meeting.

Notice what these have in common: high volume, clear inputs, and a human decision at the end.

How do you keep an AI teammate from going off script?

One member shared a cautionary tale: Claude quietly rewrote a client's mission statement inside a presentation, the team missed it, and the client caught it. Divya's fix is guardrails. "Part of AI is telling it what to do, and part of AI is telling it what not to do."

For client work, that means feeding the teammate the client's real website, emails, and brand language, then adding one explicit line of instruction: "Don't write anything that the client wouldn't say themselves." The same principle applies everywhere. Tell your inbox agent which messages it must never answer alone. Tell your writing agent what it may never change. If you want your teammate to sound like you rather than a generic bot, our guide on making AI write like you pairs well with this step.

How do you keep an AI teammate sharp over time?

Divya offered a tip most people learn the hard way: long chats degrade. "Things fall out of context, guys, so don't ever do long chats." Her rule of thumb is 15 to 18 responses, then a reset. Prompt: "Take everything you learned in this chat, create a handoff document, and give me the handoff document." Open a fresh chat in the same project, hand over the document, and keep going with full context and none of the drift.

The deeper principle behind her whole system is about where your attention goes. "Outstanding businesses are built by outstanding people, but you also need to get out of their way," she told the group, and she treats her own energy the same way: if a task does not need her personally, she does not touch it. Or in her words: "Anything that doesn't bring you joy, please use AI to take it off your plate."

Key takeaways

• Give each AI teammate one job. Specialists outperform generalists, so build a separate project per role instead of one do-everything assistant.
• Context is the multiplier. Have the AI interview you, feed it real files and examples, and let it learn your voice from things you actually wrote.
• Skills are downloadable expertise. Free skill files on GitHub can turn a general AI into a credible marketer, analyst, or editor in minutes.
• Connectors turn advice into action. Link your email, calendar, and docs so the teammate can do the work, not just describe it, with drafts and approvals as the safety net.
• Keep chats short and guardrails explicit. Hand off to a fresh chat after 15 to 18 responses, and always tell the AI what it must not do, not just what it should.

The best way to learn this is to build one alongside people doing the same thing. Willa hosts live sessions like this one about every week, taught by women who are actually running these systems in their businesses and careers. Come to the next one at our events page, and if you want the deeper workshops, member matching, and a community that cheers when you ship, we would love to have you: join Willa here.

• AI teammate
• AI agents
• Claude
• connectors
• delegation

Join Willa, the community where women learn AI together.

All posts