Five teammates, one by one · role, skills, and the line where they hand back

A chat window can do a little of everything, and that is precisely its problem. Drop a general-purpose assistant into a team and two weeks later you get the same explanation for why nobody uses it: people were not sure what to use it for, and for their own work it knew too little.

A teammate is the opposite design. It has a name, a role, a remit, access to specific systems, and a line where it hands back. That line is not a limitation. It is the reason a team actually uses it.

The five, one by one

• Willi, in full William Salespear, is the sales teammate. It researches new contacts, prepares outreach, and keeps the CRM current.
• Rita, in full Margarita Tenderhall, is the tender teammate. It reads tender documents, checks eligibility, and drafts the bid.
• Aml, in full Amelia Compliantia, is the compliance teammate. It supports anti-money-laundering reporting, keeps the audit trail, and classifies use cases under the AI Act.
• Doro, in full Dorothea Knowsmore, is the knowledge teammate. It finds what the company already knows, analyzes contracts, and curates the glossary.
• Reggi, in full Reginald Chartwright, is the reporting teammate. It builds KPI overviews, prepares board reports, and spots a trend before it is obvious.

The names are not decoration. A team that talks about a role talks differently from a team that talks about a tool: it asks what Rita said about eligibility, and it means a result somebody has read and signed off.

How a teammate reaches your data

A teammate holds no data of its own. It works through the MCP connectors on the systems you already run. Four rules apply without exception.

• Access is limited to the sources cleared for that role. Willi sees the CRM, not the personnel file.
• Every query and every action leaves an audit entry.
• Every answer names the source it came from. An answer without a source is a lead, not a result.
• The models are never trained on your content.

Rules two and three sound like paperwork and are the actual difference. A result you can look up gets checked. A result you cannot look up is either taken on faith or ignored, and both are bad outcomes.

What a teammate is not

It is not a person, and it does not sign. A bid stays a draft until somebody approves it. A classification under the AI Act stays preparation until the responsible function confirms it. A suspicious activity report is filed by a human being, because the law requires a human being.

It is also not a substitute for a well-kept system of record. A teammate reads what is there. If the CRM has been full of nonsense for two years, it reads nonsense, and it reads it faster than before.

The line where it hands back

Every role has a point where it returns the case. That point is set by the business, not by the technology, and it belongs in writing before the rollout rather than being discovered during it.

• Anything said to the outside world: if it leaves the building and binds the company, a person approves it.
• Anything with legal effect: classifications, filings, and deadlines are prepared, not decided.
• Contradictory sources: when two systems disagree, the teammate reports the contradiction instead of picking one version.
• Missing ground: if no source can be found, the answer is that no source can be found.

The last point matters most, and it is the one that separates usable systems from unusable ones. Anything that would rather invent something plausible than report a gap has no place in an audited business.

How a team starts

With one role, not five. Ideally the one whose work most visibly piles up today: the research before a client meeting, the first pass over a tender, the hunt for the contract clause somebody half remembers.

After two weeks you can tell whether the output holds. Only then is the second role worth adding, and then it is worth double, because both work off the same index and the same permission model and can pass along what the other one found.

What it does to jobs

The honest answer is that it moves work, and it moves it from assembling to judging. What falls away is the gathering. What grows is the checking, and checking is the more demanding of the two.

That is exactly why Article 4 of Regulation (EU) 2024/1689 has required sufficient AI literacy among your own staff since 2 February 2025. Anyone expected to judge results has to understand how they came about. How a team gets there is covered in the article on the training duty under Article 4.

The full profiles with skills and terms are on the AI teammates page.