# Leading AI Agents as Team Members

> AI agents are led like junior hires, not deployed like magic: with an agent charter (role, mandate, KPI), quality gates before execution, and human sign-off at the decision, the guardrails are the accelerator, not the brake.

- Canonical URL: https://www.convios.com/en/methods/ai-agents-as-team-members
- Language version: https://www.convios.com/de/methodik/ai-agents-team-members
- Status: Convios practice framework
- Method library: https://www.convios.com/en/methods — Markdown: https://www.convios.com/en/methods.md

## Problem

You want to deploy AI agents in sales and go-to-market, but the results are unpredictable: agents hallucinate, produce scrap, or act autonomously in the wrong places. Without defined roles and control points you get either sprawl or, out of caution, nothing at all. Both waste the real lever: routine work to agents, judgment and relationships with humans.

## Approach

1. Write a charter per agent: role, mandate, metric, permitted and forbidden actions, an explicit hand-off point to a human.
2. Install quality gates: human sign-off before the agent acts externally; one measurable KPI per phase.
3. Concentrate human-in-the-loop on the decision: agents handle research and preparation, the human reviews, judges, and owns the relationship.
4. Review agents like employees: weekly inspection of outputs, tightening charter and gates based on what you find.

## Example

A typical case: a B2B SaaS team deploys an agent for account research ahead of first calls. Initially without rules, the dossiers are sometimes brilliant, sometimes invented, and no one trusts the output. The team writes a charter (mandate: research and call preparation; forbidden: sending messages on its own), defines a quality gate (the AE reviews every dossier before use), and measures time saved per preparation. Extended research turns into a short read, while a human still runs every conversation and makes every decision.

## Limits

The approach is a synthesis from practice and not independently validated, operationally plausible, but not a proven framework with comparative studies. It presupposes a clean data and context foundation; agents on bad data merely scale the scrap. For teams without any AI groundwork (poor data quality, no shared playbook) it is a premature step, foundation first, agents second.

## Metric

Output quality at the quality gates (approval versus rework rate) and working time saved per task at constant or better result quality.

## Sources

- Thomas B. Sheridan, William L. Verplank: Human and Computer Control of Undersea Teleoperators. Technical Report, MIT Man-Machine Systems Laboratory für das Office of Naval Research, 1978, 340 S. (DTIC ADA057655), 1978 · academic and scholarly literature · describes the method. Defines human supervisory control as the operating model for semi-autonomous systems, a system capable of autonomous decision-making over short periods and in restricted conditions is remotely monitored and intermittently operated directly or reprogrammed by a person, and names the human supervisor's functions: command (program and commit a task to action), plan, monitor, intervene and trust; Table 8.2 grades the levels of automation in man-computer decision-making, i.e. the threshold at which the machine must involve the human. (https://archive.org/details/DTIC_ADA057655)
- NIST AI Risk Management Framework (AI RMF 1.0), National Institute of Standards and Technology, 2023 — National Institute of Standards and Technology (NIST), U.S. Department of Commerce, 2023 · investment, consulting and analyst firms, industry bodies and public agencies · describes the method. Supports the guardrails of the method: the voluntary framework lists as intended outcomes that roles and responsibilities for human-AI configurations are defined and differentiated (GOVERN 3.2), that human oversight procedures are documented and tested, and that mechanisms are in place to supersede, disengage or deactivate a system (MANAGE 2.4). It creates no obligation. (https://www.nist.gov/itl/ai-risk-management-framework)
- Klaus Christoffersen, David D. Woods: How to make automated systems team players. In: Advances in Human Performance and Cognitive Engineering Research, Bd. 2, Emerald, 2002, S. 1–12, 2002 · academic and scholarly literature · supports the underlying mechanism. States the requirement to design automated agents as team players and names two preconditions: observability (users must perceive what the automated agents are doing and will do next) and directability (users must be able to re-direct machine activities fluently). Also shows that apparent system performance often exists only because humans compensate for machine deficiencies. (https://www.emerald.com/insight/content/doi/10.1016/S1479-3601(02)02003-9/full/pdf)
- Gary Klein, David D. Woods, Jeffrey M. Bradshaw, Robert R. Hoffman, Paul J. Feltovich: Ten Challenges for Making Automation a "Team Player" in Joint Human-Agent Activity. IEEE Intelligent Systems 19(6), 2004, S. 91–95, DOI 10.1109/MIS.2004.74, 2004 · academic and scholarly literature · supports the underlying mechanism. Condenses the conditions under which a machine agent can function as a team member into ten challenges, among them a basic compact to work toward shared goals, adequate models of others' capabilities and intentions, predictability, directability, revealing status and intentions, interpreting signals, goal negotiation, attention management and controlling the costs of coordination. (https://fermatslibrary.com/s/ten-challenges-for-making-automation-a-team-player-in-joint-human-agent-activity)
- Primary source: Thomas B. Sheridan, William L. Verplank: Human and Computer Control of Undersea Teleoperators. Technical Report, MIT Man-Machine Systems Laboratory für das Office of Naval Research, 1978, 340 S. (DTIC ADA057655) (https://archive.org/details/DTIC_ADA057655)

## Related

- Method: [AI-GTM Maturity (4 Levels)](https://www.convios.com/en/methods/ai-gtm-maturity-levels)
- Method: [GTM Stack Signal Routing](https://www.convios.com/en/methods/gtm-stack-signal-routing)
- Method: ["Context: Moat or Wall?" Test](https://www.convios.com/en/methods/context-moat-or-wall-test)
- Tool: [Claude Code](https://www.convios.com/en/toolbox/claude-code)
- Tool: [Infisical](https://www.convios.com/en/toolbox/infisical)
- Tool: [Linear](https://www.convios.com/en/toolbox/linear)