AI as a Teammate
The expression "AI as teammate" does not mean that artificial intelligence is a person, has its own professional judgement, or can take responsibility for the outcome. It is a working metaphor: it helps treat AI as a collaborator that needs onboarding, context, limits, feedback, and supervision.
This approach is especially useful for agile teams that are starting to integrate generative AI and want to do so in an orderly way, without redesigning their entire work system from day one.
Unlike an improvised use of AI—where each person tries out tools, prompts, and models on their own—AI as a teammate proposes treating AI integration as a team practice.
What problem it solves
Many organizations adopt AI in a fragmented way: one person uses it to summarize meetings, another to write user stories, another to generate code, and another to review texts. The result tends to be uneven: outputs that are hard to compare, doubts about privacy, dependence on individual skills, and a lack of common criteria.
The AI as a teammate approach helps teams move from isolated experimentation to more disciplined integration.
It solves, or at least reduces, four common problems:
- opportunistic use: each person uses AI differently and without shared learning;
- lack of context: AI responds with generalities because it does not understand the product, the team, or its standards;
- inconsistent quality: results depend too heavily on improvised prompts;
- diffuse accountability: no one knows who validates, who is answerable, or what to do when AI gets it wrong.
Why "teammate" and not just "tool"
Calling AI a "teammate" can be useful if it is properly understood. Not because AI is a real member of the team, but because some practices we use with new people also work with AI.
A team doesn't expect a newly onboarded person to get things right without context. It explains the product, the users, the architecture, the working rules, the quality standards, the communication channels, and the decision-making limits.
Something similar happens with AI. If you ask it for work without context, it will produce plausible but generic answers. If you give it context, examples, criteria, limits, and feedback, its outputs improve.
The critical difference is that a person can develop judgment, responsibility, and social understanding. AI cannot. That is why the teammate metaphor should be used as an operational discipline, not as anthropomorphism.
The four pillars of the approach
The AI as a teammate approach is usually organized into four complementary practices.
| Practice | Core question | Risk if neglected |
|---|---|---|
| Model management | Which model or tool do we use for each type of task? | Using the same model for everything, even when it is not suitable. |
| Context management | What information does the AI need to produce relevant outputs? | Generic, incomplete, or product-disconnected responses. |
| Prompt engineering | How do we write clear, repeatable, and verifiable instructions? | Inconsistent results and too many iterations. |
| Governance management | What limits, permissions, policies, and responsibilities apply? | Privacy, security, cost, quality, or accountability risks. |
These four practices work together. Managing models without context produces poor responses. Improving prompts without governance can increase risk. Adding governance without hands-on learning can turn AI into useless bureaucracy.
1. Model management
Model management means selecting and managing the right models, tools, or agents for each type of task.
Not all models are equally suited to every task. Some are better for writing, others for code, others for reasoning, others for document analysis, and others for quick, low-cost tasks. In addition, they change frequently: capabilities, pricing, limits, privacy, tool integration, and performance.
In an agile team, model management involves deciding:
- which tools are permitted;
- which tasks each one is recommended for;
- what data can be entered into each system;
- which tasks require more capable models;
- which tasks can be handled with cheaper or local models;
- when to review whether the chosen model is still suitable.
The goal is not to always chase the most advanced model, but to use the option that is sufficient and safe for each job.
2. Context management
Context management consists of providing the AI with the context it needs to work well.
AI can produce very fluent responses with little context, but that does not mean it understands the product, the strategy, the users, the technical constraints, or the team's previous decisions. If it does not receive that context, it will fill the gaps with assumptions.
Context may include:
- product vision;
- active hypotheses;
- user stories;
- specs;
- acceptance criteria;
- architecture;
- APIs;
- coding standards;
- Definition of Done;
- previous decisions;
- anonymized user data;
- glossaries;
- examples of good and bad outputs.
Managing context does not mean putting all the documentation into every prompt. It means selecting the minimum sufficient information for the AI to produce a relevant result without unnecessary noise.
3. Prompt engineering
Prompt engineering is the practice of designing clear, structured, and reusable instructions to obtain consistent outputs.
In the AI as a teammate approach, the prompt is not an improvised sentence. It is a way of communicating work to the artificial collaborator. It should include the objective, context, task, format, constraints, and quality criteria.
Example of the difference:
| Improvised prompt | Teammate-oriented prompt |
|---|---|
| "Improve this story." | "Review this user story for a backlog refinement session. Return a table with: detected ambiguity, risk to the sprint, clarifying question, and suggested acceptance criterion. Do not invent new requirements." |
The difference is not just length. The second prompt makes it possible to check whether the output is fit for purpose.
4. Governance management
Governance management consists of establishing usage rules, limits, permissions, and responsibilities.
Without governance, AI can introduce significant risks: use of sensitive data, untraceable decisions, runaway costs, reliance on unverified outputs, security errors, or loss of accountability.
Governance should answer questions such as:
- which tools are approved;
- which data must not be shared;
- which tasks require human review;
- who validates an output before it is used;
- how AI use is documented;
- what to do if the AI makes a mistake;
- what spending or usage limits apply;
- which outputs can reach production;
- which policies apply to code, data, customers, or intellectual property.
Governance should not be used to block all experimentation. It should create a safe space to experiment without shifting invisible risks onto the product or the organization.
AI as a Teammate in Scrum
In Scrum, AI as a teammate does not add a new official role. The Scrum Team retains the human responsibilities that belong to it. AI can support the work, but it does not take on accountability.
It can provide support at different points:
- in Product Backlog refinement, detecting ambiguities or proposing questions;
- in Sprint Planning, helping to identify dependencies, risks, or tasks that can be delegated;
- during the sprint, generating drafts, code, tests, documentation, or analyses;
- in the Sprint Review, helping to synthesize feedback or prepare evidence;
- in the Sprint Retrospective, helping to organize patterns, tensions, or improvement experiments.
The key is that AI does not decide for the team. It provides options, analysis, or execution. The team decides.
What Changes in the Roles
In AI-enabled teams, roles are not replaced, but some responsibilities are transformed.
| Role | What changes with AI as a teammate |
|---|---|
| Product Owner / product architect | Defines what deserves to be built, validates hypotheses, and decides when AI can accelerate exploration or delivery without losing focus on value. |
| Developers / product builders | Orchestrate AI, write specs, review outputs, validate technical quality, and ensure that what is generated meets the Definition of Done. |
| Scrum Master / agile enabler | Facilitates the human-AI system: collaboration flows, governance, trust, learning, boundaries, and continuous improvement. |
The biggest change is not that AI "joins the team," but that the team needs to design how to work with it.
How to onboard AI
If AI is treated as an operational teammate, it needs onboarding. That onboarding is not emotional or social, but contextual.
A minimal onboarding might include:
- product purpose;
- main users;
- problems the product solves;
- value criteria;
- architecture or technical constraints;
- Definition of Done;
- documentation tone and standards;
- privacy policies;
- examples of good responses;
- examples of unacceptable responses;
- rules on when it should ask for clarification;
- limits on what it can and cannot do.
This onboarding can take the form of project instructions, a prompt library, specs, style guides, context files, or reusable templates.
Good tasks to start with
The AI as a teammate approach works best when the team starts with low- or medium-risk tasks whose results can be reviewed.
Good initial use cases:
- summarizing user feedback;
- identifying ambiguities in user stories;
- proposing questions for refinement;
- turning vague criteria into testable criteria;
- generating documentation drafts;
- creating review checklists;
- proposing test cases;
- reviewing texts for clarity;
- synthesizing decisions from a meeting;
- preparing a first version of a spec;
- detecting inconsistencies between documents.
These tasks enable fast learning without directly putting the product, the data, or the users at risk.
Tasks that require more oversight
Some tasks can be valuable but require stricter controls:
- generating code that will reach production;
- modifying permissions, authentication, or security;
- analyzing sensitive data;
- responding to customers;
- making roadmap decisions;
- interpreting legal, contractual, or regulatory evidence;
- approving deliverables;
- publishing external content;
- executing irreversible actions.
In these cases, AI can assist, but there must be human review, traceability, and clear acceptance criteria.
Relationship with human/AI attribution
AI as a teammate becomes practical in Sprint Planning when the team explicitly decides what human intelligence will do and what artificial intelligence will do.
Not all work should be delegated to AI. A simple classification might be:
| Type of work | Best owner | Example |
|---|---|---|
| Strategic judgment | People | Deciding whether a hypothesis merits investment. |
| Rapid exploration | People + AI | Creating prototypes or alternatives to learn. |
| Mechanical or repetitive generation | Supervised AI | Generating drafts, scaffolding, initial tests, or variants. |
| Review and integration | People, with AI support | Validating quality, architecture, security, and coherence. |
Human/AI attribution avoids two opposite mistakes: using AI for everything, or not using it where it could actually take work off the team's plate.
Relationship with Definition of Ready for AI
An item shouldn't be assigned to AI just because it "seems automatable." It must be ready.
A Definition of Ready for AI may include:
- clear objective;
- available context;
- sufficient spec or instructions;
- verifiable acceptance criteria;
- permitted data and excluded data;
- limits of action;
- suitable tool or model;
- person responsible for review;
- review time included in planning.
AI without context is fast at producing outputs but slow at generating value.
Relationship with a reinforced Definition of Done
When an AI output is incorporated into an increment, the Definition of Done must explicitly cover it.
It is not enough for the agent to say it has finished. Controls appropriate to the risk must be applied:
- human review;
- tests;
- static analysis;
- security review;
- verification against the spec;
- source checking;
- user validation, if applicable;
- traceability of relevant decisions.
This is especially important for code, technical documentation, data analysis, and decisions that affect users or customers.
Benefits
When properly applied, the AI as a teammate approach can bring:
- faster execution of repetitive tasks;
- better preparation for working sessions;
- greater capacity for exploration;
- less mechanical work;
- support for analysis and synthesis;
- generation of alternatives;
- improved consistency when templates are used;
- early detection of ambiguities;
- greater team learning about its own criteria.
The real advantage is not in "using AI," but in improving the team's ability to think, decide, and deliver with AI support.
Risks
The approach also carries risks if it is misunderstood.
Anthropomorphism
AI can express itself as if it understood, wanted, or remembered things. This makes interaction easier, but it can lead the team to attribute to it judgement, intent, or accountability that it does not have.
Excessive trust
A fluent output can seem correct. AI can be wrong in a convincing way, especially when it works with incomplete information.
Displacement of human judgment
If the team accepts results without discussion, it stops learning and starts depending on the tool.
Loss of shared context
If each person talks to the AI separately and does not share what they learn, knowledge ends up scattered across individual chats.
Lack of traceability
Without a record of prompts, sources, decisions, and reviews, it is difficult to reconstruct why a decision was made or how a result was generated.
Privacy and security risks
The team may enter sensitive data into unauthorized tools or allow actions without adequate permissions.
Useful metrics
To know whether AI is working as a teammate, it is worth measuring more than just frequency of use.
Possible metrics:
- time saved on repetitive tasks;
- reduction in iterations needed to obtain a useful output;
- number of ambiguities detected before the sprint;
- defects found before production;
- human review time per AI output;
- percentage of outputs rejected or reworked;
- hypotheses validated with AI support;
- cost of use versus value obtained;
- team satisfaction with human-AI workflows;
- privacy, security, or quality incidents.
The most misleading metric is "how much the team uses AI." Using it a lot does not mean using it well.
Best practices
To work with AI as a teammate:
- Make onboarding explicit. AI needs context, examples, boundaries, and criteria.
- Create shared templates. Good prompts should move from individual habit to team practice.
- Separate generation from decision-making. AI can propose; the team decides.
- Review critical outputs. The higher the risk, the stricter the controls should be.
- Define permissions. Not every tool should have access to all data.
- Maintain traceability. Record important decisions, sources, and prompts.
- Use retrospectives to improve collaboration with AI. The team can inspect what worked, what failed, and what to change.
- Avoid blind dependence. AI should augment the team's judgement, not replace it.
- Start with reviewable tasks. It is better to learn at low risk before delegating critical tasks.
- Keep practices up to date. Models change, and usage rules must be revisited as well.
When not to use this approach
AI as a teammate is not suitable in every context.
It may not be advisable when:
- the team does not have secure access to AI tools;
- there is not enough technical maturity to review outputs;
- the work falls within highly regulated domains without specific controls;
- the organization does not have clear data policies;
- the cost of review outweighs the benefit;
- the task is deterministic and can be better automated with rules or scripts;
- the team expects AI to make up for a lack of strategic clarity.
AI does not fix a confusing way of working. It speeds it up, and sometimes amplifies its problems.
Difference from AI agent
AI agent and AI as a teammate are not the same thing.
An AI agent is a system capable of executing tasks with a certain degree of autonomy using instructions, context, tools, and controls. AI as a teammate is an integration approach: how the team incorporates those systems into its way of working.
Put simply:
- the agent is the system;
- the teammate is the way of integrating it into the team.
Common mistake
Taking the teammate metaphor literally. AI can collaborate, generate alternatives, and carry out tasks, but it does not belong to the team in the same sense as a person does. It has no professional responsibility, no full contextual judgement, and no commitment to the product. The human team remains responsible for deciding, validating, and answering for the results.
Resources
🏦 IA aplicada al trabajo ágil [AI Applied to Agile Work, in Spanish]Skill Arena · Scrum Manager
📄 Scrum Master en equipos con IA [Scrum Master in AI-Enabled Teams, in Spanish]Free download · Scrum Manager
📄 AI Teammate Framework: A Four-Step Framework for Product TeamsScrum.org framework based on model management, context management, prompt engineering, and governance management · 2025
📄 Managing Your AI Teammate: Turning AI from Experiment to Strategic PartnerScrum.org article on the structured integration of AI into teams · 2025
📄 The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and ExpertiseWorking paper on generative AI and team collaboration · Harvard Business School · 2025
📄 The Cybernetic TeammateNBER version of the study on generative AI, performance, and team dynamics · 2025
📄 AI Risk Management FrameworkAI risk management framework · NIST · 2023
References
- Dell'Acqua, Fabrizio; Ayoubi, Charles; Lifshitz-Assaf, Hila; Sadun, Raffaella; Mollick, Ethan R.; Mollick, Lilach; Han, Yi; Goldman, Jeff; Nair, Hari; Taub, Stew; Lakhani, Karim R. (2025). "The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise," Harvard Business School Working Paper No. 25-043.
- NIST. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). National Institute of Standards and Technology.
- Scrum Manager. (2026). IA aplicada al trabajo ágil [AI applied to agile work, in Spanish]. Scrum Manager.
- Scrum Manager. (2026). Scrum en equipos con IA [Scrum in teams with AI, in Spanish]. Scrum Manager.
- Scrum.org. (2025). "AI Teammate Framework: A Four-Step Framework for Product Teams," Scrum.org.
- Scrum.org. (2025). "Managing Your AI Teammate: Turning AI from Experiment to Strategic Partner," Scrum.org.
See also
Want to advance in agility? You can search for course and exam calls or go at your own pace by becoming a member of the Agile Club. This membership includes exclusive resources, e-learning classrooms, and access to Skill Arena: a space to practice and measure your agile skills at your own pace.