Design the Team Around the Work
How to combine broad product skills, specialist depth, review capacity and learning in an AI-enabled team.

On this page
A small team can prototype an assistant, inspect sample requests and test an interaction without starting a separate project for every step. It may still need help deciding policy, accessing customer records and operating the service.
When shaping the team, account for that work as well as the prototype. Who has authority to make each decision? Who has the expertise to check it? Who has time to maintain the result?
Map the work before the roles
For a new initiative, list the decisions the team will make repeatedly: understanding the customer, deciding the product boundary, setting the quality bar, approving access, handling exceptions and reviewing the result. Name who is accountable and who contributes expertise.
Some work needs close collaboration from the start. A designer may shape a novel interaction before the prototype is built; a domain specialist may identify exceptions that change the entire approach. Other expertise can be shared across several teams when the work is less frequent and the support is available in time.
A role title does not prove a capability is covered. Ask for the work and evidence: who can review this failure, who decides whether it is acceptable and who can change the system?
Give the team authority and support
An outcome alone does not give a team room to act. Agree on the customer and workflow it owns, which decisions it can make locally, the boundaries on data, budget and consequential actions, and the person who can respond to an escalation. Make sure it also has the evidence, access, tools and specialist support needed to exercise that authority.
A team might be free to test historical cases while needing approval to use live customer records or send a response. Record the decision, assumptions and review point in proportion to the consequence. Leaders should explain strategic constraints and honour the decisions they delegate; if a boundary changes, say why.
After a pilot or incident, inspect the decision path as well as the result. Which approvals improved the decision? Where did the team wait without learning? Could anyone stop an unsafe action? Adjust the boundary or support accordingly.
Combine breadth and depth
AI can help a product manager create a working flow, a designer explore data or an engineer inspect customer feedback. That makes collaboration more flexible. It does not make the disciplines interchangeable.
| Expertise | Bring it close when the work involves... |
|---|---|
| Design and accessibility | A consequential interaction, a difficult handoff or a user group with different access needs |
| Research and data | Hard-to-reach users, ambiguous behaviour, measurement validity or causal claims |
| Engineering and operations | Integration, reliability, performance, maintainability or production recovery |
| Security and privacy | Sensitive data, untrusted inputs, new permissions or actions in connected systems |
| Domain, legal or risk | Decisions with professional, regulatory, financial, health or safety consequences |
Involve specialists while the approach is still open to change. A late approval can reveal a problem, but it may arrive after the team has committed to an avoidable design.
Choose a structure that fits

A small cross-functional group can own a bounded problem when it has the necessary access and support. Several product teams may benefit from a shared platform for identity, evaluation and deployment. A high-consequence workflow may need domain, security and operational expertise embedded in the work every day. An exploration group can be temporary, with a specific question and decision date.
These are options rather than maturity stages. The right structure depends on how often expertise is needed, how tightly systems are coupled and what happens when a decision is wrong. A specialist who is spread across too many teams can become an accidental queue even when the organisation chart looks efficient.
Account for the work agents create
An agent needs someone to define its job, maintain context and permissions, inspect quality, handle escalations and retire it when the job ends. Those responsibilities can sit with existing product, operations or platform roles, but they need time and named owners.
As a team generates more changes, design critique, code review or domain approval may become the limiting step. Automate checks where they are reliable, sample where the consequences allow it and keep individual review where the work demands it.
Introduce a changed workflow
Access to a tool is not the same as adoption. Choose a task and a group, map the existing steps and controls, and identify the improvement people should experience. Start with a group that represents the intended users, including people who are not already enthusiastic. Give them current context, a quality standard, time to learn and a way to get help.
Compare repeat use with the baseline, quality, review time, customer impact and cost. Low usage may be a rational response to a workflow that does not fit the job. Before expanding, check whether the result is repeatable, someone owns it and the next group has the support it needs. A successful pilot is evidence for further testing, not proof that every group should adopt the same process.
Consider a hypothetical reporting team that wants help assembling a monthly operational report. Before a pilot, the team records how long preparation and checking take, how often errors need correction, and whether the report reaches its recipients on time. It agrees that the tool can prepare a draft from approved sources, while a person checks the figures and publishes the report.
The pilot includes people with different levels of experience. After several reporting cycles, suppose they use the draft repeatedly and finish the first version sooner, but reviewers spend extra time tracing figures back to their source. The team should not count draft speed alone as a successful adoption. It can improve source references, compare the full preparation and review time with the baseline, and ask reviewers whether the checks are still reliable.
Expand to another reporting group only if the team can maintain the source connections, the quality is acceptable and the next group has the access and review capacity it needs. If tracing remains too difficult, keep the workflow limited or stop it. The time saved in one step must be weighed against the work created in another.
Decide what happens to any capacity returned. Track review waiting time and correction effort, and limit concurrent work to what can be checked. Managers need to make the resulting workload and trade-offs explicit.
Preserve the way people learn
Junior practitioners learn by seeing work develop, encountering mistakes and hearing experienced people explain a decision. Private agent sessions can hide that process. Pair on unfamiliar or high-risk tasks, review failures together and give people bounded work with a clear escalation path.
Learning can flow in both directions. Someone with less experience in the domain may discover a useful way to work with a new tool. Give them room to show it, while making sure a specialist checks assumptions where the consequence warrants it.
During a team review, look at where expertise arrives too late, what is waiting for review and whether newer practitioners can explain why a result is correct. Change the support or workload where needed. The competency profile can help identify strengths and gaps across the team.
v3.1 · Updated Oct 2026