Imagine receiving a green score from every team, then hearing privately that people are afraid to question AI-assisted work. A team health check needs to help you spot and resolve problems. Treating answers as grades can work against that purpose. The useful output is a problem the team understands and a change it can try.
Why a team health check can go wrong
Think of a thermometer used as an exam. If the reading decides your bonus, you have a reason to hold it near the radiator. An instrument cannot tell you much if everyone is trying to produce the approved reading. Before asking for answers, make clear how they will be used.
The idea underneath is psychological safety. Amy Edmondson defined team psychological safety in 1999 as “a shared belief that the team is safe for interpersonal risk taking”. In plain terms, people expect to be listened to, rather than embarrassed or punished, when they raise a problem, ask for help or admit a mistake. That still leaves room to challenge work and expect it to be checked.
Her original peer-reviewed field study covered 51 teams at one office-furniture manufacturer. It linked safety with learning behaviour, such as asking for help and admitting errors, and that behaviour with outside managers’ performance ratings. The study compared teams at a point in time; it cannot establish cause and effect. It also provides no direct evidence about AI-enabled software teams.
What the wider research adds
A 2023 review of 185 studies by Edmondson and Bransby describes the field as mature theory, reflecting agreement on its definition and measurement. That is not the same as settled evidence about which intervention works. The review reports a tendency for safety to vary more between groups than within them. For a leader of several teams, that is a reason to examine differences, rather than rely on a departmental average.
A 2025 software-workplaces review by Santana and colleagues covers 28 studies; 24 of them use Edmondson’s definition. It discusses code review, architectural decisions and task estimation as points of interpersonal tension. That discussion draws particularly on one included study, rather than independent confirmation from every study in the review.
The software review relies heavily on surveys and includes no studies following teams over time. Its search ran in June 2024; it does not establish what changes in AI-assisted work. The counts here were verified in its preprint, not the final published text. The 2023 review selected work published up to August 2021. The exercise below is practical advice to investigate your own teams, not a validated measure of psychological safety or a tested AI intervention.
Agree who will see the answers
Start with the purpose: finding something in the work that the team can improve. Agree who facilitates, who attends, who sees responses and what will leave the room. As the leader of several teams, check whether your presence will help people speak. The person who normally facilitates the team’s discussions may be better placed to run this one.
Do not promise anonymity simply because a form has no name field. Handwriting, a distinctive incident or the details of a role may identify someone. Reading paper answers aloud does not remove that risk. If the proposed method cannot protect a sensitive contribution, change the method or leave that contribution out of the group exercise.
Let people skip a prompt or decline to share an example. Agree a trusted route for sensitive concerns outside the session, using your organisation’s existing arrangements. Before reporting anything beyond the team, check that the wording and detail are acceptable to the people affected. Share agreed themes and actions, without names or identifying stories. Explain any limits to confidentiality before collecting answers.
Team health check questions that open conversations
Useful team health check questions describe work people can discuss and change. Start with a small selection. These prompts are original illustrations for this guide, not items from a research scale:
- What happens when someone questions an AI-assisted change during review?
- What would make it easier to say that AI output needs another check?
- When a mistake surfaces, what helps the team learn from it?
- How much of your week goes on working problems through with other people, rather than alone with a tool?
- Is the current pace one you could keep up for six months?
Ask about patterns in the work, not who is hardest to work with. Invite people to describe what would improve the situation without naming a colleague or recounting an identifiable incident. If you use a scale, explain what its answers mean and ask what would change the team’s interpretation. A number without that context gives you little to act on.
Fitting a team health check into agile routines
In agile teams, a retrospective is a regular meeting to look back at recent work and agree improvements. It can provide a useful slot, but you can arrange a short conversation this week if the next retrospective is further away. Agree the approach with the team first:
- Choose the questions and the method. A team health check survey is one option; a facilitated conversation is another. Select the method around the privacy needs already discussed, rather than a fixed team-size rule.
- Offer a suitable example of your own. Share non-sensitive feedback you received and what you changed, or something AI got wrong in your own work. Speak about a real uncertainty if there is no suitable example. Invite contributions without requiring others to match your disclosure.
- Let the team interpret the picture. Ask what the answers mean in the context of its work. Check which themes people are comfortable discussing or sharing.
- Choose one change. Name an owner and a review date. Make it small enough to start before the week ends.
- Close the loop. At the review, say what happened, what remains difficult and whether the next experiment needs to change.
One relevant finding concerns leaders sharing feedback. Coutifaris and Grant’s 2021 paper reports an experiment that assigned leaders by chance to share or seek feedback. Sharing feedback they had received improved team psychological safety at follow-up; asking for feedback did not. This is reported in the publisher’s abstract, which omits the sample size and setting. It supports considering feedback sharing, but does not validate the exercise here or guarantee an effect in your teams.
What to report across your teams
Agree useful reporting before collecting responses. You may need to see recurring obstacles across several teams, or whether an agreed change helped over time. Keep the explanation alongside any trend and check that teams have interpreted the questions consistently. Different work and circumstances can make direct score comparisons misleading.
Ask for the problem, the action and the support needed. Avoid a league table or raw comments that identify people. If several teams raise the same obstacle, investigate whether a decision above team level needs to change. A team cannot fix a constraint it does not control.
What to look for this week
Start with one willing team and two prompts. By the end of the week, aim to have an agreed issue, one action that has begun, an owner and a date to review it. Check privately with the facilitator whether the method put anyone on the spot. That is a test of the process, not proof that the team now feels safe.
Over later rounds, look for examples of concerns being heard and acted on, and ask what still goes unsaid. Imagine a team that raises an AI-review problem and agrees a clearer way to ask for help. Follow whether the change happened and helped. Neither a greener score nor more criticism alone establishes success.
For a practical example of the work being discussed, read what code review is for when agents draft the code. The free AI Delivery Health Check offers a wider starting point from your own view as a leader: a score, an area to focus on and one thing to try this week. It is separate from the team conversation above. What the Health Check covers explains how to use that view with your teams.

