Team AI Collaboration: Faster, but More Tired?
AI has made everyone faster, but what about the team? A recent set of interviews gives a slightly uncomfortable answer: individual output is up, but teamwork isn't necessarily better, and some people even feel more tired. This is the cost of team AI collaboration that's easiest to overlook: when someone else makes something with AI, you have to spend time checking it.
Last week we talked about not letting AI be just a personal plug-in. This week we go one step further and look at the new problems that appear once AI output starts flowing through a team, and how to set the rules.
Today's Highlights: Three hidden costs of team AI collaboration
Interviews find: individuals got faster, teams didn't
On October 8, email and docs company Superhuman published a summary of qualitative interviews with US knowledge workers who use AI frequently and mainly work with others, including corporate employees, consultants and small-business owners. Nearly all said AI lets them get more done on their own, but helps teamwork only a little. The sticking point appears the moment someone else needs to take over or get involved.
The report identifies three new collaboration costs.
The first is "verification friction": when a colleague hands you something made with AI's help, you have to spend extra time checking it.
The second is "attention fatigue": AI produces output faster than the team can review it.
The third is "unclear contribution": you can't tell which parts a person wrote and which parts AI changed. Interviewees cope by handing over their prompts, instructions and sources along with the work, so whoever takes over knows the underlying assumptions.
Note that the article doesn't disclose how many people were interviewed, and it closes by promoting the company's own new document product, so its conclusions naturally line up with its product positioning.
Research roundup: how you use AI matters more than how powerful it is
In August, Michael Hogan's team at the University of Galway in Ireland summarized several studies on teams and AI in Psychology Today, mostly involving university students and student teachers. One study found that higher-performing groups checked external sources more than twice as often as lower-performing groups, verifying AI content before using it; lower-performing groups kept re-prompting and pasted AI text directly. Another study observed that when members used AI privately and only brought results to meetings, group discussion and joint decision-making actually decreased; when AI was used in front of everyone, it became shared material for discussion.
Tool update: Claude joins the team channel
Tools are also moving toward "open use."
Anthropic's Claude Tag (beta) lets you add Claude to Slack channels to work alongside the team. A September 24 update lets members use the calendars, drives or CRMs connected to their own accounts when making requests in a channel; personal documents can only be opened by their owner, but Claude's replies are visible to the whole channel. Members can choose to review each reply before it's posted, or use auto mode, where Claude stops for human confirmation only on sensitive content; Enterprise admins can also require everyone to review first. The feature launched first on the Team plan, with Enterprise following.
Jesse's Take: Use AI out in the open and you'll save on verification
What I see when advising small businesses closely matches these interviews. A sales rep uses AI to write a polished quotation note and drops it into the LINE group. The manager senses something is off but can't tell whether the customer data is wrong or AI made it up. In the end both spend half an hour checking Excel, and the time saved is given right back.
Suggestion 1: Attach a "three-line source note" to AI output
A common scenario: an equipment distributor with a dozen-plus staff requires that any document made with AI's help and handed to someone else ends with a three-line note: what data was used, what instructions were given, and which parts AI wrote. One extra minute of writing saves the colleague taking over most of their checking time. Verification friction can't be eliminated, but it can be made cheaper.
Suggestion 2: For important matters, ask AI together in the open
The research reminds us that using AI privately weakens team discussion. For decisions that need consensus, like pricing strategy or replies to complaints, try projecting AI's answer in the meeting so everyone reviews it and picks it apart together. If your team already uses Slack, try adding Claude to a channel. Most companies in Taiwan use LINE, so start with a simple agreement: when posting AI output in the group, label it "AI draft, pending review."
Suggestion 3: Decide first who reviews, and how deeply
Not everything needs a word-by-word check. Internal meeting notes can be spot-checked, but external quotes and contracts must be signed off by an owner. Claude Tag lets users choose between "review before posting" and auto mode, a design small companies can borrow: tier the risk, so review effort goes where it matters most.
To be fair: Superhuman's article paves the way for its own product, and Claude Tag's description is Anthropic's official account; both present the ideal case. The academic studies mostly involved students, so apply them to the workplace with some caution. Still, the principle that AI output should be traceable and reviewed together is especially useful for small companies, because what they lack most is time to redo work.
A first step you can take today
Open the LINE group or shared folder your company uses most and post an announcement: starting today, any document made with AI's help should end with one line, "AI assisted: data used ___, key instructions ___." Check back in a week and see whether fewer files get sent back for rework.The first step in team AI collaboration isn't buying a new tool. It's making AI's contribution visible.
Further Reading
Related posts on JCXO
Original sources
Comments